Diff-ETS: Learning a Diffusion Probabilistic Model for Electromyography-to-Speech Conversion.

Area of Science:

• Biomedical Engineering
• Speech Technology
• Machine Learning

Background:

• Electromyography-to-Speech (ETS) conversion enables silent speech interfaces by translating muscle signals into audible speech.
• Current ETS models struggle with naturalness due to limited data and signal noise.

Purpose of the Study:

• To enhance the naturalness of synthesized speech in Electromyography-to-Speech (ETS) conversion.
• To introduce Diff-ETS, a novel ETS model incorporating diffusion probabilistic models.

Main Methods:

• Proposed Diff-ETS, utilizing a score-based diffusion probabilistic model to refine acoustic features from an EMG encoder.
• Evaluated fine-tuning the diffusion model and end-to-end training.
• Compared Diff-ETS against a baseline ETS model.

Main Results:

• Diff-ETS demonstrated a significant improvement in speech naturalness compared to the baseline.
• Objective metrics and listening tests confirmed the enhanced quality of synthesized speech.

Conclusions:

• The proposed Diff-ETS model effectively improves the naturalness of speech synthesized from Electromyography signals.
• Diffusion models offer a promising approach for enhancing ETS conversion quality.