[
  {
    "content": {
      "TLDR": "We present MusicSem, a dataset of 32,493 language\u2013audio music descriptions derived from organic discussions on Reddit. What sets MusicSem apart is its focus on capturing a broad spectrum of musical semantics, reflecting how listeners naturally describe music in nuanced, human-centered ways. To structure these expressions, we propose a taxonomy of five semantic categories: descriptive, atmospheric, situational, metadata-related, and contextual. Our motivation for releasing MusicSem stems from the observation that music representation learning models often lack sensitivity to these semantic dimensions, due to the limited expressiveness of existing training datasets. MusicSem addresses this gap by serving as a novel semantics-aware resource for training and evaluating models on tasks such as cross-modal music generation and retrieval.",
      "abstract": "We present MusicSem, a dataset of 32,493 language\u2013audio music descriptions derived from organic discussions on Reddit. What sets MusicSem apart is its focus on capturing a broad spectrum of musical semantics, reflecting how listeners naturally describe music in nuanced, human-centered ways. To structure these expressions, we propose a taxonomy of five semantic categories: descriptive, atmospheric, situational, metadata-related, and contextual. Our motivation for releasing MusicSem stems from the observation that music representation learning models often lack sensitivity to these semantic dimensions, due to the limited expressiveness of existing training datasets. MusicSem addresses this gap by serving as a novel semantics-aware resource for training and evaluating models on tasks such as cross-modal music generation and retrieval.",
      "authors": [
        "Rebecca Salganik",
        "Teng Tu",
        "Fei-Yueh Chen",
        "Xiaohao Liu",
        "Kaifeng Lu",
        "Ethan Luvisia",
        "Zhiyao Duan",
        "Guillaume Salha-Galvan",
        "Anson Kahng",
        "Yunshan Ma",
        "Jian Kang"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Dataset"
      ],
      "paper_link": "https://drive.google.com/file/d/1oqH5HvsVFoOfFJyGcopyW4BsOlu-CYEy/preview",
      "poster_link": "https://drive.google.com/file/d/1pdwu9QcX5Ts77JX7OaEihhYGNWc_2qbP/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1_KMa2u00XDOOnbzWROxf_y0mZG560iXp&usp=drive_copy",
      "title": "MusicSem: A Dataset of Music Descriptions on Reddit Capturing Musical Semantics",
      "youtube_id": ""
    },
    "forum": "388",
    "id": "388",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We present MaskBeat, a transformer-based approach for loopable drum pattern generation. Rather than predicting drum hits sequentially, our method uses bidirectional attention with iterative refinement, allowing instruments to be generated in parallel while maintaining musical coherence. Additionally, we introduce custom loss functions that capture drum-specific musical relationships. Our experiments show that MaskBeat generates higher quality and more musically coherent drum patterns than baseline approaches.",
      "abstract": "We present MaskBeat, a transformer-based approach for loopable drum pattern generation. Rather than predicting drum hits sequentially, our method uses bidirectional attention with iterative refinement, allowing instruments to be generated in parallel while maintaining musical coherence. Additionally, we introduce custom loss functions that capture drum-specific musical relationships. Our experiments show that MaskBeat generates higher quality and more musically coherent drum patterns than baseline approaches.",
      "authors": [
        "Luca Lanzendoerfer",
        "Florian Gr\u00f6tschla",
        "Karim Galal",
        "Roger Wattenhofer"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/13DitchhCtZyibocfAngENry9KU58b7Cg/preview",
      "poster_link": "https://drive.google.com/file/d/1FLEWCofRMkuBM4Tpi4EEGqSMbu-HHFjJ/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=11WFURGXejkBhdszYUi3btcIpXDYrpZLp&usp=drive_copy",
      "title": "MaskBeat: Loopable Drum Beat Generation",
      "youtube_id": "https://drive.google.com/file/d/1rnuwRkzrZ_2cHiuRDfiJzFBwL9tUS2RY/preview"
    },
    "forum": "390",
    "id": "390",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We present a novel framework for generating expressive MIDI performances by automatically adding ornaments using PerTok tokenization and a Transformer trained from scratch. In contrast to fine-tuning large language models, our approach employs a domain-specific Transformer designed exclusively for ornamentation, leveraging PerTok's music-aware representation to capture micro-timing and duration nuances critical for expressive performance. The model learns to transform simplified scores into richly ornamented renditions through paired data from the MAESTRO dataset. By training directly on musical data without non-musical pre-training, the system acquires ornamentation patterns in a more stylistically faithful manner. Future work includes comparative evaluation with pre-trained models, exploring alternative tokenization schemes, and conducting systematic evaluation of musicality and stylistic authenticity",
      "abstract": "We present a novel framework for generating expressive MIDI performances by automatically adding ornaments using PerTok tokenization and a Transformer trained from scratch. In contrast to fine-tuning large language models, our approach employs a domain-specific Transformer designed exclusively for ornamentation, leveraging PerTok's music-aware representation to capture micro-timing and duration nuances critical for expressive performance. The model learns to transform simplified scores into richly ornamented renditions through paired data from the MAESTRO dataset. By training directly on musical data without non-musical pre-training, the system acquires ornamentation patterns in a more stylistically faithful manner. Future work includes comparative evaluation with pre-trained models, exploring alternative tokenization schemes, and conducting systematic evaluation of musicality and stylistic authenticity",
      "authors": [
        "Weixi Zhai"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1X02t79AwYR6GnHYDbFqhQHj7RLxP2k2w/preview",
      "poster_link": "https://drive.google.com/file/d/1JxrsYRUNrG2gb7_v0lpoHA6QlQclhX_C/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1Sr3MN7M3HTLrhId1ELvDZtHwlqFhcpVb&usp=drive_copy",
      "title": "Leveraging PerTok and Domain-Specific Transformer Design for Expressive MIDI Ornament Generation",
      "youtube_id": "https://drive.google.com/file/d/10rgc7VRWy_tTUlW6_DLYBPE0Z4ZtWrmE/preview"
    },
    "forum": "392",
    "id": "392",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We introduce PANAMA, an active learning framework for the training of end-to-end parametric guitar amp models using a WaveNet-like architecture. With PANAMA, one can create a virtual amp by recording samples that are determined by an active learning strategy to use a minimum amount of datapoints (i.e., amp knob settings). We show that gradient-based optimization algorithms can be used to determine the optimal datapoints to sample, and that the approach helps under a constrained number of samples.",
      "abstract": "We introduce PANAMA, an active learning framework for the training of end-to-end parametric guitar amp models using a WaveNet-like architecture. With PANAMA, one can create a virtual amp by recording samples that are determined by an active learning strategy to use a minimum amount of datapoints (i.e., amp knob settings). We show that gradient-based optimization algorithms can be used to determine the optimal datapoints to sample, and that the approach helps under a constrained number of samples.",
      "authors": [
        "Florian Gr\u00f6tschla",
        "Luca Lanzend\u00f6rfer",
        "Longxiang Jiao",
        "Roger Wattenhofer"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1175nsE3QsXoGxhLRIk5eM2_kE9j714A0/preview",
      "poster_link": "https://drive.google.com/file/d/16XO8QSPmh5j_xRq-AgcoKL_QZxOcYOn_/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1q7oQCk9c8SYZ6lgv7kf9B3r398OVI48v&usp=drive_copy",
      "title": "Parametric Neural Amp Modeling with Active Learning",
      "youtube_id": "https://drive.google.com/file/d/1noDaOm4UojZpcwUO6TdnYKlJm0BnTVHt/preview"
    },
    "forum": "393",
    "id": "393",
    "position": ""
  },
  {
    "content": {
      "TLDR": "In music, unlike general audio data, there are characteristics according to progression such as structural information and BPM. However, in many MIR models, audio is often preprocessed under fixed information in consideration of compatibility with existing models. This work investigates BPM-aware and structure-based segmentation for music genre classification. We compare three approaches: 30-second whole track, fixed 5-second segments, and musically-informed 2-bar segments on GTZAN dataset. Our experiments demonstrate the feasibility and trade-offs of structure-aware preprocessing. The codes are available online.",
      "abstract": "In music, unlike general audio data, there are characteristics according to progression such as structural information and BPM. However, in many MIR models, audio is often preprocessed under fixed information in consideration of compatibility with existing models. This work investigates BPM-aware and structure-based segmentation for music genre classification. We compare three approaches: 30-second whole track, fixed 5-second segments, and musically-informed 2-bar segments on GTZAN dataset. Our experiments demonstrate the feasibility and trade-offs of structure-aware preprocessing. The codes are available online.",
      "authors": [
        "Seonghyeon Go",
        "Yumin Kim"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1Zk-13od-1-dApEyfkcRobpb95gauoyN2/preview",
      "poster_link": "https://drive.google.com/file/d/1BUuvCZ9WwlSu8tHPTapkecFj9Xxsj3FV/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1x6Qkivj2zSWdS2QO5Aaxtqu8SzHx_vds&usp=drive_copy",
      "title": "FROM FIXED WINDOWS TO MUSICAL BARS: STRUCTURE-AWARE PREPROCESSING IN MIR",
      "youtube_id": ""
    },
    "forum": "394",
    "id": "394",
    "position": ""
  },
  {
    "content": {
      "TLDR": "In this study, we develop a semi-automatic instrument-playing robot, Robo-Sax, which performs key fingerings automatically, while a human player performs the blowing motion. Additionally, we develop four types of Robo-Saxs (soprano, alto, tenor, and baritone) and propose a system that enables ensemble playing using a saxophone quartet. We seek to design the hardware and thoroughly develop the system software such that it can be utilized practically.",
      "abstract": "In this study, we develop a semi-automatic instrument-playing robot, Robo-Sax, which performs key fingerings automatically, while a human player performs the blowing motion. Additionally, we develop four types of Robo-Saxs (soprano, alto, tenor, and baritone) and propose a system that enables ensemble playing using a saxophone quartet. We seek to design the hardware and thoroughly develop the system software such that it can be utilized practically.",
      "authors": [
        "Gou Koutaki",
        "Masatoshi Hamanaka"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1tvtQTc-Y846qIpyiiyqSPA4H39p_uJpt/preview",
      "poster_link": "https://drive.google.com/file/d/1sJatO2vpd4MGIXWPv-ilfxo9G5l6R-uB/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1U3OGHheW75OqybGnyCo7ovgdUC_pQo0D&usp=drive_copy",
      "title": "Robo-sax ensemble system with melody morphing",
      "youtube_id": "https://drive.google.com/file/d/1LJhsfoUumzPZCKdABWuNPwGR-LzXvIk1/preview"
    },
    "forum": "395",
    "id": "395",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Many automatic guitar playing robots have been developed. These include fully-automatic and semi-automatic types, but in this research, we are developing a device in which the robot performs the fingerings to press the strings, while the picking is done by a human. While fully automatic robots are only for viewing, this device is semi-automatic, allowing the user to experience playing. Although there have been devices that can press chords at the 4th fret, we have developed a guitar device that can press up to the 22nd fret by designing it. This allows the user to play the lead guitar part. Users can easily enjoy guitar solos, etc.",
      "abstract": "Many automatic guitar playing robots have been developed. These include fully-automatic and semi-automatic types, but in this research, we are developing a device in which the robot performs the fingerings to press the strings, while the picking is done by a human. While fully automatic robots are only for viewing, this device is semi-automatic, allowing the user to experience playing. Although there have been devices that can press chords at the 4th fret, we have developed a guitar device that can press up to the 22nd fret by designing it. This allows the user to play the lead guitar part. Users can easily enjoy guitar solos, etc.",
      "authors": [
        "Gou Koutaki",
        "Akira Maezawa"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Other"
      ],
      "paper_link": "https://drive.google.com/file/d/18Q7zBNyanIVnHw0txpKnoGVz36uw5JL5/preview",
      "poster_link": "https://drive.google.com/file/d/1KRxxJ8pXsGEJOMiz7V_Umr3hguIx9rU1/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1qV7A3pKh7LZVhg-Bdmq2v4B7sAkxaO97&usp=drive_copy",
      "title": "Robo Lead Guitar: Robotic devices to support fingering of many frets",
      "youtube_id": "https://drive.google.com/file/d/1euMCK_7X0syrmNE4YdgA8wmLSyMFMWlV/preview"
    },
    "forum": "396",
    "id": "396",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We introduce a novel technique for creative audio resynthesis that operates by reworking the concept of granular synthesis at the latent vector level. Our approach creates a \"granular codebook\" by encoding a source audio corpus into latent vector segments, then matches each latent grain of a target audio signal to its closest counterpart in the codebook. The resulting hybrid sequence is decoded to produce audio that preserves the target's temporal structure while adopting the source's timbral characteristics. This technique requires no model training, works with diverse audio materials, and naturally avoids the discontinuities typical of traditional concatenative synthesis through the codec's implicit interpolation during decoding. We include supplementary material, as well as a proof-of-concept implementation to allow users to experiment with their own sounds.",
      "abstract": "We introduce a novel technique for creative audio resynthesis that operates by reworking the concept of granular synthesis at the latent vector level. Our approach creates a \"granular codebook\" by encoding a source audio corpus into latent vector segments, then matches each latent grain of a target audio signal to its closest counterpart in the codebook. The resulting hybrid sequence is decoded to produce audio that preserves the target's temporal structure while adopting the source's timbral characteristics. This technique requires no model training, works with diverse audio materials, and naturally avoids the discontinuities typical of traditional concatenative synthesis through the codec's implicit interpolation during decoding. We include supplementary material, as well as a proof-of-concept implementation to allow users to experiment with their own sounds.",
      "authors": [
        "Nao Tokui",
        "Tom Baker"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1_uxHuWld8822d6AVyPAedx4xdjNvSUS7/preview",
      "poster_link": "https://drive.google.com/file/d/1ZMsT-uVDzXDE3T261-AKNLNRBZSaJHXi/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1rlmfZKFnkYGTLo-RCbsoMnslzNbsUMIx&usp=drive_copy",
      "title": "Latent Granular Resynthesis using Neural Audio Codecs",
      "youtube_id": "https://drive.google.com/file/d/1pBXaJoU0DxlmvJ557uMUNWW6pVyqbGzX/preview"
    },
    "forum": "398",
    "id": "398",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Rhythm complexity has been a central topic in music research for decades. However, most studies have focused on binary onset patterns, neglecting the role of intensity. In this work, we investigate how dynamic accents influence the perception and computational modeling of rhythm complexity. We conducted two listening experiments, examining monophonic and polyphonic patterns featuring both artificial accent sequences and expressive drumming performances. Our findings reveal that metrics originally developed for binary patterns correlate with human ratings just as well or better when considering intensity-varying rhythms compared to those with constant velocity. These results suggest that, despite their simplifying assumptions, classic rhythm complexity measures remain relevant for music information retrieval tasks involving more realistic and expressive rhythmic content.",
      "abstract": "Rhythm complexity has been a central topic in music research for decades. However, most studies have focused on binary onset patterns, neglecting the role of intensity. In this work, we investigate how dynamic accents influence the perception and computational modeling of rhythm complexity. We conducted two listening experiments, examining monophonic and polyphonic patterns featuring both artificial accent sequences and expressive drumming performances. Our findings reveal that metrics originally developed for binary patterns correlate with human ratings just as well or better when considering intensity-varying rhythms compared to those with constant velocity. These results suggest that, despite their simplifying assumptions, classic rhythm complexity measures remain relevant for music information retrieval tasks involving more realistic and expressive rhythmic content.",
      "authors": [
        "AlessandroIlic Mezza",
        "Luca Comanducci",
        "Gabriele Maucione",
        "Massimiliano Zanoni"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1BqGKF286oGRC06MXjXNcRRFZvdiNqDRJ/preview",
      "poster_link": "https://drive.google.com/file/d/16dVTJ7SmzccsVPqet9DUCYQdSRo8hB65/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=18f711ow52nEs4lDTyqFTI2lyg_j-KnE4&usp=drive_copy",
      "title": "Beyond Binary Patterns: A Preliminary Study on the Effect of Dynamic Accents on Rhythm Complexity",
      "youtube_id": ""
    },
    "forum": "399",
    "id": "399",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Wind instruments enable highly expressive performances through playing techniques such as vibrato, slurs, and growl. In this demo, we introduce PULSE-IT, a lightweight synthesis method that combines simple signal processing components with data-driven control signals. Based on a pulsetable oscillator and time-variant filtering, PULSE-IT supports both re-synthesis and cross-synthesis (timbre transfer) of expressive wind instrument playing at low computational cost. Despite its simplicity, the method yields convincing results and provides a practical alternative to more complex neural approaches. Audio examples and interactive visualizations are available at our accompanying website.",
      "abstract": "Wind instruments enable highly expressive performances through playing techniques such as vibrato, slurs, and growl. In this demo, we introduce PULSE-IT, a lightweight synthesis method that combines simple signal processing components with data-driven control signals. Based on a pulsetable oscillator and time-variant filtering, PULSE-IT supports both re-synthesis and cross-synthesis (timbre transfer) of expressive wind instrument playing at low computational cost. Despite its simplicity, the method yields convincing results and provides a practical alternative to more complex neural approaches. Audio examples and interactive visualizations are available at our accompanying website.",
