For questions, email mlforaudioworkshop@gmail.com
Machine learning research for audio applications has experienced a surge of innovation in recent years, with prominent and widely relevant advancements rapidly emerging and momentum continuing to build. There are numerous key problems within the audio research domain that continue to attract widespread attention. This ongoing relevance, alongside the success of the Machine Learning for Audio workshop at NeurIPS 2023 and ICML 2025, has inspired us to bring this workshop at ICML 2026. We believe that bringing this workshop to a wider audience will provide a good opportunity to bring together both practitioners of audio tools along with machine learning researchers interested in audio, in order to foster community, discussion, and future collaboration. In addition, with the field moving so rapidly, we believe this workshop will provide a dedicated space for the crucial ethical discussions that must be facilitated among researchers around applications of generative machine learning for audio.
The Machine Learning for Audio workshop at ICML 2026 will cover a broad range of tasks and challenges involving audio data. These include, but are not limited to: methods of speech modeling, environmental sound generation or other forms of ambient sound, novel generative models, music generation in the form of raw audio, text-to-speech methods, denoising of speech and music, data augmentation, classification of acoustic events, transcription, source separation, and multimodal problems.
We plan to solicit original extended abstracts (up to 4 pages) in these areas, which will be reviewed by the organizers and an additional set of reviewers. We anticipate approximately 30 accepted submissions. To avoid potential conflicts of interest, no organizer or reviewer will review a submitted paper from the same organization as the organizer or reviewer, enforced by CMT. We also plan to run a demo session alongside the poster session, where contributors will be able to present live demos of their work.
Our team of organizers were involved with two separate audio-related workshops at ICML 2022: the Workshop on Machine Learning for Audio Synthesis and ICML Expressive Vocalizations Workshop and Competition. We then combined our organizing committees and offered a workshop at NeurIPS 2023 entitled the Workshop on Machine Learning for Audio. Last year, we added new organizers to the team and hosted a workshop at ICML 2025. This year, we plan to improve upon previous iterations of the workshop with a lineup of prominent in-person invited speakers, more accessible data distribution (as outlined below), and more.
Recognizing the scarcity of free, publicly available audio data, Modulate and Hume AI will contribute several datasets in the speech domain alongside the workshop, all of large scale for their respective domains. These datasets, accessible via Google Drive, will include acted speech (professionally acted scripts), spontaneous speech (streamer content), mimicked speech (short-form emotive recordings), and mimicked non-verbal speech. The organizers hope this allows researchers from smaller research groups and academia to work with and validate findings on larger, more generalizable datasets. In previous iterations, multiple submissions utilized versions of provided data in their work, and a corresponding white paper was subsequently posted on arXiv.
Further details on available data described here.
We are calling for extended abstracts up to 4 pages excluding references. Accepted submissions will be posted on the workshop website but not published/archived. Several submissions will be chosen for 15-minute contributed talks and the remaining selected submissions will participate in the poster & demo session. Please make sure submissions adhere to the ICML format. The review process will be double-blind so please make sure not to put any author information in your submission. Authors may also submit supplementary materials along with their papers if they wish (e.g., a preview of a potential demo). Reviewers will not be required to read/view/listen to said supplementary material.
Timeline
Submission deadline (main paper & all supplementary material): May 25 23:59:59 AOE
Accept/Reject notification date: June 1 AOE
We plan for the workshop to be an 8-hour event. Below is an approximate timetable of the workshop schedule, subject to change. We have been careful to facilitate ample time for informal discussion during the coffee break, poster & demo session, and open conversation session, as well as time for audience participation during the panel discussion and Q&A sections following invited talks.
