Estimation of Speech Features Using a Wearable Inertial Sensor

Area of Science:

• Biomedical Engineering
• Speech Processing
• Digital Health

Background:

• Speech features are emerging digital biomarkers for psychiatric and neurocognitive diseases.
• Microphone-based speech recording faces limitations like privacy leakage and environmental noise, hindering clinical use, especially for long-term monitoring.

Purpose of the Study:

• To investigate the feasibility of extracting speech features from chest wall acceleration.
• To assess sternum acceleration as a viable alternative to microphones for capturing speech characteristics.

Main Methods:

• Ten healthy subjects performed speech tasks (sentence repetition, reading) under varying rates and loudness.
• Simultaneous recording of voice (microphone) and chest vibrations (sternum accelerometer).
• Extraction and comparison of acoustic and time-related prosodic features from both signals using linear fit and correlation analysis.

Main Results:

• High agreement was found between acceleration and microphone features for all six time-related prosodic features.
• Good agreement was observed for 19 (task 1) and 17 (task 2) acoustic features, primarily related to loudness and pitch.
• Sternum acceleration effectively tracks time-related speech prosody, loudness, and pitch.

Conclusions:

• Deriving speech features from sternum acceleration is feasible.
• This method offers a promising, privacy-preserving approach for developing digital biomarkers.
• Potential applications exist for diseases linked to prosodic and loudness speech characteristics.