      "authors": [
        "Christian Dittmar",
        "Johannes Zeitler",
        "Stefan Balke",
        "Simon Schw\u00e4r",
        "Meinard M\u00fcller"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1DdwnSHaPZegCe8Ml7X-o4l1sT6NEJlMB/preview",
      "poster_link": "https://drive.google.com/file/d/1JHWaGeR6RKxDlJ3KEajkQ-2wzazhgR6j/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=10zKFUUJsS0w9WDbZHkWBHVbcVL8zctOI&usp=drive_copy",
      "title": "PULSE-IT: Lightweight and Expressive Synthesis of Wind Instrument Playing",
      "youtube_id": "https://drive.google.com/file/d/1DC34i7K7XWHDYWpzPzN_wZu7mbQTgRKg/preview"
    },
    "forum": "403",
    "id": "403",
    "position": ""
  },
  {
    "content": {
      "TLDR": "The recent surge in State Space Models (SSMs), particularly the emergence of Mamba, has established them as strong alternatives or complementary modules to Transformers across diverse domains. In this work, we aim to explore the potential of Mamba-based architectures for music generation. We adopt discrete tokens of Residual Vector Quantization (RVQ) as the modeling representation and empirically find that a single-layer codebook suffices to capture the majority of semantic information in music. Motivated by this observation, we focus on modeling a single-codebook representation and adapt SiMBA, originally designed as a Mamba-based encoder, to function as a decoder for sequence modeling. We compare its performance against a standard Transformer-based decoder. Our results suggest that SiMBA achieves faster convergence and generates outputs closer to ground truth under limited-resource settings, highlighting the promise of SSMs for efficient and expressive music generation. We put audio examples on Github\\footnote{https://lonian6.github.io/web-exploring-ssm/}.",
      "abstract": "The recent surge in State Space Models (SSMs), particularly the emergence of Mamba, has established them as strong alternatives or complementary modules to Transformers across diverse domains. In this work, we aim to explore the potential of Mamba-based architectures for music generation. We adopt discrete tokens of Residual Vector Quantization (RVQ) as the modeling representation and empirically find that a single-layer codebook suffices to capture the majority of semantic information in music. Motivated by this observation, we focus on modeling a single-codebook representation and adapt SiMBA, originally designed as a Mamba-based encoder, to function as a decoder for sequence modeling. We compare its performance against a standard Transformer-based decoder. Our results suggest that SiMBA achieves faster convergence and generates outputs closer to ground truth under limited-resource settings, highlighting the promise of SSMs for efficient and expressive music generation. We put audio examples on Github\\footnote{https://lonian6.github.io/web-exploring-ssm/}.",
      "authors": [
        "Wei-Jaw Lee",
        "Fang-Chih Hsieh",
        "Xuanjun Chen",
        "Fang-Duo Tsai",
        "Yi-Hsuan Yang"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/16XaoJHCl2JEtYcdvnHePa7D-XXiKHrdz/preview",
      "poster_link": "https://drive.google.com/file/d/1-ido-NlIEOUGz92cMAi_9hq_Gc9CVnWU/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1xTScalIBL-s5pahw1ooylNDh4tjxz3ML&usp=drive_copy",
      "title": "Exploring State-Space-Model Based Language Model in Music Generation",
      "youtube_id": ""
    },
    "forum": "404",
    "id": "404",
    "position": ""
  },
  {
    "content": {
      "TLDR": "The music information retrieval community has long benefited from open and royalty-free datasets, both academically and commercially. The artists whose works have become part of these datasets, however, often see little to no tangible compensation for their work, even when their music has become an integral part of a commercially successful task such as music source separation. In this work, we introduce a simple web application that allows users to discover these artists and support their work. This is by no means a sustainable attempt to rectify this long-standing economic disparity, but hopefully will be a small first step in keeping ourselves accountable and conscious of the creators whose creative outputs we so often unfairly rely on.",
      "abstract": "The music information retrieval community has long benefited from open and royalty-free datasets, both academically and commercially. The artists whose works have become part of these datasets, however, often see little to no tangible compensation for their work, even when their music has become an integral part of a commercially successful task such as music source separation. In this work, we introduce a simple web application that allows users to discover these artists and support their work. This is by no means a sustainable attempt to rectify this long-standing economic disparity, but hopefully will be a small first step in keeping ourselves accountable and conscious of the creators whose creative outputs we so often unfairly rely on.",
      "authors": [
        "Karn Watcharasupat",
        "Jo Chung",
        "Yonghyun Kim",
        "Amy Hung",
        "Fabian-Robert St\u00f6ter"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Other"
      ],
      "paper_link": "https://drive.google.com/file/d/1hljqPLyGKuZPmBQiNOQ8q-e6RMnc_Haz/preview",
      "poster_link": "https://drive.google.com/file/d/1XQybsIywJrZ2kIJbBEtfMNiUKydmDJ50/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1Dqc89Jmr5OTbtYLdv6n8H4eOGpTZoKsL&usp=drive_copy",
      "title": "The Secret Source: Supporting the Artists Whose Works Have Made Our Works Possible",
      "youtube_id": ""
    },
    "forum": "406",
    "id": "406",
    "position": ""
  },
  {
    "content": {
      "TLDR": "MusicDocs is an ongoing web-based application designed to support a broad spectrum of Optical Music Recognition (OMR) tasks based on the application of state-of-the-art models. MusicDocs is designed for flexibility: it supports model inference both on server-side and client-side, including in-browser execution via Open Neural Network Exchange (ONNX). This hybrid architecture ensures scalability and allows for privacy-preserving local processing when needed. The application is available at: musicdocs.ai",
      "abstract": "MusicDocs is an ongoing web-based application designed to support a broad spectrum of Optical Music Recognition (OMR) tasks based on the application of state-of-the-art models. MusicDocs is designed for flexibility: it supports model inference both on server-side and client-side, including in-browser execution via Open Neural Network Exchange (ONNX). This hybrid architecture ensures scalability and allows for privacy-preserving local processing when needed. The application is available at: musicdocs.ai",
      "authors": [
        "Joan Cerveto-Serrano",
        "Antonio R\u00edos-Vila",
        "David Rizo",
        "Jorge Calvo-Zaragoza"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1GZxGcuLBzj3KH4HVLb_uennb6pSxbCn2/preview",
      "poster_link": "https://drive.google.com/file/d/1IuvIvUAyqZYt1J-aBMM8Y5OWri6_Z4zM/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=10G0tR_32wy5n98ESnTqg9oJ6VctZRFW4&usp=drive_copy",
      "title": "MUSICDOCS: A web-based application for Optical Music Recognition",
      "youtube_id": ""
    },
    "forum": "407",
    "id": "407",
    "position": ""
  },
  {
    "content": {
      "TLDR": "This work investigates the performance and robustness of lyrics alignment models under varying vocal conditions. While recent models perform well on either mixtures or separated vocals, real-world applications require robustness without assuming access to a specific separation method. We first extend a previously proposed alignment model to the multilingual setting by expanding the phoneme vocabulary and training on multilingual data. Evaluation on the Multi-Lang Jamendo dataset shows strong performance across multiple languages. To support more detailed evaluation, we introduce a word-level timestamp extension of the MUSDB18 test set. Using this resource, we compare model performance on clean vocal stems and vocals processed by three widely used source separation tools. The results reveal substantial variation depending on the separation method, with alignment accuracy generally improving as vocal quality increases. These findings highlight the limitations of current models under realistic audio conditions and the importance of evaluating alignment systems in diverse scenarios. All resources are publicly released to support further research.",
      "abstract": "This work investigates the performance and robustness of lyrics alignment models under varying vocal conditions. While recent models perform well on either mixtures or separated vocals, real-world applications require robustness without assuming access to a specific separation method. We first extend a previously proposed alignment model to the multilingual setting by expanding the phoneme vocabulary and training on multilingual data. Evaluation on the Multi-Lang Jamendo dataset shows strong performance across multiple languages. To support more detailed evaluation, we introduce a word-level timestamp extension of the MUSDB18 test set. Using this resource, we compare model performance on clean vocal stems and vocals processed by three widely used source separation tools. The results reveal substantial variation depending on the separation method, with alignment accuracy generally improving as vocal quality increases. These findings highlight the limitations of current models under realistic audio conditions and the importance of evaluating alignment systems in diverse scenarios. All resources are publicly released to support further research.",
      "authors": [
        "Jiawen Huang",
        "Emmanouil Benetos"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1K-ozHj3d2rqFTp2ESwQy3c4I_MUwJ9w1/preview",
      "poster_link": "https://drive.google.com/file/d/13Lr7wfDf5CAxVy4kxTmR3APXKOofD_L9/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1GQRC0FOmotB8W4XTQgNKpGS6hGWuC-ON&usp=drive_copy",
      "title": "Evaluating Lyrics Alignment under Source Separated Conditions",
      "youtube_id": ""
    },
    "forum": "412",
    "id": "412",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We develop a system that enables a beginner of musical performance to easily participate elaborated ensemble performances with an experienced musician. This system is composed of two key elements: a robo-instrument, which enables the beginner player to take a part in the ensemble by automatic mechanical controls of fingering motions for musical instruments such as saxophone, and a realtime music alignment (score following) method that takes the performance signal of the experienced musician as input and estimates the current position in the performed musical score. To realize a seamless and flexible ensemble, we study a MIDI-based score following method that can treat performance errors, repeats, and large skips, and a method for predictive transmission of MIDI events to the robo-instrument. We demonstrate this system with pieces for piano and saxophone, where the latter part is played with a robo-saxophone, and discuss new problems of music information processing required for extending the research on such realtime interactive systems.",
      "abstract": "We develop a system that enables a beginner of musical performance to easily participate elaborated ensemble performances with an experienced musician. This system is composed of two key elements: a robo-instrument, which enables the beginner player to take a part in the ensemble by automatic mechanical controls of fingering motions for musical instruments such as saxophone, and a realtime music alignment (score following) method that takes the performance signal of the experienced musician as input and estimates the current position in the performed musical score. To realize a seamless and flexible ensemble, we study a MIDI-based score following method that can treat performance errors, repeats, and large skips, and a method for predictive transmission of MIDI events to the robo-instrument. We demonstrate this system with pieces for piano and saxophone, where the latter part is played with a robo-saxophone, and discuss new problems of music information processing required for extending the research on such realtime interactive systems.",
      "authors": [
        "Eita Nakamura",
        "Gou Koutaki",
        "Masatoshi Hamanaka"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1fYKBGqAf9bfVkfZr84rMDtUYhBgQb3Rk/preview",
      "poster_link": "https://drive.google.com/file/d/1JxHdVBJPhiC_k9n6oYUKOa7gPU4Mx6_9/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=181W6lasUWExAoRH0PncZpPK0L3G7Axa0&usp=drive_copy",
      "title": "Ensemble System for Robo-Instrument Player and Experienced Musician Using Score Following",
      "youtube_id": ""
    },
    "forum": "419",
    "id": "419",
    "position": ""
  },
  {
    "content": {
      "TLDR": "This poster presents an ongoing study comparing rule-based and deep learning approaches to modeling 13th-century motet composition. The rule-based path uses the harMo13 dataset, symbolic analysis tools (e.g., Music21, Humdrum), and encodes the data in **kern format. Unsupervised learning is applied to extract potential compositional rules. The deep learning path explores fine-tuning, polyphonic adaptation, and custom training. While still in progress, the study proposes a multi-dimensional evaluation framework to assess stylistic validity, rule compliance, and perceptual coherence, aiming to inform both reconstruction and generative modeling of early polyphony.",
      "abstract": "This poster presents an ongoing study comparing rule-based and deep learning approaches to modeling 13th-century motet composition. The rule-based path uses the harMo13 dataset, symbolic analysis tools (e.g., Music21, Humdrum), and encodes the data in **kern format. Unsupervised learning is applied to extract potential compositional rules. The deep learning path explores fine-tuning, polyphonic adaptation, and custom training. While still in progress, the study proposes a multi-dimensional evaluation framework to assess stylistic validity, rule compliance, and perceptual coherence, aiming to inform both reconstruction and generative modeling of early polyphony.",
      "authors": [
        "Xuhong Qiu",
        "Emilia Parada-Cabaleiro",
        "Frank Hentschel"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1w5tDythQ-1vBLaLHGABSFjnUIVnnlMvM/preview",
      "poster_link": "https://drive.google.com/file/d/1OigaltRtyOIxXEaJiIm1JhMAVOmoS3Y0/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1FD6aaR-UTFlM2gwFGXwC3JNMdEiivDG9&usp=drive_copy",
      "title": "Reconstructing 13th-Century Motets: A Comparative Study of Rule-Based and Deep Learning Models",
      "youtube_id": ""
    },
    "forum": "420",
    "id": "420",
    "position": ""
  },
  {
    "content": {
      "TLDR": "In this work, we explore the use of Osu!, a community-based rhythm game, as an alternative source of beat and downbeat annotations. Osu! beatmaps are created and refined by a large, diverse community and span underrepresented genres such as anime, Vocaloid, and video game music. We introduce a pipeline for extracting annotations from Osu! beatmaps and partition them into meaningful subsets. Through manual analysis, we find that beatmaps with a single timing point or widely spaced multiple timing points (\u22655 seconds apart) provide reliable annotations, while closely spaced timing points (<5 seconds apart) often require additional curation. We also observe high consistency across multiple annotations of the same song. This study demonstrates the potential of Osu! data as a scalable, diverse, and community-driven resource for MIR research. We release our pipeline and a high-quality subset osu2beat2025 to support further exploration: https://github.com/ziyunliu4444/osu2mir.",
      "abstract": "In this work, we explore the use of Osu!, a community-based rhythm game, as an alternative source of beat and downbeat annotations. Osu! beatmaps are created and refined by a large, diverse community and span underrepresented genres such as anime, Vocaloid, and video game music. We introduce a pipeline for extracting annotations from Osu! beatmaps and partition them into meaningful subsets. Through manual analysis, we find that beatmaps with a single timing point or widely spaced multiple timing points (\u22655 seconds apart) provide reliable annotations, while closely spaced timing points (<5 seconds apart) often require additional curation. We also observe high consistency across multiple annotations of the same song. This study demonstrates the potential of Osu! data as a scalable, diverse, and community-driven resource for MIR research. We release our pipeline and a high-quality subset osu2beat2025 to support further exploration: https://github.com/ziyunliu4444/osu2mir.",
      "authors": [
        "Ziyun Liu",
        "Chris Donahue"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Dataset"
      ],
      "paper_link": "https://drive.google.com/file/d/1tzBzFqnh7cgpjBtCY0TKaiZHGgzF8oYD/preview",
      "poster_link": "https://drive.google.com/file/d/1avsJRV-orBEdPAawtLgo4QoVzcv86cG-/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1332MMmD21c723axKdUzF4VzbOBr3sjz3&usp=drive_copy",
      "title": "Osu2MIR: Beat Tracking Dataset Derived From Osu! Data",
      "youtube_id": ""
    },
    "forum": "421",
    "id": "421",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We present \\texttt{DiffFx}, an open-source toolkit for audio effects research. The toolkit provides differentiable audio effects processors built in PyTorch for seamless integration into deep learning training frameworks. It can be used for multiple purposes, including mixing and mastering, style transfer of audio effects, audio effects parameter matching. The code is available at \\url{https://github.com/ytsrt66589/diffFx-pytorch}.",
      "abstract": "We present \\texttt{DiffFx}, an open-source toolkit for audio effects research. The toolkit provides differentiable audio effects processors built in PyTorch for seamless integration into deep learning training frameworks. It can be used for multiple purposes, including mixing and mastering, style transfer of audio effects, audio effects parameter matching. The code is available at \\url{https://github.com/ytsrt66589/diffFx-pytorch}.",
      "authors": [
        "Yen-Tung Yeh",
        "Chung-Jui Chan",
        "Yi-Hsuan Yang"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1y6OfrUK2Q-0vn2N-ZHaw90-BPoM7nVs-/preview",
      "poster_link": "https://drive.google.com/file/d/1kdG6kn4SP7u2KZirzALmmAfGIqfDnBLc/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1ilpH671O2M_BWlaVuhxMxtieSEU5wNr5&usp=drive_copy",
      "title": "DiffFx: A Toolkit for Differentiable Audio Effects Processors",
      "youtube_id": ""
    },
    "forum": "422",
    "id": "422",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Digital annotations can be used by MIR systems to record, contextualise, and visualise analytical outputs, or to encode and structure ground truth for processing as an input. We present two related recent developments for the annotation of digital music scores. The first concerns Verovio, a popular music engraving software embedded within many web-based notation tools, previously lacking the capability to flexibly process and display annotations. We describe a new implementation for handling Music Encoding Initiative (MEI) XML <annot> elements in Verovio, providing default visualisations alongside identification of annotated elements in SVG graphical output for extensible customisation. Secondly, we introduce mei-friend, an interactive editor for MEI with an embedded Verovio notation viewer, with additional interface support for the creation and editing of <annot> annotations.",
