Accelerometer-based real-time voice activity detection using neck surface vibration measurement

Abstract

Speech analysis has a growing number of clinical and industry applications, all of which rely on Voice Activity Detection (VAD). Common VAD applications use microphones, which can be problematic in the presence of background noise and additional voices. Recent studies have utilized accelerometers instead of microphones as voice transducers. As part of a larger research project on impaired speech in the voce disorder spasmodic dysphonia (SD), this study aimed to explore the use of wearable accelerometers to detect speech. These accelerometers would be part of a real-time VAD system embedded in a wearable neck collar for patients with SD. This collar would deliver vibro-tactile stimulation (VTS) to the laryngeal muscles during speech as a therapy for these patients. The aims of this research concerned a) finding the ideal location on the neck to place the accelerometers and b) developing a VAD algorithm that reliably detects the onset and offset of speech based on these accelerometer signals. Methods: 6 healthy adult participants (M/F = 3/3, 26 ± SD = 5.1 years) vocalized 20 sample sentences under 12 conditions from a combination of 3 variables: 1) Normal or slow speed of speech, 2) Three accelerometer attachment locations: thyroid cartilage, sternocleidomastoid, and superior to the C7 vertebra, and 3) Application of VTS during speech in two locations. Time-synchronized acceleration and audio were recorded in each condition. Results: Number of onsets of voice activity and total time voiced, as calculated from application of the VAD algorithm to the acceleration data, were measured. The thyroid cartilage attachment location had over 90% accuracy detecting speech in both measures on average. The average accuracy of the sternocleidomastoid location was below 75% accuracy and was below 15% for C7. Discussion: Placing of an accelerometer at the thyroid cartilage for real-time detection of speech was shown to be feasible. The obtained usability data document that accelerometer signals at this anatomical landmark provide the most reliable data to detect speech. The other two locations tested were too variable in accuracy for implementing VAD. With respect to using the established VAD algorithm in the wearable collar device to treat voice symptoms in spasmodic dysphonia, one needs to state that the algorithm can be improved in robustness to filter out the noise caused by vibration. The use of advanced processing methods such as adaptive filtering will likely deliver the desired result.