Skip to content

lucystrauss/MLMLMLM

Repository files navigation

MLMLMLM

Muscle-Listening Machine Learning Model for Live Music

Lucy Strauss, Prashanth Thattai Ravikumar, and Matthew Yee-King. 2026. Cross-Modal Sig2Sig Machine Translation with Deep Generative Modeling for NIME Design. Proceedings of the International Conference on New Interfaces for Musical Expression. DOI: 10.5281/zenodo.20784411

MLMLMLM is a custom model architecture for translating electromyographic (EMG) signals into probable audio in live music performance settings. This is implemented through EMG-conditioned sequence generation of audio signals.

The MLMLMLM model architecture is composed of two RVQ-VAEs and a decoder-only Transformer. Each model is trained separately. One RVQ-VAE models EMG signals and the other RVQ-VAE models audio signals. The Transformer is implemented in latent space, using quantized latent vectors of the audio RVQ-VAE for self-attention, and quantized latent vectors of the EMG RVQ-VAE for cross-attention.

Requirements

Training Requirements

stable-audio-tools

Inference Requirements

pybela

Additional Requiremets

See setup.py for additional requirements.

To run training, you will also need a time-aligned dataset of audio and EMG signals. We have not made our dataset public because we consider this to be part of the artwork and the musician did not wish to make the dataset publicly available. However, you can reach out to us here. We share this repo for research purposes, but do not intend for our exact model training to be reproducible. Take the script, remix it and make your own version! Please cite the repo and paper if you do.

Install

First, install stable-audio-tools from PyPI with:

$ pip install stable-audio-tools

Then:

$ pip install .

Shoutouts

The scripts for live interaction in performance settings are built on pybela.

The models in this repo are adapted from stable-audio-tools. Our main changes are to adapt the model architecture capacity and hyperparameters for deep generative modeling of a 6-channel EMG dataset. Additionally, we added functionality to training scripts to allow for causal, autoregressive sequence generation with streaming conditioning, necessary for live performance scenarios. Our training procedure and architecture composition is also different. For full details, check out our paper.

Please Cite:

Lucy Strauss, Prashanth Thattai Ravikumar, and Matthew Yee-King. 2026. Cross-Modal Sig2Sig Machine Translation with Deep Generative Modeling for NIME Design. Proceedings of the International Conference on New Interfaces for Musical Expression. DOI: 10.5281/zenodo.20784411

BibTeX Entry:

@inproceedings{nime2026_133,
abstract = {},
address = {London, United Kingdom},
articleno = {133},
author = {Lucy Strauss and Prashanth Thattai Ravikumar and Matthew Yee-King},
booktitle = {Proceedings of the International Conference on New Interfaces for Musical Expression},
doi = {10.5281/zenodo.20784411},
editor = {Benedict Gaster and João Tragtenberg and Anna Xambó and Tom Mitchell},
issn = {2220-4806},
month = {June},
numpages = {18},
pages = {1084--1101},
presentation-video = {https://youtu.be/Z7-ySfuF7lg},
title = {Cross-Modal Sig2Sig Machine Translation with Deep Generative Modeling for NIME Design},
track = {paper},
url = {http://nime.org/proceedings/2026/nime2026_133.pdf},
year = {2026}
}

TODO:

  • train instructions
  • run instructions

About

muscle-sound machine translation for live music performance

Resources

License

Stars

0 stars

Watchers

0 watching

Forks

Releases

No releases published

Packages

 
 
 

Contributors

Languages