      "abstract": "Digital annotations can be used by MIR systems to record, contextualise, and visualise analytical outputs, or to encode and structure ground truth for processing as an input. We present two related recent developments for the annotation of digital music scores. The first concerns Verovio, a popular music engraving software embedded within many web-based notation tools, previously lacking the capability to flexibly process and display annotations. We describe a new implementation for handling Music Encoding Initiative (MEI) XML <annot> elements in Verovio, providing default visualisations alongside identification of annotated elements in SVG graphical output for extensible customisation. Secondly, we introduce mei-friend, an interactive editor for MEI with an embedded Verovio notation viewer, with additional interface support for the creation and editing of <annot> annotations.",
      "authors": [
        "Werner Goebl",
        "David Lewis",
        "Kevin Page",
        "Laurent Pugin",
        "David Weigl"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1SM-QnsT2u1VMF3yICAuDecBfC8avfb-x/preview",
      "poster_link": "https://drive.google.com/file/d/1yWd1NO7UuHBEdmXU-QAXaaR8jah9DtSU/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1-DlzMaleEy26MBGE8VZI_4byIN2j8JxE&usp=drive_copy",
      "title": "Annotations for MIR with Verovio and mei-friend",
      "youtube_id": ""
    },
    "forum": "423",
    "id": "423",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Interpolating music samples has the potential to enhance both the research workflow and the creativity of musicians. For the former, it can be used for augmentation, as data scarcity is a frequent hindrance to MIR research, whereas musicians can discover interesting soundscapes to incorporate in their performances and compositions. There are vast possibilities to perform the same musical phrases not only with different timbre but also with different tempo, timing, and accents. We approach this problem through the lens of the embedding space Music2Latent. Our preliminary work empirically explores how interpolated latents correspond to different musical properties. To ensure temporal consistency, we apply classical interpolation algorithms that are devised to handle sequences of different lengths.",
      "abstract": "Interpolating music samples has the potential to enhance both the research workflow and the creativity of musicians. For the former, it can be used for augmentation, as data scarcity is a frequent hindrance to MIR research, whereas musicians can discover interesting soundscapes to incorporate in their performances and compositions. There are vast possibilities to perform the same musical phrases not only with different timbre but also with different tempo, timing, and accents. We approach this problem through the lens of the embedding space Music2Latent. Our preliminary work empirically explores how interpolated latents correspond to different musical properties. To ensure temporal consistency, we apply classical interpolation algorithms that are devised to handle sequences of different lengths.",
      "authors": [
        "Dmitrii Gavrilev"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1qQyzTmAbufSq1stMAbj6iK2DZIj33s3D/preview",
      "poster_link": "https://drive.google.com/file/d/1O_Xhl836fMNX-qSftOcyTBU2423GaGPK/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1eMin7FtB-Ha3kD9Z-BFlIIgS5UHxqrRK&usp=drive_copy",
      "title": "Towards Time-Aware Music Interpolation in Latent Space",
      "youtube_id": ""
    },
    "forum": "425",
    "id": "425",
    "position": ""
  },
  {
    "content": {
      "TLDR": "With the rapid advancement of generative audio models, distinguishing between human-composed and generated music is becoming increasingly challenging. As a response, models for detecting fake music have been proposed. In this work, we explore the robustness of such systems under audio augmentations. To evaluate model generalization, we constructed a dataset consisting of both real and synthetic music generated using several systems. We then apply a range of audio transformations and analyze how they affect classification accuracy. We test the performance of a recent state-of-the-art musical deepfake detection model in the presence of audio augmentations. The performance of the model decreases significantly even with the introduction of light augmentations.",
      "abstract": "With the rapid advancement of generative audio models, distinguishing between human-composed and generated music is becoming increasingly challenging. As a response, models for detecting fake music have been proposed. In this work, we explore the robustness of such systems under audio augmentations. To evaluate model generalization, we constructed a dataset consisting of both real and synthetic music generated using several systems. We then apply a range of audio transformations and analyze how they affect classification accuracy. We test the performance of a recent state-of-the-art musical deepfake detection model in the presence of audio augmentations. The performance of the model decreases significantly even with the introduction of light augmentations.",
      "authors": [
        "Tomasz Sroka",
        "Tomasz W\u0119\u017cowicz",
        "Dominik Sidorczuk",
        "Mateusz Modrzejewski"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1Uw9q257_Xp51-uNOh8KsNO2XLfudwoxJ/preview",
      "poster_link": "https://drive.google.com/file/d/1enmhemGrZjzq2Di9okakgR9xQPWTJXvk/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1mPGh6Ip6HvdDc7bu9QHZHxN-FNSdZ-tE&usp=drive_copy",
      "title": "Evaluating Fake Music Detection Performance Under Audio Augmentations",
      "youtube_id": "https://drive.google.com/file/d/1n7TpGIXgWCCtp9rm5tSbTE3BhQr2B7g5/preview"
    },
    "forum": "426",
    "id": "426",
    "position": ""
  },
  {
    "content": {
      "TLDR": "harMo13 is a computational corpus studies project that investigates the sonorities of thirteenth-century motets. Our goal is to uncover whether there is an underlying grammar to the sonorities of motets, a question that has hitherto eluded modern scholarly analysis. We focus on motets in which phrases of the tenor (termed cursus) are repeated, and we use these repetitions to compare polyphonic settings of the same melodic material, thereby identifying recurring patterns in the use of sonorities. Our corpus comprises 73 three- and four-voice motets from four major medieval manuscripts. Parts of the following abstract are summarized from the research materials of the Hentschel Group, especially a recent forthcoming publication by R. Eisinger et al. X. Qiu is presenting author of the proposed poster.",
      "abstract": "harMo13 is a computational corpus studies project that investigates the sonorities of thirteenth-century motets. Our goal is to uncover whether there is an underlying grammar to the sonorities of motets, a question that has hitherto eluded modern scholarly analysis. We focus on motets in which phrases of the tenor (termed cursus) are repeated, and we use these repetitions to compare polyphonic settings of the same melodic material, thereby identifying recurring patterns in the use of sonorities. Our corpus comprises 73 three- and four-voice motets from four major medieval manuscripts. Parts of the following abstract are summarized from the research materials of the Hentschel Group, especially a recent forthcoming publication by R. Eisinger et al. X. Qiu is presenting author of the proposed poster.",
      "authors": [
        "Xuhong Qiu",
        "Robert Eisinger",
        "Philippa Ovenden",
        "David Ubber",
        "Frank Hentschel"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1grDnl4WaQXVJ7CsWRbUPalidYj0O8VxK/preview",
      "poster_link": "https://drive.google.com/file/d/1-EliDGbYVGZR-9Piimx5NcUmR6UuqQzq/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1MNpva7gkPil8xIJT5-jwwQs77KrCp9e3&usp=drive_copy",
      "title": "harMo13 \u201cHarmony\u201d in Thirteenth-Century Three- and Four-Voice Motets A Corpus-Based Analysis",
      "youtube_id": ""
    },
    "forum": "427",
    "id": "427",
    "position": ""
  },
  {
    "content": {
      "TLDR": "In this paper, we propose an Optimal Transport objective for learning one-dimensional translation-equivariant systems and demonstrate its applicability to single pitch estimation. Our method provides a theoretically grounded, more numerically stable, and simpler alternative for training state-of-the-art self-supervised pitch estimators.",
      "abstract": "In this paper, we propose an Optimal Transport objective for learning one-dimensional translation-equivariant systems and demonstrate its applicability to single pitch estimation. Our method provides a theoretically grounded, more numerically stable, and simpler alternative for training state-of-the-art self-supervised pitch estimators.",
      "authors": [
        "Bernardo Torres",
        "Alain Riou",
        "Ga\u00ebl Richard",
        "Geoffroy Peeters"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1Don0yZfBLwLYBWcGF7XpNxoZ4IuSWaDK/preview",
      "poster_link": "https://drive.google.com/file/d/16z9KfO8ao9dlbnJx34qNM5cYrUnJQXt1/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=15GS5WfgP2Iadpirk6NDvVMyllXHaBMfS&usp=drive_copy",
      "title": "Translation-equivariant Self-supervised learning For Pitch Estimation with Optimal Transport",
      "youtube_id": ""
    },
    "forum": "428",
    "id": "428",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Neural audio synthesis has become a central topic of research given the high quality of generation currently obtained by deep probabilistic models. However, these models promote task-centric approaches that solve a given type of generation with limited interactivity and control. Hence, those models lack transformative operations that can allow to craft the sound iteratively without a fixed objective. In this paper we propose the idea of latent wavetable synthesis, in which we consider the latent variables of audio models as signals that can be manipulated with operations pertaining to wavetable synthesis. This allows to obtain a new form of audio synthesis that is performative and real-time, while maintaining the quality and novelty of neural generation. We implement this approach through a Max4Live device that relies on specific externals coded in C++ to allow for integration in creative workflows. The device and all corresponding open-source code and results are available on our supporting webpage",
      "abstract": "Neural audio synthesis has become a central topic of research given the high quality of generation currently obtained by deep probabilistic models. However, these models promote task-centric approaches that solve a given type of generation with limited interactivity and control. Hence, those models lack transformative operations that can allow to craft the sound iteratively without a fixed objective. In this paper we propose the idea of latent wavetable synthesis, in which we consider the latent variables of audio models as signals that can be manipulated with operations pertaining to wavetable synthesis. This allows to obtain a new form of audio synthesis that is performative and real-time, while maintaining the quality and novelty of neural generation. We implement this approach through a Max4Live device that relies on specific externals coded in C++ to allow for integration in creative workflows. The device and all corresponding open-source code and results are available on our supporting webpage",
      "authors": [
        "Philippe Esling",
        "David Genova"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1zO1Bm-wWfSBR0gWZjRU_2DBanSTCBM68/preview",
      "poster_link": "https://drive.google.com/file/d/1j7CvyF9BcfovQS8CS95D93dPUq8Tksi-/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1drObk6Re1OBj8RHxYGyvaxz-27wnd9tO&usp=drive_copy",
      "title": "Ravetable: Neural latent wavetable audio synthesis",
      "youtube_id": ""
    },
    "forum": "430",
    "id": "430",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Recent large pre-trained music models have demonstrated competitive results on a variety of tasks. However, the potential benefits of large pre-training remain underexplored for many specific music tasks, including pitch estimation. This work systematically investigates the performance of Data2Vec-music when fine-tuned for pitch estimation. Our findings reveal the utility of its convolutional backbone, which we leverage to construct a computationally efficient, general-purpose pitch tracker. Our model achieves competitive performance with CREPE on monophonic data and NMP on polyphonic data. Along with first results, we compile a benchmark from various datasets to assess diverse aspects of pitch estimation.",
      "abstract": "Recent large pre-trained music models have demonstrated competitive results on a variety of tasks. However, the potential benefits of large pre-training remain underexplored for many specific music tasks, including pitch estimation. This work systematically investigates the performance of Data2Vec-music when fine-tuned for pitch estimation. Our findings reveal the utility of its convolutional backbone, which we leverage to construct a computationally efficient, general-purpose pitch tracker. Our model achieves competitive performance with CREPE on monophonic data and NMP on polyphonic data. Along with first results, we compile a benchmark from various datasets to assess diverse aspects of pitch estimation.",
      "authors": [
        "Silvan Laube",
        "Brandon Panos"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1wIWzdiY1qIMk1Xdsk7B6wxSseROI3t2R/preview",
      "poster_link": "https://drive.google.com/file/d/1SD7cyvi6XrGWPVrBUvqNoy2bbbT1W1ey/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=13cZP7M7wHqiexys_KdDwe1jdklrh54qa&usp=drive_copy",
      "title": "COREPIT: A Convolutional Real-Time Pitch Tracker",
      "youtube_id": ""
    },
    "forum": "431",
    "id": "431",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Piano performance is a multimodal activity that intrinsically combines physical actions with the acoustic rendition. Despite growing research interest in analyzing the multimodal nature of piano performance, the laborious process of acquiring large-scale multimodal data remains a significant bottleneck, hindering further progress in this field. To overcome this barrier, we present an integrated web toolkit comprising two graphical user interfaces (GUIs): (i) PiaRec, which supports the synchronized acquisition of audio, video, MIDI, and performance metadata. (ii) ASDF, which enables the efficient annotation of performer fingering from the visual data. Collectively, this system can streamline the acquisition of multimodal piano performance datasets.",
      "abstract": "Piano performance is a multimodal activity that intrinsically combines physical actions with the acoustic rendition. Despite growing research interest in analyzing the multimodal nature of piano performance, the laborious process of acquiring large-scale multimodal data remains a significant bottleneck, hindering further progress in this field. To overcome this barrier, we present an integrated web toolkit comprising two graphical user interfaces (GUIs): (i) PiaRec, which supports the synchronized acquisition of audio, video, MIDI, and performance metadata. (ii) ASDF, which enables the efficient annotation of performer fingering from the visual data. Collectively, this system can streamline the acquisition of multimodal piano performance datasets.",
      "authors": [
        "Junhyung Park",
        "Yonghyun Kim",
        "Joonhyung Bae",
        "Kirak Kim",
        "Taegyun Kwon",
        "Alexander Lerch",
        "Juhan Nam"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1zUcUNd6UHBJHCD_ZQ660UPzBLMjcPQQG/preview",
      "poster_link": "https://drive.google.com/file/d/1ZnjPkG6Cr6O9vu6zkvetKCHNvlpiib4I/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1oN18zXEdR-faj0SD3Vg2GjeVA1G26A5z&usp=drive_copy",
      "title": "Two Web Toolkits for Multimodal Piano Performance Dataset Acquisition and Fingering Annotation",
      "youtube_id": ""
    },
    "forum": "432",
    "id": "432",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We introduce Voices of Civilizations, the first multilingual QA benchmark for evaluating audio LLMs' cultural comprehension on full-length music recordings. Covering 380 tracks across 38 languages, our automated pipeline yields 1,190 multiple-choice questions through four stages\u2014each followed by manual verification: 1) compiling a representative music list; 2) generating cultural-background documents for each sample in the music list via LLMs; 3) extracting key attributes from those documents; and 4) constructing multiple\u2010choice questions probing language, region associations, mood, and thematic content. We evaluate models under four conditions and report per\u2010language accuracy. Our findings demonstrate that even state-of-the-art audio LLMs struggle to capture subtle cultural nuances without rich textual context and exhibit systematic biases in interpreting music from different cultural traditions. The dataset is publicly available on Hugging Face to foster culturally inclusive music understanding research.",
      "abstract": "We introduce Voices of Civilizations, the first multilingual QA benchmark for evaluating audio LLMs' cultural comprehension on full-length music recordings. Covering 380 tracks across 38 languages, our automated pipeline yields 1,190 multiple-choice questions through four stages\u2014each followed by manual verification: 1) compiling a representative music list; 2) generating cultural-background documents for each sample in the music list via LLMs; 3) extracting key attributes from those documents; and 4) constructing multiple\u2010choice questions probing language, region associations, mood, and thematic content. We evaluate models under four conditions and report per\u2010language accuracy. Our findings demonstrate that even state-of-the-art audio LLMs struggle to capture subtle cultural nuances without rich textual context and exhibit systematic biases in interpreting music from different cultural traditions. The dataset is publicly available on Hugging Face to foster culturally inclusive music understanding research.",
      "authors": [
        "Shangda Wu",
        "Ziya Zhou",
        "Yongyi Zang",
        "Yutong Zheng",
        "Dafang Liang",
        "Ruibin Yuan",
        "Qiuqiang Kong"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Dataset"
      ],
      "paper_link": "https://drive.google.com/file/d/1Z_bh4dTd1qACtF1ROigTyk6ybdLspBrA/preview",
      "poster_link": "https://drive.google.com/file/d/1QEfk6xf1lun2l-yQjLmYbFoGIOVdRfuH/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1_qhlNFwXmcM3XQXeHVVuA4m1ZAgLemWH&usp=drive_copy",
      "title": "Voices of Civilizations: A Multilingual QA Benchmark for Global Music Understanding",
      "youtube_id": ""
    },
    "forum": "433",
    "id": "433",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We present STOMP, a compendium of loss functions for self-supervised learning in music information retrieval with applications to beat and tempo estimation. The key idea is to penalize weights of a sequence-to-sequence encoder so that its output approximates a pulse train with undetermined period and phase. STOMP is effective and simpler than prior work: it is end-to-end, nonconstrative, and amenable to both single-track and multitrack settings.",