| Time (KST) | Topic |
|---|---|
| 8:25 AM | Opening remarks — Brian Kulis |
| 8:30 AM | Tara Sainath |
| 9:00 AM | Marius Miron: Animal Language Processing: AI to decode communication beyond humans |
| 9:30 AM | Flow Fake: Parametric Efficient Alternative for Transformers |
| 9:50 AM | MondegreensEval: A Phonetic Benchmark for Measuring Language-model Bias in Automatic Speech Recognition |
| 10:10 AM | PianoKontext: Expressive Performance Rendering from Deadpan Context |
| 10:30 AM | Coffee break |
| 11:00 AM | Juhan Nam |
| 11:30 AM | Minje Kim: Neural Speech and Audio Coding: Efficient Representations, Emerging Capabilities, and Open Challenges |
| 12:00 PM | Lunch |
| 1:00 PM | Poster Session |
| 2:00 PM | Heiga Zen: From Statistical Models to Foundation Capabilities: The Historical Progression of Speech Generation |
| 2:30 PM | Prosodic Differences Between Child-Directed and Adult-Directed Speech in Text-to-Speech Generation |
| 2:50 PM | Few-Shot Synthetic Accented Speech for ASR Fine-Tuning: What Helps and When? |
| 3:10 PM | Coffee break |
| 3:30 PM | Autoregressive Zero-Shot Voice Conversion |
| 3:50 PM | Decoding Insect Song: A Multitask Semisupervised Orthoptera Bioacoustic Classifier |
| 4:10 PM | Multilingual Speech Editing |
| 4:30 PM | Closing remarks — Brian Kulis |
Minje Kim is an Associate Professor in the Siebel School of Computing and Data Science at the University of Illinois Urbana-Champaign and an Amazon Scholar. His research focuses on efficient machine learning for audio, including efficient data representations (e.g., neural audio coding), intelligent signal processing (e.g., speech enhancement and source separation), and generative modeling of audio.
Marius Miron is a Senior AI Research Scientist at the Earth Species Project, where he builds machine-learning and signal-processing methods for bioacoustics to help decode animal communication. Previously, he worked in music AI and audio signal processing (including a PhD on orchestral music source separation) at the Music Technology Group at Pompeu Fabra University.
Tara Sainath is a Distinguished Research Scientist at Google DeepMind and co-lead of the Gemini Audio pillar, known for applying deep learning to advance automatic speech recognition. She earned her S.B., M.Eng., and PhD in EECS from MIT and previously worked at IBM’s T.J. Watson Research Center.
Juhan Nam is a professor at KAIST's Graduate School of Culture Technology and leads the Music and Audio Computing Lab, where he researches music information retrieval and audio/music signal processing. He also serves as an affiliate professor at the Kim Jaechul Graduate School of Artificial Intelligence and the Graduate School of Metaverse. He is a co-founder of Neutune and AudAi.
Heiga Zen is a Principal Scientist at Google DeepMind in Japan, where he researches speech technology and machine learning. He is one of the original authors and first maintainer of the HMM-based speech synthesis system (HTS), and is a Fellow of ISCA and IEEE.
| Speaker | Title | Abstract |
|---|---|---|
| Marius Miron | Animal Language Processing: AI to decode communication beyond humans | AI already supports conservation at a remarkable scale. Audio and video monitoring now generate petabytes of data for biodiversity monitoring and ecological research. These systems primarily detect and classify, they can tell us if a species is present or absent. But what if we could move beyond detecting species into understanding their communication? Animal Language Processing is an emerging interdisciplinary field that combines biology, bioacoustics, and AI to understand the meaning and function of animal communication: bird dialects, spider courtship dances, calls to label conspecifics (the animal equivalent of names), matriarchal clans that pass culture across generations. For machine learning, this is a vast test bed with profound implications for our relationship with the rest of nature. However, it requires defining new evaluation metrics and methods and reaching beyond anthropocentric assumptions to deal with the biggest challenge in the field: lack of ground truth or predefined labels. |
| Juhan Nam | LLM4FM: Empowering LLMs to Generate Yamaha DX7 Patches from Text and Audio | Programming FM synthesizers is notoriously difficult due to the highly non-linear relationship between parameters and sound. In this talk, I present LLM4FM, a framework that enables Large Language Models (LLMs) to generate Yamaha DX7 patches from text descriptions or audio examples. To support this task, we introduce DX7Caps, the first dataset pairing DX7 patches with natural-language captions, and propose Operator-Isolated Audio Grounding for CoT Distillation. We further extend the framework to sound matching by generating DX7 patches directly from audio. Results from objective evaluations, listening tests, and LLM-based assessments demonstrate the potential of LLMs as practical assistants for FM sound design. |
