Faust Autodiff: Towards Audio Domain-Specific Machine Learning

Rushton T, Orlarey Y, Michon R, Risset T, Letz S. Faust Autodiff: Towards Audio Domain-Specific Machine Learning. In: 2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio. Audio Engineering Society; 2025. Paper 28. Available from: https://aes.org/publications/elibrary-page/?id=23017

@inproceedings{Rushton2025_23017,
  author = {Rushton, Thomas and Orlarey, Yann and Michon, Romain and Risset, Tanguy and Letz, Stéphane},
  title = {{Faust Autodiff: Towards Audio Domain-Specific Machine Learning}},
  booktitle = {2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio},
  note = {Paper 28},
  year = {2025},
  month = sep,
  publisher = {Audio Engineering Society},
  url = {https://aes.org/publications/elibrary-page/?id=23017}
}

TY  - CPAPER
TI  - Faust Autodiff: Towards Audio Domain-Specific Machine Learning
AU  - Rushton, Thomas
AU  - Orlarey, Yann
AU  - Michon, Romain
AU  - Risset, Tanguy
AU  - Letz, Stéphane
T2  - 2025 AES International Conference on Artificial Intelligence and Machine Learning for Audio
M1  - Paper 28
PY  - 2025
DA  - 2025/09/02
UR  - https://aes.org/publications/elibrary-page/?id=23017
PB  - Audio Engineering Society
LA  - en
AB  - Differentiable programming via automatic differentiation (AD) is the foundation for gradient-based optimisation techniques, and forms the basis for many current approaches to machine learning. Though well catered-for in general-purpose programming languages, the availability of AD in a domain-specific language (DSL) could offer novel perspectives on optimisation problems in the field of audio. We present a general scheme for differentiable programming in the Faust programming language, a high-performance DSL tailored to audio synthesis and signal processing. Faust's rich ecosystem, coupled with a comprehensive AD implementation, can provide support for audio optimisation and machine learning applications on a multitude of platforms, from FPGAs to the web.
ER  -