      "abstract": "We present STOMP, a compendium of loss functions for self-supervised learning in music information retrieval with applications to beat and tempo estimation. The key idea is to penalize weights of a sequence-to-sequence encoder so that its output approximates a pulse train with undetermined period and phase. STOMP is effective and simpler than prior work: it is end-to-end, nonconstrative, and amenable to both single-track and multitrack settings.",
      "authors": [
        "Vincent Lostanlen",
        "Yuexuan Kong",
        "Gabriel Meseguer-Brocal",
        "Mathieru Lagrange",
        "Romain Hennequin"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1fr13Lbg2p_gu7dLbN4W3ldovVANHCOFs/preview",
      "poster_link": "https://drive.google.com/file/d/1_vx00_HP6yV11FIw7yP0NPNthzCF7amm/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1Q8ooeOgXyPy9xT6UqmPPaJmU9Hk6rKpl&usp=drive_copy",
      "title": "STOMP! Self-supervised beat induction by matching pulses",
      "youtube_id": ""
    },
    "forum": "434",
    "id": "434",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We introduce ET-MusicGen, a proof-of-concept framework for generating music directly from brain activity. Our method discretizes EEG signals from the DEAP dataset into tokens using a self-supervised encoder. These EEG tokens then condition the hierarchical prior of OpenAI\u2019s Jukebox model to generate new audio. Limited by GPU resources, our model produces 5-second clips, but achieving a promising Fr\u00e9chet Audio Distance (FAD) of 1.24 and a CLAP score of 0.55. This indicates strong audio quality and semantic alignment comparable to recent diffusion-based models. Futhermore this work establishes discrete EEG tokenization as a viable pathway toward future end-to-end, brain-controlled music generation.",
      "abstract": "We introduce ET-MusicGen, a proof-of-concept framework for generating music directly from brain activity. Our method discretizes EEG signals from the DEAP dataset into tokens using a self-supervised encoder. These EEG tokens then condition the hierarchical prior of OpenAI\u2019s Jukebox model to generate new audio. Limited by GPU resources, our model produces 5-second clips, but achieving a promising Fr\u00e9chet Audio Distance (FAD) of 1.24 and a CLAP score of 0.55. This indicates strong audio quality and semantic alignment comparable to recent diffusion-based models. Futhermore this work establishes discrete EEG tokenization as a viable pathway toward future end-to-end, brain-controlled music generation.",
      "authors": [
        "Seunghyun Cho"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/18T-olFA8UrCVGVVeEyI7nN6WzOSIQWZO/preview",
      "poster_link": "https://drive.google.com/file/d/16tSnnuNboi9THMVcaqtJh_xt_C3TgLbK/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1hkH1_NuLhn80u1aiFkRXFDv5W5S-O0jL&usp=drive_copy",
      "title": "ET-MUSICGEN: EEG TOKEN BASED MUSIC GENERATION",
      "youtube_id": "https://drive.google.com/file/d/1yo33o8H5ryMqfr9GclYaguLifCr8xtyU/preview"
    },
    "forum": "435",
    "id": "435",
    "position": ""
  },
  {
    "content": {
      "TLDR": "K-pop\u2019s global success is fueled by its dynamic performances and vibrant fan engagement. Inspired by K-pop fan culture, we propose a methodology for automatically extracting live vocals from performances. We use a combination of source separation, cross-correlation, and amplitude scaling to automatically remove pre-recorded vocals and instrumentals from a live performance. Our preliminary work introduces the task of live vocal separation and provides a foundation for future research in this topic.",
      "abstract": "K-pop\u2019s global success is fueled by its dynamic performances and vibrant fan engagement. Inspired by K-pop fan culture, we propose a methodology for automatically extracting live vocals from performances. We use a combination of source separation, cross-correlation, and amplitude scaling to automatically remove pre-recorded vocals and instrumentals from a live performance. Our preliminary work introduces the task of live vocal separation and provides a foundation for future research in this topic.",
      "authors": [
        "Yujin Kim",
        "Richa Namballa",
        "Magdalena Fuentes"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1e2UsqPceTameI7Xt1B_Yd6qi0ZMEwWoC/preview",
      "poster_link": "https://drive.google.com/file/d/1mWbRfzZdsfMLKaPa9T_wz1pIG7hquQoj/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=15UQLVyx8FsZoi_SaHkfLxNNVuAezcEIz&usp=drive_copy",
      "title": "Live Vocal Extraction from K-pop Performances",
      "youtube_id": ""
    },
    "forum": "436",
    "id": "436",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Music is traditionally analyzed at the level of discrete symbols or channels. However, it is perceived in space by listeners with diverse physiological and cognitive audio processing mechanisms. We present early results from a study quantifying listeners\u2019 ability to localize sounds in near-field virtual space. Our data help bridge gaps between symbolic, acoustic, and cognitive understandings of musical information.",
      "abstract": "Music is traditionally analyzed at the level of discrete symbols or channels. However, it is perceived in space by listeners with diverse physiological and cognitive audio processing mechanisms. We present early results from a study quantifying listeners\u2019 ability to localize sounds in near-field virtual space. Our data help bridge gaps between symbolic, acoustic, and cognitive understandings of musical information.",
      "authors": [
        "Michael Gancz",
        "Justin Berry",
        "Skylar Bartush",
        "Jake Shaker",
        "Kimberly Hieftje",
        "Asher Marks",
        "Veronica Weser"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1IQWX-QTLj_yle3jk9-IQis9nwqg3s6oR/preview",
      "poster_link": "https://drive.google.com/file/d/1PFHJ-bB1OZxCx_RpGLWhwi42jKk7tX1C/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=19c_tGm9TLjRNW4rGXvEpqDy7SF9kGRNr&usp=drive_copy",
      "title": "Spatial Perception of HRTF-Matched Binaural Audio",
      "youtube_id": ""
    },
    "forum": "437",
    "id": "437",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Differentiable digital signal processing (DDSP) incorporates signal processing modules into neural networks to enable gradient-based learning. The DDSP mixture model (DDSP-MM) builds on this idea to represent an audio mixture as the sum of source signals generated by multiple differentiable synthesizers. Each source is independently controlled and may represent a distinct musical line. By estimating control parameters to match the input audio, DDSP-MM allows flexible editing and re-synthesis of individual sources. However, it requires time-aligned score information for initialization, which is often unavailable in practice. In this paper, we address this limitation by applying dynamic time warping to align unaligned score data with the mixture in a preprocessing step. We further provide a music audio editing system to demonstrate the workflow, enabling source-level editing and re-synthesis.",
      "abstract": "Differentiable digital signal processing (DDSP) incorporates signal processing modules into neural networks to enable gradient-based learning. The DDSP mixture model (DDSP-MM) builds on this idea to represent an audio mixture as the sum of source signals generated by multiple differentiable synthesizers. Each source is independently controlled and may represent a distinct musical line. By estimating control parameters to match the input audio, DDSP-MM allows flexible editing and re-synthesis of individual sources. However, it requires time-aligned score information for initialization, which is often unavailable in practice. In this paper, we address this limitation by applying dynamic time warping to align unaligned score data with the mixture in a preprocessing step. We further provide a music audio editing system to demonstrate the workflow, enabling source-level editing and re-synthesis.",
      "authors": [
        "Kengo Takemoto",
        "Tomohiko Nakamura",
        "Hiroshi Saruwatari"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1ZQ7k1o2RdnNBD4d-cAHNuqQqi5o29u-1/preview",
      "poster_link": "https://drive.google.com/file/d/1iL8EGhHck8LjHVHZugRUbjoetNHMfyny/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=10gcHE2nkgxJPFOJPPtOykA8PwVCg_otO&usp=drive_copy",
      "title": "Toward Score-Informed Music Audio Editing System using Differentiable Digital Signal Processing Mixture Model",
      "youtube_id": ""
    },
    "forum": "438",
    "id": "438",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We present a voice editing tool based on our neural analysis synthesis (NANSY) framework that allows modification of pitch parameters such as F0, subharmonics, jitter, and shimmer. The NANSY framework splits the voice into pitch parameters, linguistic-, and speaker embeddings. In this work we propose an extension NANSY's pitch encoder to include subharmonic- and structural noise parameters by predicting a joint note and phonation-mode distribution Furthermore, we extend the self-supervised training procedure to include supervision on synthetic on-the-fly generated data and integrate Viterbi smoothing into the training process. We provide a web-based UI that allows editing the voice on a computer or tablet by redrawing the control parameter curves.",
      "abstract": "We present a voice editing tool based on our neural analysis synthesis (NANSY) framework that allows modification of pitch parameters such as F0, subharmonics, jitter, and shimmer. The NANSY framework splits the voice into pitch parameters, linguistic-, and speaker embeddings. In this work we propose an extension NANSY's pitch encoder to include subharmonic- and structural noise parameters by predicting a joint note and phonation-mode distribution Furthermore, we extend the self-supervised training procedure to include supervision on synthetic on-the-fly generated data and integrate Viterbi smoothing into the training process. We provide a web-based UI that allows editing the voice on a computer or tablet by redrawing the control parameter curves.",
      "authors": [
        "Frederik Bous",
        "Jacob Morton",
        "Joon Byun",
        "Yechan Yu",
        "Jinhyeok Yang",
        "Hyeongju Kim",
        "Juheon Lee"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/16nmL_VZglaa70fs365Z0RIp7xNbzugHB/preview",
      "poster_link": "https://drive.google.com/file/d/1uuN2mRYkKsZpmytf5pK7pFt0hCkIgndu/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1-bI65FS5osBFjr98FX9gHfdBj9r8iOD6&usp=drive_copy",
      "title": "A neural voice editing tool for transforming pitch, subharmonics, and structural noise",
      "youtube_id": "https://drive.google.com/file/d/1NSiNdg2WGWrEliYMVqV5-eXYh0lBIpyR/preview"
    },
    "forum": "439",
    "id": "439",
    "position": ""
  },
  {
    "content": {
      "TLDR": "A recent focus in the development of music recommendation systems is the incorporation of physiological signals. Among this, the possibility of using non-invasive, electroencephalography-based neural activity is of great interest. In this preliminary work, we sought to predict the preference of individuals for previously unheard music through a combination of acoustic and neural features. We developed a real-time system for preference decoding which was used to skip songs with \u223c80 ms latency according to users\u2019 desires. The results suggest that music recommendation systems could supplement acoustic features with neural activity for characterising an individual\u2019s music preferences in real time, with options to incorporate further acoustic and physiological information for improved",
      "abstract": "A recent focus in the development of music recommendation systems is the incorporation of physiological signals. Among this, the possibility of using non-invasive, electroencephalography-based neural activity is of great interest. In this preliminary work, we sought to predict the preference of individuals for previously unheard music through a combination of acoustic and neural features. We developed a real-time system for preference decoding which was used to skip songs with \u223c80 ms latency according to users\u2019 desires. The results suggest that music recommendation systems could supplement acoustic features with neural activity for characterising an individual\u2019s music preferences in real time, with options to incorporate further acoustic and physiological information for improved",
      "authors": [
        "Thomas Binns",
        "Shinichi Furuya",
        "Vincent Cheung"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/19BikU99IWIaDbmxVPXS7oq5nPrB-JDzH/preview",
      "poster_link": "https://drive.google.com/file/d/1Bl6xHgUzEKznHxJKtEBpZzQLlSw_9WW_/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1S7B3A1QTyXsbIu5cPW732LDl3BSYtNnV&usp=drive_copy",
      "title": "A real-time multimodal system for music preference decoding combining EEG and acoustic features",
      "youtube_id": ""
    },
    "forum": "440",
    "id": "440",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Automatic music transcription (AMT) has achieved remarkable progress for instruments such as the piano, largely due to the availability of large-scale, high-quality datasets. In contrast, violin AMT remains underexplored due to limited annotated data. A common approach is to fine-tune pretrained models for other downstream tasks, but the effectiveness of such transfer remains unclear in the presence of timbral and articulatory differences. In this work, we investigate whether training from scratch on a medium-scale violin dataset can match the performance of fine-tuned piano-pretrained models. We adopt a piano transcription architecture without modification and train it on the MOSA dataset, which contains approximately 30 hours of aligned violin recordings. Our experiments on URMP and Bach10 show that models trained from scratch achieved competitive or even superior performance compared to fine-tuned counterparts. These findings suggest that strong violin AMT is possible without relying on pretrained piano representations, highlighting the importance of instrument-specific data collection and augmentation strategies.",
      "abstract": "Automatic music transcription (AMT) has achieved remarkable progress for instruments such as the piano, largely due to the availability of large-scale, high-quality datasets. In contrast, violin AMT remains underexplored due to limited annotated data. A common approach is to fine-tune pretrained models for other downstream tasks, but the effectiveness of such transfer remains unclear in the presence of timbral and articulatory differences. In this work, we investigate whether training from scratch on a medium-scale violin dataset can match the performance of fine-tuned piano-pretrained models. We adopt a piano transcription architecture without modification and train it on the MOSA dataset, which contains approximately 30 hours of aligned violin recordings. Our experiments on URMP and Bach10 show that models trained from scratch achieved competitive or even superior performance compared to fine-tuned counterparts. These findings suggest that strong violin AMT is possible without relying on pretrained piano representations, highlighting the importance of instrument-specific data collection and augmentation strategies.",
      "authors": [
        "Ting-Kang Wang",
        "Yueh-Po Peng",
        "Li Su",
        "VincentK.M. Cheung"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/12c7ZO9Px_3qS0y9HqOxYkjUZ4VDF5_VQ/preview",
      "poster_link": "https://drive.google.com/file/d/1fhCcswA4eN-FTeb7WplW7Ds_sVVCTJHR/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1r88oMRkvLDO--v_mzj7X0cJuYP-AwnLI&usp=drive_copy",
      "title": "Is Transfer Learning Necessary for Violin Transcription?",
      "youtube_id": ""
    },
    "forum": "442",
    "id": "442",
    "position": ""
  },
  {
    "content": {
      "TLDR": "This paper evaluates the extent to which expertise in prompt construction influences the quality of the music generation output. We propose a Retrieval-Augmented Prompt Rewrite system (RAG) that transforms novice prompts into expert descriptions using CLAP. Our method helps preserve user intent and bypass the need for extensive domain training of the user. Given novice-level prompts, participants selected relevant terminologies from top-k most textually or audibly similar MusicCaps captions, which were fed into GPT 3.5 to create succinct, expert-level rewrites. These rewrites were then used to generate music with Stable Audio 2.0. To mitigate anchoring bias toward expert prompts, we implemented a counterbalanced design and conducted a subjective study to evaluate the effectiveness of RAG. We generated rewrites using a traditional LoRA fine-tuning method as our baseline. Participants evaluated the expertness, musicality, production quality and preference of music generated from novice and expert prompts. Both RAG and LoRA rewrites significantly improve music generation across all NLP and subjective metrics, with RAG outperforming LoRA overall. Finally, the subjective results largely align with Meta\u2019s Audiobox Aesthetics metrics.",
      "abstract": "This paper evaluates the extent to which expertise in prompt construction influences the quality of the music generation output. We propose a Retrieval-Augmented Prompt Rewrite system (RAG) that transforms novice prompts into expert descriptions using CLAP. Our method helps preserve user intent and bypass the need for extensive domain training of the user. Given novice-level prompts, participants selected relevant terminologies from top-k most textually or audibly similar MusicCaps captions, which were fed into GPT 3.5 to create succinct, expert-level rewrites. These rewrites were then used to generate music with Stable Audio 2.0. To mitigate anchoring bias toward expert prompts, we implemented a counterbalanced design and conducted a subjective study to evaluate the effectiveness of RAG. We generated rewrites using a traditional LoRA fine-tuning method as our baseline. Participants evaluated the expertness, musicality, production quality and preference of music generated from novice and expert prompts. Both RAG and LoRA rewrites significantly improve music generation across all NLP and subjective metrics, with RAG outperforming LoRA overall. Finally, the subjective results largely align with Meta\u2019s Audiobox Aesthetics metrics.",
      "authors": [
        "Meiying Ding",
        "Brian McFee",
        "Chenkai Hu",
        "Sunny Yang",
        "Juhua Huang"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1hb0nfV-scmFy-5_BMTIsOE3Yi8tq_MfR/preview",
      "poster_link": "https://drive.google.com/file/d/1ssVpDnY7hKASqkZa7LercVQKtzQWNY7A/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1Ujt83aTmBFQcgdPHnXDq-2uHoNS3XRq0&usp=drive_copy",
      "title": "ENHANCING TEXT-TO-MUSIC GENERATION THROUGH RETRIEVAL-AUGMENTED PROMPT REWRITE LBD",
      "youtube_id": ""
    },
    "forum": "443",
    "id": "443",
    "position": ""
  },
  {
    "content": {
      "TLDR": "The musical experience does not arise from the sound itself but from how the listener interprets it. This perspective has inspired artistic practices that recontextualize non-musical sounds as a piece of art. Motivated by this view, we define a novel task, General Audio to Symbolic Music Transformation (GASMT), which aims to generate symbolic music with both high input accordance and musical quality from arbitrary audio. We also present \\textit{Sori}, a real-time system for GASMT, built on a streaming encoder-decoder architecture with domain adversarial training and classifier-free guidance that controls the trade-off between accordance and musicality. Audio demos are available at: https://tinyurl.com/sori-demo",
