Real-Time Voice Activity Detection Using Neck-Mounted Accelerometers for Controlling a Wearable Vibration Device to Treat Speech Impairment

Abstract

Abstract Speech analysis using microphones can be problematic for Voice Activity Detection (VAD) in the presence of background noise. This study explored the use of wearable accelerometers instead of microphones. We assessed if accelerometers placed on the neck can be part of a VAD system embedded in a wearable collar-like device that delivers vibro-tactile stimulation (VTS) to the larynx during speech as a therapy for patients with the voice disorder spasmodic dysphonia. Specifically, we aimed to a) find the ideal location for placing accelerometers to the neck, and b) develop a VAD algorithm that detects the onset and offset of speech. Six healthy adult participants (M/F = 3/3, age = 26 (5.1)) vocalized 20 sample sentences with and without VTS at three neck locations: 1) thyroid cartilage, 2) sterno-cleidomastoid, and 3) posterior neck above C7. Based on time-synchronized acceleration and audio signals, VAD algorithm identified the Number of Onsets of Speech and Total Time Voiced. The thyroid cartilage attachment location had over 90% accuracy detecting speech in both measures. The average accuracy of the sternocleidomastoid and C7 locations were below 75% and 15% respectively. VAD accuracy decreased with the presence of VTS trials at all locations. We conclude that accelerometer signals due to tissue motion at thyroid cartilage are most suitable for real-time VAD. These findings support the feasibility of accelerometer-based voice detection for the use in medical devices that target speech and voice disorders.