| Minje Kim | Neural Speech and Audio Coding: Efficient Representations, Emerging Capabilities, and Open Challenges | This talk will provide an overview of the rapidly evolving landscape of neural speech and audio coding (NSAC). Recent progress has shown that data-driven coding systems, when paired with appropriate learning objectives and model architectures, can achieve substantial gains in coding efficiency over conventional approaches. Beyond improved representational efficiency, NSAC also opens the door to new capabilities that are difficult to realize with traditional codecs, including personalized speech coding, task-specific representations for audio coding for machines (ACoM), and cascaded residual learning frameworks that make neural codecs more flexible and expressive. The talk will also discuss key challenges that are specific to NSAC, including the computational cost of neural encoders and decoders, the risk of hallucination and other generative artifacts, and the difficulty of balancing perceptual quality with faithful signal reconstruction. Finally, I will highlight recent efforts to address these issues through more efficient model architectures, improved training strategies, and semantic loss functions that better align codec behavior with human perception and downstream machine-listening tasks. |
| Heiga Zen | From Statistical Models to Foundation Capabilities: The Historical Progression of Speech Generation | The field of speech generation has undergone massive transformations, evolving from physical model simulations and concatenative systems to advanced foundational models. This talk will trace the historical progression of generative approaches for speech generation. |
Alice Baird is a senior AI research scientist at Hume AI, NY, USA, where she works on modeling expressive human behaviors from audio and other modalities. She earned her Ph.D. at the University of Augsburg in 2022. Her work on emotion understanding from auditory, physiological, and multimodal data has been widely published in leading journals and conferences. She has co-organized several machine learning competitions, including the 2022 ICML Expressive Vocalizations Workshop and the 2023 NeurIPS Workshop on Machine Learning for Audio.
Sander Dieleman is a research scientist at DeepMind in London, UK, where he contributed to the development of AlphaGo and WaveNet. His research focuses on generative modeling of perceptual signals at scale, including audio (speech & music) and visual data. He has co-organized multiple workshops, including the NeurIPS workshop on machine learning for creativity and design (2017-2020), the Recsys workshop on deep learning for recommender systems (2016-2018), the Machine Learning for Audio Synthesis workshop at ICML 2022, and the Workshop on Machine Learning for Audio at NeurIPS 2023.
Chris Donahue is an assistant professor at Carnegie Mellon University and a research scientist at Google DeepMind. His research focuses on developing and responsibly deploying generative AI for music and creativity to unlock and augment human creative potential. His work includes improving machine learning methods for controllable generative modeling for music, audio, and sequential data, as well as deploying interactive systems that allow a broad audience—including non-musicians—to harness generative music AI through intuitive controls.
Brian Kulis is an associate professor at Boston University and a former Amazon Scholar who worked on Alexa. His research focuses on machine learning, particularly applications in audio problems such as detection and generation. He has won best paper awards at ICML and CVPR and has organized multiple workshops at ICCV, NeurIPS, and ICML. He has also served as an area or senior area chair at major AI conferences and has organized tutorials at ICML and ECCV.
David Liu is a Ph.D. student in the Department of Computer Science at Boston University. His research focuses on deep learning for audio, with a particular emphasis on state-space models. He earned his bachelor’s degree in computer science, data science, and mathematics from the University of Wisconsin - Madison in 2023.
Rachel Manzelli is the Machine Learning Team Lead at Modulate, where she leads the development of audio generation and classification models supporting moderation teams in detecting harms in voice conversations (ToxMod) and real-time voice conversion (VoiceWear). Previously, she worked at Macro as a machine learning engineer, focusing on source separation models. She has co-organized the Machine Learning for Audio Synthesis workshop at ICML 2022 and the Workshop on Machine Learning for Audio at NeurIPS 2023. She earned her bachelor’s degree in computer engineering from Boston University in 2019, where she conducted research in structured music generation and MIR.
Shrikanth Narayanan is a University Professor and holder of the Niki and Max Nikias Chair in Engineering at the University of Southern California (USC). Shri is a Fellow of the National Academy of Inventors (NAI), the Acoustical Society of America (ASA), the Institute of Electrical and Electronics Engineers (IEEE), the International Speech Communication Association (ISCA), the Association for Psychological Science (APS), the American Association for the Advancement of Science (AAAS), American Institute for Medical and Biological Engineering (AIMBE) and the Association for the Advancement of Affective Computing (AAAC). Shri is a member of the European Academy of Sciences and Arts and a 2022 Guggenheim Fellow.
The Microsoft CMT service was used for managing the peer-reviewing process for this conference. This service was provided for free by Microsoft and they bore all expenses, including costs for Azure cloud services as well as for software development and support.