      "abstract": "The musical experience does not arise from the sound itself but from how the listener interprets it. This perspective has inspired artistic practices that recontextualize non-musical sounds as a piece of art. Motivated by this view, we define a novel task, General Audio to Symbolic Music Transformation (GASMT), which aims to generate symbolic music with both high input accordance and musical quality from arbitrary audio. We also present \\textit{Sori}, a real-time system for GASMT, built on a streaming encoder-decoder architecture with domain adversarial training and classifier-free guidance that controls the trade-off between accordance and musicality. Audio demos are available at: https://tinyurl.com/sori-demo",
      "authors": [
        "Kyungsu Kim",
        "Yejin Kim",
        "Kyogu Lee"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1CjVcHI7LNJ3ojBrodD_UNo2OlCsXd6Ew/preview",
      "poster_link": "https://drive.google.com/file/d/1giuJIjjpvCwRuFOtU1W382wVOnN38GFV/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1EFPVTl7yFWIQD0RpNVaipmi9hkx1XmuK&usp=drive_copy",
      "title": "Sori: Real-time General Audio to MIDI Transformation with Accordance-Musicality Trade-off Controllability",
      "youtube_id": "https://drive.google.com/file/d/1t1B1VHGTkmA8WUoKWqcBJH_a76rC3Cca/preview"
    },
    "forum": "444",
    "id": "444",
    "position": ""
  },
  {
    "content": {
      "TLDR": "MIDI serves as an important symbolic representation of performance and sheet music. With the availability of millions of MIDI files in public datasets, content-based analysis of MIDI files has become increasingly important. In this extended abstract, we present a self-supervised approach for chord and key estimation in score MIDI files. The core idea is extremely simple: We locate MIDI tracks that look useful to the task and create noisy psuedo-labels, which are used to fine-tune a MIDI foundation model. Our method achieves state-of-the-art performance entirely without human annotation or MIDI metadata.",
      "abstract": "MIDI serves as an important symbolic representation of performance and sheet music. With the availability of millions of MIDI files in public datasets, content-based analysis of MIDI files has become increasingly important. In this extended abstract, we present a self-supervised approach for chord and key estimation in score MIDI files. The core idea is extremely simple: We locate MIDI tracks that look useful to the task and create noisy psuedo-labels, which are used to fine-tune a MIDI foundation model. Our method achieves state-of-the-art performance entirely without human annotation or MIDI metadata.",
      "authors": [
        "Junyan Jiang",
        "Bilei Zhu",
        "Gus Xia"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1g2edBP-VAsJXLR_dDgNKND_NizI9bN7K/preview",
      "poster_link": "https://drive.google.com/file/d/1nXin85YEygzq8lArYEKMuRvw5M9PIJe4/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1fZth7-KzFiE6Led25z0WXhiombQzygkF&usp=drive_copy",
      "title": "The Music Labels Itself: Content-based MIDI Chord and Key Estimation Without Human Labels",
      "youtube_id": "https://drive.google.com/file/d/1OeoC73Ph4Dre4yi7-aKdLOuRBA9f7pz9/preview"
    },
    "forum": "445",
    "id": "445",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Automatic Music Transcription (AMT) systems produce MIDI files of varying quality for audio samples with different acoustic conditions. Without filtering or postprocessing, transcription errors can propagate into downstream tasks. In this work, we present a data-driven approach to assessing the quality of piano MIDI. We analyze note alignment between MIDI performances and musical scores in ASAP and ATEPP datasets and show that alignment quality is a good proxy for music quality. Based on this insight, we label performances in ATEPP dataset and train a sequence classifier to categorize MIDI files into five classes: musical score, recorded, high quality, low quality, and corrupted. The final model demonstrates a perceptually meaningful understanding of music quality, with errors constrained to adjacent classes. This approach can be used to create cleaner datasets of transcribed music.",
      "abstract": "Automatic Music Transcription (AMT) systems produce MIDI files of varying quality for audio samples with different acoustic conditions. Without filtering or postprocessing, transcription errors can propagate into downstream tasks. In this work, we present a data-driven approach to assessing the quality of piano MIDI. We analyze note alignment between MIDI performances and musical scores in ASAP and ATEPP datasets and show that alignment quality is a good proxy for music quality. Based on this insight, we label performances in ATEPP dataset and train a sequence classifier to categorize MIDI files into five classes: musical score, recorded, high quality, low quality, and corrupted. The final model demonstrates a perceptually meaningful understanding of music quality, with errors constrained to adjacent classes. This approach can be used to create cleaner datasets of transcribed music.",
      "authors": [
        "Ilya Borovik"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1f0Co9LDWKmeyxPpmr1ZWkfrRDodTrid0/preview",
      "poster_link": "https://drive.google.com/file/d/10HsAtQ0j2N4rulB_T_RDwcv3CEtwLL1O/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1x7DuaWqHt6WpNgCHgMRhG6P131_opK4h&usp=drive_copy",
      "title": "Automatic Quality Assessment of Transcribed Piano MIDI",
      "youtube_id": ""
    },
    "forum": "446",
    "id": "446",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Abstract instructions in piano education, such as \"raise your wrist\" and \"relax your tension,\" lead to varying interpretations among learners, preventing instructors from effectively conveying their intended pedagogical guidance. To address this problem, this study conducted systematic interviews with a piano professor with 18 years teaching experience, and two researchers derived seven core need groups through cross-validation. Based on these findings, we developed a web-based dashboard prototype integrating video, motion capture, and musical scores, enabling instructors to provide concrete, visual feedback instead of relying solely on abstract verbal instructions. Technical feasibility was validated through 109 performance datasets.",
      "abstract": "Abstract instructions in piano education, such as \"raise your wrist\" and \"relax your tension,\" lead to varying interpretations among learners, preventing instructors from effectively conveying their intended pedagogical guidance. To address this problem, this study conducted systematic interviews with a piano professor with 18 years teaching experience, and two researchers derived seven core need groups through cross-validation. Based on these findings, we developed a web-based dashboard prototype integrating video, motion capture, and musical scores, enabling instructors to provide concrete, visual feedback instead of relying solely on abstract verbal instructions. Technical feasibility was validated through 109 performance datasets.",
      "authors": [
        "Joonhyung Bae",
        "Hyeyoon Cho",
        "Kirak Kim",
        "Dawon Park",
        "Taegyun Kwon",
        "Yoon-Seok Choi",
        "Hyeon Hur",
        "Shigeru Kai",
        "Yohei Wada",
        "Satoshi Obata",
        "Akira Maezawa",
        "Jaebum Park",
        "Jonghwa Park",
        "Juhan Nam"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1ZvS2exo4Tc_53g_Bdo7ykzhk5C_g-_yH/preview",
      "poster_link": "https://drive.google.com/file/d/18HUHIFLwom8siDH_f2Bi-1wG8hTSKOZH/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=13Z94pAj0oyqH5gNole9_Wttu_XEE1Fdr&usp=drive_copy",
      "title": "Designing a Multimodal Viewer for Piano Performance Analysis - a Pedagogy-First Approach",
      "youtube_id": ""
    },
    "forum": "447",
    "id": "447",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Despite growing interest in video-to-music generation systems, their application in film production remains limited, primarily due to the lack of large-scale datasets containing aligned pairs of movie clips and soundtracks. Although prior work has attempted to construct such a dataset, this comprises only 36.5 hours of data, which is insufficient for training robust models. In this study, we present Open Screen Soundtrack Library Version 2, a novel dataset comprising pairs of video clips from films and their corresponding soundtracks, curated with a novel methodology that automatically identifies and extracts soundtrack segments from video clips. This dataset consists of 552.70 hours and 76,408 video clips sourced from both public domain movies as well as commercial ones from a publicly available dataset.",
      "abstract": "Despite growing interest in video-to-music generation systems, their application in film production remains limited, primarily due to the lack of large-scale datasets containing aligned pairs of movie clips and soundtracks. Although prior work has attempted to construct such a dataset, this comprises only 36.5 hours of data, which is insufficient for training robust models. In this study, we present Open Screen Soundtrack Library Version 2, a novel dataset comprising pairs of video clips from films and their corresponding soundtracks, curated with a novel methodology that automatically identifies and extracts soundtrack segments from video clips. This dataset consists of 552.70 hours and 76,408 video clips sourced from both public domain movies as well as commercial ones from a publicly available dataset.",
      "authors": [
        "Haven Kim",
        "Leduo Chen",
        "Bill Wang",
        "Hao-Wen Dong",
        "Julian McAuley"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Dataset"
      ],
      "paper_link": "https://drive.google.com/file/d/1UNJqK7hApHCimies1lSKu52V9vmQwgzj/preview",
      "poster_link": "https://drive.google.com/file/d/1AKvL3Ic1i3qNcs2r3FSQQKpp8t9nEF_z/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1bcqrlMkQpIyOBfnDkqbixP3GPL9vW74H&usp=drive_copy",
      "title": "Open Screen Soundtrack Library Version 2",
      "youtube_id": ""
    },
    "forum": "448",
    "id": "448",
    "position": ""
  },
  {
    "content": {
      "TLDR": "This paper proposes a method for altering the meter of melodies from the 4/4 meter to another irregular meter (specifically 5/4). Melody transformation is studied in the context of melodic style transfer, but transformation of the meter of melodies have not bean dealt with. To incorporate the difficulty that a dataset of melodies in the target (5/4) meter cannot be available, we propose a method based on a notewise Markov model, which can be trained only with melodies in the 4/4 meter. Preliminary experimental results implied that melodies generated with our model has close melodic features to existing melodies.",
      "abstract": "This paper proposes a method for altering the meter of melodies from the 4/4 meter to another irregular meter (specifically 5/4). Melody transformation is studied in the context of melodic style transfer, but transformation of the meter of melodies have not bean dealt with. To incorporate the difficulty that a dataset of melodies in the target (5/4) meter cannot be available, we propose a method based on a notewise Markov model, which can be trained only with melodies in the 4/4 meter. Preliminary experimental results implied that melodies generated with our model has close melodic features to existing melodies.",
      "authors": [
        "Sota Nagi"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1BxvInEqprEVObl2wU0JIW5_0YO7NtrnN/preview",
      "poster_link": "https://drive.google.com/file/d/1FHdJpAM007L6cVZazTaMAeJ924gbIlyv/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1xHgOrkVdNTmrb72ck_4De1qgCxejuXyf&usp=drive_copy",
      "title": "Meter Transformation of Melody Using Markov Model",
      "youtube_id": ""
    },
    "forum": "449",
    "id": "449",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We demo Peachnote Piano Tutor, a prototype that listens to a student's playing, aligns the stream to a reference score in real-time, and delivers teacher\u2011style feedback. While commercial tutors often focus on note accuracy, our system addresses expressive nuance using a multimodal LLM. By concurrently analyzing the performance as it happens, the tutor responds immediately after the student pauses, combining spoken advice with synchronized visual highlights on the score. Although the prototype accepts both microphone audio and MIDI, the live demo employs a MIDI keyboard to guarantee clean note detection in a noisy exhibition space. The lightweight client\u2011server\u2011services stack built on WebRTC transport enables this responsive interaction.",
      "abstract": "We demo Peachnote Piano Tutor, a prototype that listens to a student's playing, aligns the stream to a reference score in real-time, and delivers teacher\u2011style feedback. While commercial tutors often focus on note accuracy, our system addresses expressive nuance using a multimodal LLM. By concurrently analyzing the performance as it happens, the tutor responds immediately after the student pauses, combining spoken advice with synchronized visual highlights on the score. Although the prototype accepts both microphone audio and MIDI, the live demo employs a MIDI keyboard to guarantee clean note detection in a noisy exhibition space. The lightweight client\u2011server\u2011services stack built on WebRTC transport enables this responsive interaction.",
      "authors": [
        "Vladimir Viro"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "",
      "poster_link": "",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "",
      "title": "From Notes to Nuance: A Conversational AI for Affective Piano Performance",
      "youtube_id": ""
    },
    "forum": "451",
    "id": "451",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Computational analysis of folk song audio is challenging due to structural irregularities and the need for manual annotation. We propose a method for automatic motive segmentation in Korean folk songs by fine-tuning a speech transcription model on audio lyric with motif boundary annotation Applying this to 856 songs, we extracted motif count and duration entropy as structural features. Statistical analysis revealed that these features vary systematically according to the songs\u2019 social functions. Songs associated with collective labor, for instance, showed different structural patterns from those for entertainment or personal settings. This work offers a scalable approach for quantitative structural analysis of oral music traditions.",
      "abstract": "Computational analysis of folk song audio is challenging due to structural irregularities and the need for manual annotation. We propose a method for automatic motive segmentation in Korean folk songs by fine-tuning a speech transcription model on audio lyric with motif boundary annotation Applying this to 856 songs, we extracted motif count and duration entropy as structural features. Statistical analysis revealed that these features vary systematically according to the songs\u2019 social functions. Songs associated with collective labor, for instance, showed different structural patterns from those for entertainment or personal settings. This work offers a scalable approach for quantitative structural analysis of oral music traditions.",
      "authors": [
        "danbinaerin Han",
        "Dasaem Jeong",
        "Juhan Nam"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1yqvgyyJ4cTRYpj4eLetsUQgpehdDoA-k/preview",
      "poster_link": "https://drive.google.com/file/d/1kTYoXDYPVxPS1OTh_3kQ09g_ukx1Bc4x/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1OMgLI4jL8n1dtY9WrE733Cq_H8YvZQLK&usp=drive_copy",
      "title": "Motive-level Analysis of Form-functions Association in Korean Folk Song",
      "youtube_id": ""
    },
    "forum": "452",
    "id": "452",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We present a real-time musical interaction system for two AI agents using a reactive synchronization model. Extending human\u2013AI ensemble work, the architecture enables bidirectional interaction in which each agent adjusts tempi to incoming input without knowing whether it originates from a human or another machine. Experiments vary initial tempi and reactivities and record intended and actual",
      "abstract": "We present a real-time musical interaction system for two AI agents using a reactive synchronization model. Extending human\u2013AI ensemble work, the architecture enables bidirectional interaction in which each agent adjusts tempi to incoming input without knowing whether it originates from a human or another machine. Experiments vary initial tempi and reactivities and record intended and actual",
      "authors": [
        "Kit Armstrong",
        "Tzu-Chi Lin",
        "Nien-Heng Yu",
        "Yi-Wen Liu"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1Pc-2CxfBK8OCy1YAKzqkURU6SmLoYWj-/preview",
      "poster_link": "https://drive.google.com/file/d/1CxDCkQSEfYYbDPB9HzhAU7LiL37fYtNw/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1RwkCWkj1cVdyi4Ur0bPmc_FE4h8rlJQ8&usp=drive_copy",
      "title": "Real-time piano accompaniment system performing with itself",
      "youtube_id": "https://drive.google.com/file/d/1Dnn7yb2zAvO9awAFsdpPBaO1b1f7RRS3/preview"
    },
    "forum": "453",
    "id": "453",
    "position": ""
  },
  {
    "content": {
      "TLDR": "This paper presents an optical music recognition (OMR) system tailored for PDF scores of string quartet pieces. The approach begins with a SegNet-based segmentation model to localize musical regions, followed by convolutional neural networks and image processing techniques to classify and extract musical symbols. To improve rhythmic precision, dynamic programming methods enforce time signature constraints and explicitly handle tuplets. Evaluation using a longest common subsequence metric across multiple feature matrices demonstrates robust accuracy in recognizing pitch, onset, duration, and accidentals on a benchmark dataset. Code and data are available at \\url{https://github.com/lattellie/25-omr}.",
      "abstract": "This paper presents an optical music recognition (OMR) system tailored for PDF scores of string quartet pieces. The approach begins with a SegNet-based segmentation model to localize musical regions, followed by convolutional neural networks and image processing techniques to classify and extract musical symbols. To improve rhythmic precision, dynamic programming methods enforce time signature constraints and explicitly handle tuplets. Evaluation using a longest common subsequence metric across multiple feature matrices demonstrates robust accuracy in recognizing pitch, onset, duration, and accidentals on a benchmark dataset. Code and data are available at \\url{https://github.com/lattellie/25-omr}.",
      "authors": [
        "Jin-Shiang Hu",
        "Bo-Yuan Cheng",
        "Li Su"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1q1F8UJHSzsozwY6Sg79AfTQXqe4BVLcZ/preview",
      "poster_link": "https://drive.google.com/file/d/10oiRKlRWBGlCPVs9r_OhjhMKIptalwGE/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1XpUp5eIk1p8raLGNnd2d_KbmvhZgvxev&usp=drive_copy",
      "title": "An Optical Music Recognition Tool for String Quartets",
      "youtube_id": ""
    },
    "forum": "454",
