Liquid Neural Dynamics for Temporal Signal Modeling: Applications in EMG-Based Speech Synthesis

Overview

This is a research-driven project aimed at developing a novel system for silent speech recognition using mainly Electromyography (EMG) signals & Liquid Neural Networks. The goal is to recognize and synthesize speech based on the electrical activity of facial muscles without the need for vocalization.

Paper

You can find the paper associated to this code here.

Setup

Recommended Setup (using Miniconda)

Just run the following command to install the recommended setup:

./setup.sh --env-name IAR_SSR --python-version 3.10

Venv

Alternatively, you can create a virtual environment using venv:

./setup.sh --env-name IAR_SSR --python-version 3.10 --venv

System Specifications (for reproducibility)

• Model: ROG Zephyrus G16 GU605MZ
• GPU: NVIDIA GeForce RTX 4080
  • Memory: 12 GB GDDR6
  • AI Performance: 542 AI TOPs
  • Boost Clock: 1920 MHz
  • Dynamic Boost: 115W
• Processor: Intel® Core™ i9-185H Ultra
  • Base Frequency: 2.3 GHz
  • Boost Frequency: Up to 5.1 GHz
  • Cores/Threads: 16 cores, 22 threads
  • Cache: 24 MB
• Memory (RAM): 32 GB LPDDR5X 7467 MHz (16 GB x 2)
• Storage: Dual SSD M.2 NVMe PCIe® 4.0
• Operating System: Ubuntu 24.04
• Kernel Version: 6.11.0-19-generic
• NVIDIA Driver Version: 550.120
• CUDA Version: 12.4

License

This project is licensed under the Apache License (Version 2.0).

See LICENSE for details.

Citation

If you use or reference this work, please cite it as follows:

@misc{maubras2025emg,
  title={Liquid Neural Dynamics for Temporal Signal Modeling: Applications in EMG-Based Speech Synthesis},
  author={Juan Maubras, Lili M. Rampre and Alexandre Pitti and Sylvain Reynal},
  year={2025},
  howpublished={\url{https://github.com/Elesdes/EMG2Speech-LiquidNet}},
  note={Unpublished manuscript},
}