    "id": "454",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Taiwanese opera (Kua-\u00e1-h\u00ec), a major form of local theatrical tradition, underwent extensive television adaptation notably by pioneers like I\u00fbnn L\u00afe-hua. These videos, while potentially valuable for in-depth studies of Taiwanese opera, often have low quality and require substantial manual effort during data preparation. To streamline this process, we developed an interactive system for realtime OCR correction and a two-step approach integrating OCR-driven segmentation with Speech and Music Activity Detection (SMAD) to efficiently identify vocal segments from archival episodes with high precision. The resulting dataset, consisting of vocal segments and corresponding lyrics, can potentially supports various MIR tasks such as lyrics identification and tune retrieval.",
      "abstract": "Taiwanese opera (Kua-\u00e1-h\u00ec), a major form of local theatrical tradition, underwent extensive television adaptation notably by pioneers like I\u00fbnn L\u00afe-hua. These videos, while potentially valuable for in-depth studies of Taiwanese opera, often have low quality and require substantial manual effort during data preparation. To streamline this process, we developed an interactive system for realtime OCR correction and a two-step approach integrating OCR-driven segmentation with Speech and Music Activity Detection (SMAD) to efficiently identify vocal segments from archival episodes with high precision. The resulting dataset, consisting of vocal segments and corresponding lyrics, can potentially supports various MIR tasks such as lyrics identification and tune retrieval.",
      "authors": [
        "Tzu-Hung Huang",
        "Yun-En Tsai",
        "Chih-Wei Wu",
        "I-Chieh Wei",
        "Li Su",
        "Yun-Ning Hung"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Dataset"
      ],
      "paper_link": "https://drive.google.com/file/d/1y4iiCOM33FWzVaWSrJO-EkPKYe-8aWZd/preview",
      "poster_link": "https://drive.google.com/file/d/1XGWquHoVCP8Rilg5_ABrAHAW3pJju9e5/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=15H9qjiprp_a50U6QvMrCm1bjQeU1RmF3&usp=drive_copy",
      "title": "Timed text extraction from Taiwanese kua-\u00e1-h\u00ec TV series",
      "youtube_id": "https://drive.google.com/file/d/1VaXahqoe9n1iFn8ZW9i1d4zoYSZYxI52/preview"
    },
    "forum": "455",
    "id": "455",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Research in drum classification often trends in two key directions; (1) audio based single instrument classification, or (2) automatic drum transcription which makes exploration of drum audio style compelling, especially in comparison to a CNN trained on the GMD (Groove Midi Dataset) with a pretrained transformer based model PaSST (Patchout Audio Spectrogram Transformer) with frozen general audio embeddings from AudioSet. This comparison reveals the ways in which general audio knowledge can affect drum audio style classification. Experiments with model depth, augmentation and padding show that PaSST with these frozen embeddings reduces performance in terms of accuracy but reveals robust feature representation distinct from the CNN.",
      "abstract": "Research in drum classification often trends in two key directions; (1) audio based single instrument classification, or (2) automatic drum transcription which makes exploration of drum audio style compelling, especially in comparison to a CNN trained on the GMD (Groove Midi Dataset) with a pretrained transformer based model PaSST (Patchout Audio Spectrogram Transformer) with frozen general audio embeddings from AudioSet. This comparison reveals the ways in which general audio knowledge can affect drum audio style classification. Experiments with model depth, augmentation and padding show that PaSST with these frozen embeddings reduces performance in terms of accuracy but reveals robust feature representation distinct from the CNN.",
      "authors": [
        "Trent Eriksen",
        "Edwin VanderHeide",
        "Robert Saunders"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/19K1iM9bkmBrGTfuiSojR7AB4hu2mGGsr/preview",
      "poster_link": "https://drive.google.com/file/d/1BGsAgPaXzDCYFlBS-JzAdAxn-aBWsxfV/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1hsh2iQzHR9YB0ZgQ7lSBloiywxzPpyav&usp=drive_copy",
      "title": "A comparative study of Transfer Learning for Drum audio Style Classification",
      "youtube_id": "https://drive.google.com/file/d/1XOqPDEdD9Ax51Wgmwt8bfO7bGgyKF_k8/preview"
    },
    "forum": "456",
    "id": "456",
    "position": ""
  },
  {
    "content": {
      "TLDR": "WaveRoll is an interactive JavaScript library that enables comparative visualization and synchronized playback of multiple MIDI piano rolls on a browser. It addresses a specific evaluation need in Automatic Music Transcription (AMT), contrasting multiple MIDI outputs produced from the same input. The library displays multiple MIDI tracks on a single, time-aligned grid with synchronized audio, allowing users to compare pitch and timing, identify missed or extra notes, and observe onset and offset differences, as well as section-level patterns. We expect that such comparisons would assist in model evaluation and error analysis, and help readers to understand the model behavior better. The open-source library is available at https://github.com/crescent-stdio/wave-roll",
      "abstract": "WaveRoll is an interactive JavaScript library that enables comparative visualization and synchronized playback of multiple MIDI piano rolls on a browser. It addresses a specific evaluation need in Automatic Music Transcription (AMT), contrasting multiple MIDI outputs produced from the same input. The library displays multiple MIDI tracks on a single, time-aligned grid with synchronized audio, allowing users to compare pitch and timing, identify missed or extra notes, and observe onset and offset differences, as well as section-level patterns. We expect that such comparisons would assist in model evaluation and error analysis, and help readers to understand the model behavior better. The open-source library is available at https://github.com/crescent-stdio/wave-roll",
      "authors": [
        "Hannah Park",
        "Dasaem Jeong"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/18xrhIzo86qOxYFbGGCchARWmqKCgfIk3/preview",
      "poster_link": "https://drive.google.com/file/d/1JoTVrvd0llvuLLvJkXGRGLQIQgQz2lno/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1NFSl6jATpswSIXSPag-q9r1bfVSmT8o5&usp=drive_copy",
      "title": "WaveRoll: JavaScript Library for Comparative MIDI Piano-Roll Visualization",
      "youtube_id": ""
    },
    "forum": "459",
    "id": "459",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We fit interpretable models, including Elastic Net and Generalized Additive Models (GAMs), to investigate which audio features align with human aesthetic ratings of music. Our analysis focuses on a subset of the SongEval dataset, which includes subjective annotations across five perceptual dimensions: overall coherence, memorability, vocal naturalness, structural clarity, and overall musicality. Using summary statistics of pitch, spectral, energy, and rhythmic features, we train the models on a balanced sample of songs with high and low scores. Both models consistently identify zero crossing rate and Root Mean Square (RMS) energy as dominant predictors, followed by onset envelope and other spectral attributes. These findings provide promising, although not definitive, insights into the perceptual dimensions of music aesthetics.",
      "abstract": "We fit interpretable models, including Elastic Net and Generalized Additive Models (GAMs), to investigate which audio features align with human aesthetic ratings of music. Our analysis focuses on a subset of the SongEval dataset, which includes subjective annotations across five perceptual dimensions: overall coherence, memorability, vocal naturalness, structural clarity, and overall musicality. Using summary statistics of pitch, spectral, energy, and rhythmic features, we train the models on a balanced sample of songs with high and low scores. Both models consistently identify zero crossing rate and Root Mean Square (RMS) energy as dominant predictors, followed by onset envelope and other spectral attributes. These findings provide promising, although not definitive, insights into the perceptual dimensions of music aesthetics.",
      "authors": [
        "Yiting Xia"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1rPuBlpKVjzOYOkgRauHUvsEDTvIEywyB/preview",
      "poster_link": "https://drive.google.com/file/d/1vuK0jd5kwLCJ4qmBtD2v3ikm5K7BzNfI/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1XfQ11zE9jrChCQo1Re2T9BVHnnE1piQR&usp=drive_copy",
      "title": "Interpretable Modeling of Music Aesthetics: Insights from Elastic Net and GAMs on the SongEval Dataset",
      "youtube_id": ""
    },
    "forum": "460",
    "id": "460",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Vocal percussion (VP) is a vocal technique that emulates drum sounds and plays a crucial rhythmic role in contemporary a cappella music. Despite its importance, synthesizing VP sounds remains difficult due to their noisy, non-linguistic characteristics, which conventional speech and singing voice synthesis methods fail to handle effectively. Our previous work framed VP sound synthesis as a timbre transfer task from drum to VP sounds, leveraging their functional correspondence. It also introduced an offline method using a variational autoencoder-based model called RAVE. In this paper, we propose a real-time drum-to-VP sound conversion system based on this offline method. The system processes input audio in chunks of 46 ms, enabling online operation. We demonstrate that the proposed system operates in real time on the central processing unit of a modern laptop computer.",
      "abstract": "Vocal percussion (VP) is a vocal technique that emulates drum sounds and plays a crucial rhythmic role in contemporary a cappella music. Despite its importance, synthesizing VP sounds remains difficult due to their noisy, non-linguistic characteristics, which conventional speech and singing voice synthesis methods fail to handle effectively. Our previous work framed VP sound synthesis as a timbre transfer task from drum to VP sounds, leveraging their functional correspondence. It also introduced an offline method using a variational autoencoder-based model called RAVE. In this paper, we propose a real-time drum-to-VP sound conversion system based on this offline method. The system processes input audio in chunks of 46 ms, enabling online operation. We demonstrate that the proposed system operates in real time on the central processing unit of a modern laptop computer.",
      "authors": [
        "Rinka Nobukawa",
        "Tomohiko Nakamura",
        "Shinnosuke Takamichi",
        "Hiroshi Saruwatari"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/17I0rzEnSt0atJ5t-tdnehonoVeYxE0kW/preview",
      "poster_link": "https://drive.google.com/file/d/1hFpof7a5OPje-UREZwoj60GpHJhSTMDf/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1rLjnAE-i3bJG7idK3gRpQyP9ZOpANUxx&usp=drive_copy",
      "title": "Real-Time Drum-to-Vocal Percussion Sound Conversion System",
      "youtube_id": ""
    },
    "forum": "461",
    "id": "461",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Pansori is a traditional Korean narrative musical form that combines sung passages with spoken narration (aniri) under explicit rhythmic cycles (jangdan). While recent text to-music models achieve strong results on mainstream genres, their behavior on underrepresented traditions remains largely unexplored. We adapt ACE-Step to pansori via LoRA-based fine-tuning and compare its adaptation to Korean pop as a mainstream baseline. We evaluate our results through quantitative analysis with Fr\u00e9chet Audio Distance (FAD) and qualitative assessment by an expert pansori per former. Our study highlights challenges and partial successes in adapting to pansori. Future work will extend text-to-music models toward cultural diversity and traditional music.",
      "abstract": "Pansori is a traditional Korean narrative musical form that combines sung passages with spoken narration (aniri) under explicit rhythmic cycles (jangdan). While recent text to-music models achieve strong results on mainstream genres, their behavior on underrepresented traditions remains largely unexplored. We adapt ACE-Step to pansori via LoRA-based fine-tuning and compare its adaptation to Korean pop as a mainstream baseline. We evaluate our results through quantitative analysis with Fr\u00e9chet Audio Distance (FAD) and qualitative assessment by an expert pansori per former. Our study highlights challenges and partial successes in adapting to pansori. Future work will extend text-to-music models toward cultural diversity and traditional music.",
      "authors": [
        "Seola Cho",
        "Minjun Kim",
        "Dasaem Jeong"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1plgonEcw3AT7oJK5wkolIflFdaCVVFbL/preview",
      "poster_link": "https://drive.google.com/file/d/1A2tNhQjc-nUlekLj6i48z6gIvxEoUoj1/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1xmgnbLIkpMlAeSy-vzl09RYx-deTz4sC&usp=drive_copy",
      "title": "Exploring Pansori Generation with ACE-Step",
      "youtube_id": "https://drive.google.com/file/d/1inB9KhliLghvo6HHH0sVS28DYAfxOZ75/preview"
    },
    "forum": "463",
    "id": "463",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We present a bidirectional framework, S2H2S, for modelling the relationship between piano performance audio and Hackkey signals\u2014high-resolution, continuous measurements of vertical key position captured during performance. Unlike symbolic representations such as MIDI, Hackkey encodes fine-grained piano gesture dynamics, including key press depth, velocity, and release timing, offering a physically grounded perspective on expressive piano technique. The framework comprises two core components: sound-to-Hackkey (S2H) transcription and Hackkey-to-sound (H2S) synthesis. For both tasks, we adapt and fine-tune state-of-the-art neural models to predict continuous Hackkey trajectories from audio input and to synthesise expressive audio directly from Hackkey signals, respectively. The system demonstrates promising results on recordings of Hanon exercises, achieving accurate gesture reconstruction and faithful audio synthesis, supporting new use cases in expressive performance analysis, music performance synthesis, and music education.",
      "abstract": "We present a bidirectional framework, S2H2S, for modelling the relationship between piano performance audio and Hackkey signals\u2014high-resolution, continuous measurements of vertical key position captured during performance. Unlike symbolic representations such as MIDI, Hackkey encodes fine-grained piano gesture dynamics, including key press depth, velocity, and release timing, offering a physically grounded perspective on expressive piano technique. The framework comprises two core components: sound-to-Hackkey (S2H) transcription and Hackkey-to-sound (H2S) synthesis. For both tasks, we adapt and fine-tune state-of-the-art neural models to predict continuous Hackkey trajectories from audio input and to synthesise expressive audio directly from Hackkey signals, respectively. The system demonstrates promising results on recordings of Hanon exercises, achieving accurate gesture reconstruction and faithful audio synthesis, supporting new use cases in expressive performance analysis, music performance synthesis, and music education.",
      "authors": [
        "Jingjing Tang",
        "Shinichi Furuya",
        "George Fazekas",
        "Vincent Cheung"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1SYrnRmLuMWEnXzB3qopW_QQnvJmdgGUy/preview",
      "poster_link": "https://drive.google.com/file/d/1x963LkKg-Uky7U40WIRmmPqyaBz3QGe7/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1ZafZCnMbc0N-sq84V1BP31jDm0l0R1GE&usp=drive_copy",
      "title": "S2H2S: Modelling Expressive Piano Performances via Continuous Piano Key Motion Signals",
      "youtube_id": ""
    },
    "forum": "464",
    "id": "464",
    "position": ""
  },
  {
    "content": {
      "TLDR": "This study develops a deep learning model to detect the use of iZotope\u2019s automated mastering plugin, Ozone, and analyzes the difference in its usage rate between amateur and professional artists. We constructed a dataset of 135 pairs of human-mastered and Ozone-mastered audio (totaling 553 minutes and 20 seconds) and designed a classification model using feature extraction from a Demucs encoder. After data preprocessing, which included 5-second chunking and silence removal, the model achieved an accuracy of 88.2 % and an F1 score of 0.88 on the test set. The model was then applied to EDM tracks from two platforms: SoundCloud to represent amateur artists and Spotify for professionals. For the amateur cohort, we established the criteria that tracks must be non-commercially released and have fewer than 100 likes on SoundCloud. The results showed that the Ozone usage rate among amateur artists was 20.8 % (16 out of 77 tracks), whereas for professional artists, it was 2.0 % (2 out of 98 tracks), revealing a nearly 10-fold difference. This suggests that while amateur artists tend to rely on accessible and cost-effective AI mastering tools, professional artists continue to prefer dedicated human mastering engineers. ",
      "abstract": "This study develops a deep learning model to detect the use of iZotope\u2019s automated mastering plugin, Ozone, and analyzes the difference in its usage rate between amateur and professional artists. We constructed a dataset of 135 pairs of human-mastered and Ozone-mastered audio (totaling 553 minutes and 20 seconds) and designed a classification model using feature extraction from a Demucs encoder. After data preprocessing, which included 5-second chunking and silence removal, the model achieved an accuracy of 88.2 % and an F1 score of 0.88 on the test set. The model was then applied to EDM tracks from two platforms: SoundCloud to represent amateur artists and Spotify for professionals. For the amateur cohort, we established the criteria that tracks must be non-commercially released and have fewer than 100 likes on SoundCloud. The results showed that the Ozone usage rate among amateur artists was 20.8 % (16 out of 77 tracks), whereas for professional artists, it was 2.0 % (2 out of 98 tracks), revealing a nearly 10-fold difference. This suggests that while amateur artists tend to rely on accessible and cost-effective AI mastering tools, professional artists continue to prefer dedicated human mastering engineers. ",
      "authors": [
        "Jaehyun On",
        "Chaeho Myung",
        "Dasaem Jeong"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1fTDV9kVdJYKEOGO2dmEEQUbGqHD_CRa3/preview",
      "poster_link": "https://drive.google.com/file/d/1xbCLcs2SOFzLwlACPyoThBOfNEIL4807/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=18WTApZ_JpiqrBftbgdUHE91ngG2u79c9&usp=drive_copy",
      "title": "AI MASTERING DETECTOR: ANALYZING THE IMPACT OF AI ON MUSIC POST-PRODUCTION",
      "youtube_id": ""
    },
    "forum": "465",
    "id": "465",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Streaming\u2010era music playlist curation demands tools beyond metadata and heuristics. Talaria is an exploratory, segment\u2010based audio retrieval system that lets curators search by sound\u2014humming, singing, referencing tracks, or supplying audio clips\u2014to uncover music sharing timbral, rhythmic, and emotive affinities. Co\u2010developed with professional curators, Talaria fosters creative agency, transparency, and interpretability via an interactive visual interface. By foregrounding how music feels and why it resonates, Talaria reframes retrieval as a generative, human\u2010centered process\u2014merging machine learning with curator insight for more intuitive, inspiring discovery.",
      "abstract": "Streaming\u2010era music playlist curation demands tools beyond metadata and heuristics. Talaria is an exploratory, segment\u2010based audio retrieval system that lets curators search by sound\u2014humming, singing, referencing tracks, or supplying audio clips\u2014to uncover music sharing timbral, rhythmic, and emotive affinities. Co\u2010developed with professional curators, Talaria fosters creative agency, transparency, and interpretability via an interactive visual interface. By foregrounding how music feels and why it resonates, Talaria reframes retrieval as a generative, human\u2010centered process\u2014merging machine learning with curator insight for more intuitive, inspiring discovery.",
      "authors": [
        "Jing-Kai Lou",
        "Shuen-Huei Guan",
        "Lu-Chieh Chen",
        "I-Kai Wang",
        "Tsai-Tsung Chen"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1VQUku8W97aYdwakL7xlrr_YGlLwnMPZ6/preview",
      "poster_link": "https://drive.google.com/file/d/1_LDfD88mcnYl7CksyNrhJwLm8MAV1DxG/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1A3VuSpnEf8XWWR62O3BaPL8TpL4wQI7i&usp=drive_copy",
      "title": "Talaria: Segment-Based Audio Retrieval for Human-Centric Music Curation",
      "youtube_id": "https://drive.google.com/file/d/19ARvz_oXlitD7gbp9xkY6OhYXN1K3GJF/preview"
    },
    "forum": "466",
    "id": "466",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Adapting learning materials to the skill level of a student is crucial in education. In music training, sight-reading, the ability to play unfamiliar scores at first sight, benefits from progressive and level-appropriate practice. However, creating exercises with suitable difficulty often requires considerable time and expertise. We tackle this challenge as a controlled symbolic music generation task with the objective of producing piano scores at specified difficulty levels. Symbolic generation is typically conditioned through control tokens, but these often fail to influence global properties, such as difficulty. To address this, we introduce an auxiliary difficulty prediction objective that mitigates conditioning collapse, a frequent issue where models ignore control signals without explicit supervision. This objective encourages the model to learn internal representations aligned with the target difficulty, enabling more precise and adaptive score generation. Evaluations using automatic metrics and expert judgment demonstrate improved difficulty control and highlight the potential educational value of our approach, which represents a step toward personalized piano practice material.",
      "abstract": "Adapting learning materials to the skill level of a student is crucial in education. In music training, sight-reading, the ability to play unfamiliar scores at first sight, benefits from progressive and level-appropriate practice. However, creating exercises with suitable difficulty often requires considerable time and expertise. We tackle this challenge as a controlled symbolic music generation task with the objective of producing piano scores at specified difficulty levels. Symbolic generation is typically conditioned through control tokens, but these often fail to influence global properties, such as difficulty. To address this, we introduce an auxiliary difficulty prediction objective that mitigates conditioning collapse, a frequent issue where models ignore control signals without explicit supervision. This objective encourages the model to learn internal representations aligned with the target difficulty, enabling more precise and adaptive score generation. Evaluations using automatic metrics and expert judgment demonstrate improved difficulty control and highlight the potential educational value of our approach, which represents a step toward personalized piano practice material.",
      "authors": [
        "Pedro Ramoneda",
        "Masahiro Suzuki",
        "Akira Maezawa",
        "xavier Serra"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1L0UEx-PcFjfgffdJZni0LB8eGeycoulZ/preview",
      "poster_link": "https://drive.google.com/file/d/1sGn6XuQ_0GcsCIg0-UplKOvanfLHTP9F/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1xFdbnMVRqW7fiwQVmMggWA_jPRgrZjMo&usp=drive_copy",
      "title": "Towards Personalized Piano Practice Material",
      "youtube_id": ""
    },
    "forum": "467",
    "id": "467",
    "position": ""
  },
  {
    "content": {
      "TLDR": "This study investigates human brain responses to musical pitch during perception and imagery tasks using electro-encephalogram (EEG). Seven participants in the experiment were engaged in the two tasks: the first task involved passive listening to five-note melodies composed of seven pitches (C4\u2013B4), while the second required them to imagine the same melodies guided by a flashing static note cue displayed on a piano keyboard. Using event-related potential (ERP) analysis, we explored both eye-included and eye-cleaned EEG signals. ERPs reconstructed from eye-related independent components (ICs) revealed distinct hemispheric polarity, with left-right ERP polarity reversed across listening and imagery conditions. After removing eye-related ICs, statistical comparisons between the lowest (C4) and highest (B4) pitches showed significant ERP differences across specific time windows and channels in both perception and imagery. These findings suggest a complex, context-sensitive representation of musical pitch in the human EEG, influenced by both modality and the relative height difference of notes.",
      "abstract": "This study investigates human brain responses to musical pitch during perception and imagery tasks using electro-encephalogram (EEG). Seven participants in the experiment were engaged in the two tasks: the first task involved passive listening to five-note melodies composed of seven pitches (C4\u2013B4), while the second required them to imagine the same melodies guided by a flashing static note cue displayed on a piano keyboard. Using event-related potential (ERP) analysis, we explored both eye-included and eye-cleaned EEG signals. ERPs reconstructed from eye-related independent components (ICs) revealed distinct hemispheric polarity, with left-right ERP polarity reversed across listening and imagery conditions. After removing eye-related ICs, statistical comparisons between the lowest (C4) and highest (B4) pitches showed significant ERP differences across specific time windows and channels in both perception and imagery. These findings suggest a complex, context-sensitive representation of musical pitch in the human EEG, influenced by both modality and the relative height difference of notes.",
      "authors": [
        "Chaeeun Yoon"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1SQ_03lnBDSip2mD8QOhj6iyXPeR5S-hc/preview",
      "poster_link": "https://drive.google.com/file/d/1UwucsJmRpYHLIO7NAoDPpj5_TFTaQORr/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1fKUHlm_J0t9FOTRtYuiUYc3WDEB5HkYx&usp=drive_copy",
      "title": "Neural representations of musical pitch during music perception and imagery",
      "youtube_id": ""
    },
    "forum": "468",
    "id": "468",
    "position": ""
  },
  {
    "content": {
      "TLDR": "This demo demonstrates the quality and usability of 30-second music segments generated by the Talaria system, an AI music retrieval system developed by KKCompany Technologies. The Talaria system supports playlist editors in exploring extensive music records based on the semantic-level similarity. The research field of foundation models for music has been booming in recent years. KKCompany Technologies developed the Talaria system based on music foundation models. We conducted the user studies to investigate the usability and effectiveness of the selected music segments in this demo. We compared the subjective evaluation between the AI-selected 30-second music segment and the traditional first 30-second segment (or Head in short). Our findings indicate that Talaria selected segments provide better representativeness and received better user impressions than the Head segment, mainly due to their melodic and vocal features. The study demonstrates the effectiveness of the music foundation model and the potential for applying the AI system in professional editorial workflows.",
      "abstract": "This demo demonstrates the quality and usability of 30-second music segments generated by the Talaria system, an AI music retrieval system developed by KKCompany Technologies. The Talaria system supports playlist editors in exploring extensive music records based on the semantic-level similarity. The research field of foundation models for music has been booming in recent years. KKCompany Technologies developed the Talaria system based on music foundation models. We conducted the user studies to investigate the usability and effectiveness of the selected music segments in this demo. We compared the subjective evaluation between the AI-selected 30-second music segment and the traditional first 30-second segment (or Head in short). Our findings indicate that Talaria selected segments provide better representativeness and received better user impressions than the Head segment, mainly due to their melodic and vocal features. The study demonstrates the effectiveness of the music foundation model and the potential for applying the AI system in professional editorial workflows.",
      "authors": [
        "Tzu-Chen Lo",
        "Pin-Han Wang",
        "Yi-Hong Chen",
        "Yu-Hsun Lin",
        "Jing-Kai Lou",
        "Shuen-Huei Guan"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1PCX9oZN8ueezLgbNIGn3heqixaTGiU2T/preview",
      "poster_link": "https://drive.google.com/file/d/1GMnpfD-_pRIu3YABMcW9fQTXU8xIEyZy/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1kSMIKZywdGmo9SPt5y2SqAN1PpVxiOxU&usp=drive_copy",
      "title": "The Human Side of Music Segments: Perceptual Evaluation of AI-Selected Snippets",
      "youtube_id": ""
    },
    "forum": "469",
    "id": "469",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Popular Western music often includes song lyrics, with US samples indicating that nearly all personal song libraries contain lyrics. Social science theory suggests that personal values expressed in widely consumed text, like song lyrics, may align with or diverge from those of the listener, potentially shaping their response to the music. This suggests that there is potential for values estimated in",
      "abstract": "Popular Western music often includes song lyrics, with US samples indicating that nearly all personal song libraries contain lyrics. Social science theory suggests that personal values expressed in widely consumed text, like song lyrics, may align with or diverge from those of the listener, potentially shaping their response to the music. This suggests that there is potential for values estimated in",
      "authors": [
        "daniel P\u00e9ter",
        "alex Despan",
        "Andrew Demetriou",
        "jaehun Kim",
        "Cynthia Liem"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/12cVuTSlMQQggBa6CoCb-dz1OSiH-vt8U/preview",
      "poster_link": "https://drive.google.com/file/d/1SflYpXOF98s9_xH2hql-pf3Ywl9tu9OR/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1BaeQFxaowWUmQ6ewLbJwpJjrIRAo1UsS&usp=drive_copy",
      "title": "DEMO: TOWARDS AUTOMATIC PERSONAL VALUE ESTIMATION IN SONG LYRICS WITH LARGE LANGUAGE MODELS",
      "youtube_id": ""
    },
    "forum": "470",
    "id": "470",
    "position": ""
  },
  {
    "content": {
      "TLDR": "The recently introduced Model Context Protocol (MCP) provides a standardised way for LLMs to interact with external tools. We implement an MCP server for Max/MSP, a visual programming language for music and multimedia, enabling LLMs to directly understand and generate Max patches, paving the way for flexible sound generation using LLMs. The LLM agent can dynamically retrieve official documentation of Max objects before taking actions, allowing for in-context learning. For understanding, it can explain user-selected objects. For generation, it can create objects, connect patch cords, send messages to objects, and more. This offers a more seamless way to assist Max users compared to traditional approaches where LLMs respond with text-based answers or code snippets. Moreover, it eliminates syntax errors caused by generating different formats of text-based code, as the process reduces to selecting the correct functions to call and providing the appropriate arguments. Code and demonstration are available online in a repository: https://github.com/tiianhk/MaxMSP-MCP-Server.",
      "abstract": "The recently introduced Model Context Protocol (MCP) provides a standardised way for LLMs to interact with external tools. We implement an MCP server for Max/MSP, a visual programming language for music and multimedia, enabling LLMs to directly understand and generate Max patches, paving the way for flexible sound generation using LLMs. The LLM agent can dynamically retrieve official documentation of Max objects before taking actions, allowing for in-context learning. For understanding, it can explain user-selected objects. For generation, it can create objects, connect patch cords, send messages to objects, and more. This offers a more seamless way to assist Max users compared to traditional approaches where LLMs respond with text-based answers or code snippets. Moreover, it eliminates syntax errors caused by generating different formats of text-based code, as the process reduces to selecting the correct functions to call and providing the appropriate arguments. Code and demonstration are available online in a repository: https://github.com/tiianhk/MaxMSP-MCP-Server.",
      "authors": [
        "Haokun Tian",
        "Shuoyang Zheng",
        "Mathieu Lagrange",
        "Stefan Lattner",
        "Charalampos Saitis"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/15QLh9NZP-5YQTBQB5HQ8ZXw4VfQD0Dxq/preview",
      "poster_link": "https://drive.google.com/file/d/1OAaaeM7dKjjeLN-Ur1sGASN1TMzYtWmn/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1TGrdCb2mfg0Hw79S3CW_InKOIvb0grDR&usp=drive_copy",
      "title": "Agentic Max/MSP Programming with LLMs and MCP",
      "youtube_id": ""
    },
    "forum": "472",
    "id": "472",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Although sight-singing is a fundamental skill for all musicians,\u3000many struggle to acquire it effectively due to its inherent difficulty. In order to support sight-singing practice, this study proposes a color-coded musical score that visualizes the estimated difficulty of each note. The difficulty is predicted using a bi-directional LSTM model trained on performance data (SPED), which achieved a moderate correlation on test data. A preliminary application suggests that the resulting visualization is reasonably consistent with perceived difficulty, based on the author\u2019s evaluation.",
      "abstract": "Although sight-singing is a fundamental skill for all musicians,\u3000many struggle to acquire it effectively due to its inherent difficulty. In order to support sight-singing practice, this study proposes a color-coded musical score that visualizes the estimated difficulty of each note. The difficulty is predicted using a bi-directional LSTM model trained on performance data (SPED), which achieved a moderate correlation on test data. A preliminary application suggests that the resulting visualization is reasonably consistent with perceived difficulty, based on the author\u2019s evaluation.",
      "authors": [
        "Hiroki Karatsu",
        "Satoru Fukayama"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1qIqSN71evQh4n5b7PGoc_erIh_rdmOH3/preview",
      "poster_link": "https://drive.google.com/file/d/1pMQtsmV3QHN3-78jqCe0t8HeL-3P80vm/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=15E8Ubh7wXs0lhxCRmrsqZoY9LXbwBYyA&usp=drive_copy",
      "title": "Visualizing Difficulty in Sight-Singing Practice with Color-Coded Scores",
      "youtube_id": ""
    },
    "forum": "474",
    "id": "474",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Breast cancer, one of the most prevalent malignancies worldwide, poses a significant disease burden, and is characterized by its heterogeneity and variable patient outcomes. To advance precision medicine, novel approaches are needed to improve the interpretability of complex genomic data. This study introduces BRCASonics, a prototype tool that applies data sonification\u2014a method of representing data through sound\u2014to a 43-gene expression signature previously identified as predictive of breast cancer survival using The Cancer Genome Atlas Breast Cancer (TCGA-BRCA) dataset. By mapping gene expression levels to auditory properties such as pitch, rhythm, and timbre, BRCASonics transforms high-dimensional genomic data into an interactive, multisensory experience. This approach not only enhances pattern recognition and accessibility but also holds educational value by providing an engaging medium for students to explore genomic concepts. Developed as an R Shiny web application, BRCASonics allows users to explore variations in gene expression across samples through auditory cues. This work demonstrates the feasibility and utility of sonification as a complementary tool to traditional data visualization, with potential applications in both clinical interpretation and genomic education. Ultimately, this research paves the way for multisensory frameworks in data communication, offering an inclusive and intuitive paradigm for engaging with complex biological information.",
      "abstract": "Breast cancer, one of the most prevalent malignancies worldwide, poses a significant disease burden, and is characterized by its heterogeneity and variable patient outcomes. To advance precision medicine, novel approaches are needed to improve the interpretability of complex genomic data. This study introduces BRCASonics, a prototype tool that applies data sonification\u2014a method of representing data through sound\u2014to a 43-gene expression signature previously identified as predictive of breast cancer survival using The Cancer Genome Atlas Breast Cancer (TCGA-BRCA) dataset. By mapping gene expression levels to auditory properties such as pitch, rhythm, and timbre, BRCASonics transforms high-dimensional genomic data into an interactive, multisensory experience. This approach not only enhances pattern recognition and accessibility but also holds educational value by providing an engaging medium for students to explore genomic concepts. Developed as an R Shiny web application, BRCASonics allows users to explore variations in gene expression across samples through auditory cues. This work demonstrates the feasibility and utility of sonification as a complementary tool to traditional data visualization, with potential applications in both clinical interpretation and genomic education. Ultimately, this research paves the way for multisensory frameworks in data communication, offering an inclusive and intuitive paradigm for engaging with complex biological information.",
      "authors": [
        "JiaYi Lee",
        "JiaWei Lee",
        "CarolineGuatLay Lee"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1rrWR21v2qULnZJtgvjT394BTHdeFdHR-/preview",
      "poster_link": "https://drive.google.com/file/d/1rkRWzLJ1WD7G0rPrL5SDbU_n1n0beJUq/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1TIW3WBUrMhh84ru8hZickMMdMuJhWTkX&usp=drive_copy",
      "title": "BRCASonics: AI-Based Sonification of a 43-Gene Signature Predictive of Breast Cancer Prognosis",
      "youtube_id": ""
    },
    "forum": "475",
    "id": "475",
    "position": ""
  },
  {
    "content": {
      "TLDR": "The LinkedMusic Project aims to transform music re-search by creating a global digital library that integrates diverse online music databases. Unlike textual search, music search is challenged by varied content formats and incompatible metadata schemas. To address this, we con-vert metadata into RDF and store it in a central graph da-tabase. This infrastructure supports SESEMMI\u2014an open-source, multilingual metasearch engine enabling cross-database SPARQL queries. A key feature is UMIL (Uni-versal Music Instrument Lexicon), a crowdsourced plat-form collecting multilingual instrument names and imag-es to support culturally inclusive searches. SESEMMI applies Linked Data principles and draws on external sources like Wikidata to enrich results. To improve ac-cessibility, Large Language Models translate natural lan-guage into SPARQL queries. By uniting experts in eth-nomusicology, library science, and music information retrieval, LinkedMusic tackles longstanding interoperabil-ity issues. The project enables large-scale, multilingual, and demographically informed music exploration, open-ing new avenues for discovery, analysis, and preservation of global traditions.",
      "abstract": "The LinkedMusic Project aims to transform music re-search by creating a global digital library that integrates diverse online music databases. Unlike textual search, music search is challenged by varied content formats and incompatible metadata schemas. To address this, we con-vert metadata into RDF and store it in a central graph da-tabase. This infrastructure supports SESEMMI\u2014an open-source, multilingual metasearch engine enabling cross-database SPARQL queries. A key feature is UMIL (Uni-versal Music Instrument Lexicon), a crowdsourced plat-form collecting multilingual instrument names and imag-es to support culturally inclusive searches. SESEMMI applies Linked Data principles and draws on external sources like Wikidata to enrich results. To improve ac-cessibility, Large Language Models translate natural lan-guage into SPARQL queries. By uniting experts in eth-nomusicology, library science, and music information retrieval, LinkedMusic tackles longstanding interoperabil-ity issues. The project enables large-scale, multilingual, and demographically informed music exploration, open-ing new avenues for discovery, analysis, and preservation of global traditions.",
      "authors": [
        "Ichiro Fujinaga"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1_D-aZSR-EkC3EBAFv3HbmAIksWMmsYbZ/preview",
      "poster_link": "https://drive.google.com/file/d/1GyhUcfyjXidXxhQhk0UUtOc4INd5Hxqh/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1Y_fZDZdB-b-_XddVnEzEkFRYE8679c9p&usp=drive_copy",
      "title": "LINKEDMUSIC PROJECT: INTEGRATING MUSIC DATABASES WITH LINKED DATA",
      "youtube_id": ""
    },
    "forum": "476",
    "id": "476",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Selecting a soundtrack is a critical step in video editing. However, evaluating music is a slow, sequential process. Creators must listen to tracks one by one, making direct comparison difficult and forcing them to rely on auditory memory to predict a track's impact on their video. We present VidTune, a system that facilitates exploration and comparison of generative soundtracks with Visual Thumbnails. We introduce a technique for generating contextual visual thumbnails that translate a music track's character into a stylized preview of the user's own video. Our method maps analyzed musical attributes like valence and energy to visual parameters such as color and brightness applied to a keyframe. This approach transforms soundtrack selection from a slow process of sequential listening into a rapid act of parallel visual comparison, allowing creators to more intuitively imagine each track's final impact on their video.",
      "abstract": "Selecting a soundtrack is a critical step in video editing. However, evaluating music is a slow, sequential process. Creators must listen to tracks one by one, making direct comparison difficult and forcing them to rely on auditory memory to predict a track's impact on their video. We present VidTune, a system that facilitates exploration and comparison of generative soundtracks with Visual Thumbnails. We introduce a technique for generating contextual visual thumbnails that translate a music track's character into a stylized preview of the user's own video. Our method maps analyzed musical attributes like valence and energy to visual parameters such as color and brightness applied to a keyframe. This approach transforms soundtrack selection from a slow process of sequential listening into a rapid act of parallel visual comparison, allowing creators to more intuitively imagine each track's final impact on their video.",
      "authors": [
        "Mina Huh",
        "Ailie Fraser",
        "Dingzeyu Li",
        "Mira Dontcheva",
        "Bryan Wang"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1tBbXamXc4QnJ_lustGOyqR1o0Mre0Pmw/preview",
      "poster_link": "https://drive.google.com/file/d/1FF5922Nso0g-oQZ63t6l2uHpB2U8ZKC-/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1uey_VQaaK5a93eI77E7N1l7nLlqy7FiM&usp=drive_copy",
      "title": "Generating Video Soundtracks with Contextual Visual Thumbnails",
      "youtube_id": "https://drive.google.com/file/d/1kDJUyTlTmLNbzv-0trbdMpRJPqOkydME/preview"
    },
    "forum": "478",
    "id": "478",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We propose MusicViz, a graphical animation system that visualizes decision processes in transformer-based music models. The purpose of our system is to make human-artificial intelligence (AI) collaboration more transparent and intuitive. By combining piano-roll and staff notation, the system displays notes as rectangles linked by weighted arcs representing attention between note tokens. This visualization helps musicians grasp the model\u2019s \u201cthought process\u201d by revealing detected patterns such as motifs and longer sequences. Interactive features allow users to zoom, scroll through time, and toggle attention weights, enabling detailed exploration of the model\u2019s behavior. By revealing the AI's internal focus, musicians can adapt their input to guide generated responses, fostering more effective collaboration. MusicViz addresses the challenge of interpreting complex transformer computations in a human-friendly way. The current offline prototype lays the groundwork for future real-time generative music visualization, with potential applications in live performance and improvisation.",
      "abstract": "We propose MusicViz, a graphical animation system that visualizes decision processes in transformer-based music models. The purpose of our system is to make human-artificial intelligence (AI) collaboration more transparent and intuitive. By combining piano-roll and staff notation, the system displays notes as rectangles linked by weighted arcs representing attention between note tokens. This visualization helps musicians grasp the model\u2019s \u201cthought process\u201d by revealing detected patterns such as motifs and longer sequences. Interactive features allow users to zoom, scroll through time, and toggle attention weights, enabling detailed exploration of the model\u2019s behavior. By revealing the AI's internal focus, musicians can adapt their input to guide generated responses, fostering more effective collaboration. MusicViz addresses the challenge of interpreting complex transformer computations in a human-friendly way. The current offline prototype lays the groundwork for future real-time generative music visualization, with potential applications in live performance and improvisation.",
      "authors": [
        "Rachel Loh",
        "Ashley Zhang",
        "Anna Huang",
        "Eran Egozy",
        "Heidi Lei",
        "Lancelot Blanchard"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1Zsdzdkij0CBZhXdi8bIpzuM5GdZbzrl4/preview",
      "poster_link": "https://drive.google.com/file/d/1prEJFYGhN2Gv6GeNzo1zI9FEcdRJe4Yz/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1EJTP2W7afAbDsp5g5WO-pfXlnrZAisxr&usp=drive_copy",
      "title": "MusicViz: Visualizing Decision Processes in Transformer-Based Music Models",
      "youtube_id": "https://drive.google.com/file/d/1-vZw3a0x0v_PZOEDYNqYDSWQB-PEgiWf/preview"
    },
    "forum": "479",
    "id": "479",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We present the prototype of a music recommendation system enabling users to make specific requests using natural language prompts, such as \"Recommend slower hip-hop songs without drums but with guitar.\" Our approach combines audio source separation with large language model (LLM) query processing to bridge the gap between user intent and music discovery. The system uses Demucs 6-source stem separation to extract instrument presence information from 604 songs across 8 genres from the Free Music Archive, stores this metadata in a structured database, and employs Groq\u2019s Llama-3.3-70B-Versatile model to convert natural language queries into SQL statements for precise music retrieval. While current capabilities focus on instrument presence detection, the system architecture supports future expansion to other MIR tasks such as Automatic Chord Estimation (ACE) and auto-tagging stems with instrument-specific metadata. This work provides a prototype of utilizing Music Source Separation (MSS) in recommendation systems.",
      "abstract": "We present the prototype of a music recommendation system enabling users to make specific requests using natural language prompts, such as \"Recommend slower hip-hop songs without drums but with guitar.\" Our approach combines audio source separation with large language model (LLM) query processing to bridge the gap between user intent and music discovery. The system uses Demucs 6-source stem separation to extract instrument presence information from 604 songs across 8 genres from the Free Music Archive, stores this metadata in a structured database, and employs Groq\u2019s Llama-3.3-70B-Versatile model to convert natural language queries into SQL statements for precise music retrieval. While current capabilities focus on instrument presence detection, the system architecture supports future expansion to other MIR tasks such as Automatic Chord Estimation (ACE) and auto-tagging stems with instrument-specific metadata. This work provides a prototype of utilizing Music Source Separation (MSS) in recommendation systems.",
      "authors": [
        "Vansh Chugh"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1UD5BaLg5gMyEwsD1fHi8EFs2PT1baH9f/preview",
      "poster_link": "https://drive.google.com/file/d/1K9Dm-pAKB9egGo-NOO_-8_XfTkIVkXN2/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1wy1myAcUh3oVr4oeDOYXx5UIMDZ46kU_&usp=drive_copy",
      "title": "Prompt-Based Music Discovery: A Prototype Using Source Separation and LLMs",
      "youtube_id": "https://drive.google.com/file/d/1usbsVu6H0DPXwidwzndAvPPqXk3Fg4pX/preview"
    },
    "forum": "480",
    "id": "480",
    "position": ""
  },
  {
    "content": {
      "TLDR": "This late-breaking demo presents progress toward a DJ performance system that enables real-time manipulation of musical style. Drawing on Neural Style Transfer (NST) concepts, the envisioned system uses a hardware controller to interpolate between learned stylistic embed-dings during live performance. As an initial step, we adapt Magenta RealTime (Magenta RT) [1] \u2014 an open-weights, low-latency music generation model \u2014 for playlist-driven, culturally specific fine-tuning. Using a curated 88-minute playlist of Indian music, we explore a workflow where DJs upload a playlist as the dataset and fine-tune the model based on the style they intend to per-form. In informal listening tests, the fine-tuned model appeared to capture more of the melodic, rhythmic, and timbral qualities of Indian music compared to the origi-nal Magenta RT, suggesting the potential for personal-ized, culturally adaptive AI-driven DJ tools.",
      "abstract": "This late-breaking demo presents progress toward a DJ performance system that enables real-time manipulation of musical style. Drawing on Neural Style Transfer (NST) concepts, the envisioned system uses a hardware controller to interpolate between learned stylistic embed-dings during live performance. As an initial step, we adapt Magenta RealTime (Magenta RT) [1] \u2014 an open-weights, low-latency music generation model \u2014 for playlist-driven, culturally specific fine-tuning. Using a curated 88-minute playlist of Indian music, we explore a workflow where DJs upload a playlist as the dataset and fine-tune the model based on the style they intend to per-form. In informal listening tests, the fine-tuned model appeared to capture more of the melodic, rhythmic, and timbral qualities of Indian music compared to the origi-nal Magenta RT, suggesting the potential for personal-ized, culturally adaptive AI-driven DJ tools.",
      "authors": [
        "Ivana RaschChinchilla"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1zQD5KKvUuebliw0mokKZkxMzpUwMf0I6/preview",
      "poster_link": "https://drive.google.com/file/d/1EsBksSi4V9h81GemWINb-7KeX8vbGDUU/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1etZKna1K88DksgwEhvB8uUlgQvY3NV4W&usp=drive_copy",
      "title": "Bringing Global Music to Magenta RealTime: Playlist-Driven Fine-Tuning for Personalized DJ Performance",
      "youtube_id": ""
    },
    "forum": "481",
    "id": "481",
    "position": ""
  },
  {
    "content": {
      "TLDR": "We present work in progress on a score-informed automatic music transcription and performance assessment system designed to support automated musical education. The system combines transcription models with score alignment and error detection tools to provide meaningful feedback to both students and teachers. It comprises an automatic guitar transcription model trained using a combination of public and internal data, along with pre-trained transcription models for piano and violin. Following transcription, the system evaluates performances by aligning the transcribed results with reference music scores to identify various errors, such as pitch, rhythm, tempo, structural, and intonation. This work demonstrates how intelligent systems can enhance the music learning experience by making it more interactive, personalized, and data-driven.",
      "abstract": "We present work in progress on a score-informed automatic music transcription and performance assessment system designed to support automated musical education. The system combines transcription models with score alignment and error detection tools to provide meaningful feedback to both students and teachers. It comprises an automatic guitar transcription model trained using a combination of public and internal data, along with pre-trained transcription models for piano and violin. Following transcription, the system evaluates performances by aligning the transcribed results with reference music scores to identify various errors, such as pitch, rhythm, tempo, structural, and intonation. This work demonstrates how intelligent systems can enhance the music learning experience by making it more interactive, personalized, and data-driven.",
      "authors": [
        "AikateriniMaria Primenta",
        "Jackson Loth",
        "Jingjing Tang",
        "Xavier Riley",
        "Simon Dixon",
        "Emmanouil Benetos"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1YVtxtFNioJEdHdduHyfEHRupBGM8POjY/preview",
      "poster_link": "https://drive.google.com/file/d/1Xs7EHmVRtpK-K7OM6XfvHP6KultCN9eF/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1cxjzjoRWvtrleqXWWTe-_yENNFf1UGt9&usp=drive_copy",
      "title": "Towards Intelligent Music Education: Score-Informed Transcription and Performance Assessment",
      "youtube_id": ""
    },
    "forum": "482",
    "id": "482",
    "position": ""
  },
  {
    "content": {
      "TLDR": "A cappella music presents unique challenges for source separation due to its diverse vocal styles and the presence of vocal percussion (VP). Current datasets are limited in size and diversity, hindering robust model development. To address this, we curated a high-quality human-performed dataset of 50 a cappella songs and propose expanding it using AI-driven voice cloning and synthesis techniques. We introduce a two-step source separation pipeline tailored for a cappella and present preliminary results on VP separation. Our dataset and approach enable applications such as source separation for a cappella rehearsal software.",
      "abstract": "A cappella music presents unique challenges for source separation due to its diverse vocal styles and the presence of vocal percussion (VP). Current datasets are limited in size and diversity, hindering robust model development. To address this, we curated a high-quality human-performed dataset of 50 a cappella songs and propose expanding it using AI-driven voice cloning and synthesis techniques. We introduce a two-step source separation pipeline tailored for a cappella and present preliminary results on VP separation. Our dataset and approach enable applications such as source separation for a cappella rehearsal software.",
      "authors": [
        "Ting-Yu Pan",
        "Kexin Ju",
        "Hao-Wen Dong"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Early Research"
      ],
      "paper_link": "https://drive.google.com/file/d/1T0bOGEibSyplvpOibGKUvLRFgDbTs6ub/preview",
      "poster_link": "https://drive.google.com/file/d/1WesBXRVPHDRRPxRdZxTZPCtkGDBIn00R/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1G49lKphouz9MXA2b95-b-Bby04E9P7gb&usp=drive_copy",
      "title": "Curating an A Cappella Dataset for Source Separation",
      "youtube_id": ""
    },
    "forum": "483",
    "id": "483",
    "position": ""
  },
  {
    "content": {
      "TLDR": "Recent attempts to automate harmonization have been advanced by machine learning. In particular, architectures that combine Variational Autoencoder (VAE) \\cite{kingma2022autoencodingvariationalbayes} with time-series processing methods such as Transformer and LSTM are commonly employed \\cite{morphing_reharmonization, structured_and_flexible, ji2023emotionconditionedmelodyharmonizationhierarchical}. Training with a VAE enables diverse outputs through interpolation and extrapolation. However, these approaches may not fully utilize the entire learned latent space. In addition, the generated accompaniments often consist of the same chord for an entire bar or monotonous rhythmic patterns. This paper investigates a method for style control by training a LSTM-VAE on melody and chord progressions, and measuring and comparing features of generated accompaniments across the entire latent space.",
      "abstract": "Recent attempts to automate harmonization have been advanced by machine learning. In particular, architectures that combine Variational Autoencoder (VAE) \\cite{kingma2022autoencodingvariationalbayes} with time-series processing methods such as Transformer and LSTM are commonly employed \\cite{morphing_reharmonization, structured_and_flexible, ji2023emotionconditionedmelodyharmonizationhierarchical}. Training with a VAE enables diverse outputs through interpolation and extrapolation. However, these approaches may not fully utilize the entire learned latent space. In addition, the generated accompaniments often consist of the same chord for an entire bar or monotonous rhythmic patterns. This paper investigates a method for style control by training a LSTM-VAE on melody and chord progressions, and measuring and comparing features of generated accompaniments across the entire latent space.",
      "authors": [
        "Itsuki Nakada",
        "Shinji Sako"
      ],
      "bilibili_id": "",
      "day": 4,
      "keywords": [
        "Software/Library Demo"
      ],
      "paper_link": "https://drive.google.com/file/d/1in3ei9_xQc0LI-kvcK30pgA01QRJ6xao/preview",
      "poster_link": "https://drive.google.com/file/d/1tEfR2aZuVHykv_fqWTpJbUz3K_4GjWNF/preview",
      "poster_type": "",
      "session": [
        ""
      ],
      "thumbnail_link": "https://drive.google.com/open?id=1zC2TlJ1no_yR6SmGdC_vJ9aDM92yUEf8&usp=drive_copy",
      "title": "Harmonization using a Style-Controllable Latent Space Based on VAE",
      "youtube_id": ""
    },
    "forum": "484",
    "id": "484",
    "position": ""
  }
]
