Health monitoring with voice analysis: Acoustic correlates of heart failure, irregular pitch periods, and dysphonia

Abstract

Voice and speech production relies on complex interactions of linguistic and cognitive systems, neuromotor pathways, respiration, and airflow through the vocal tract. Voice can reveal disruptions to any of those systems, so it can be used to non-invasively detect and monitor illness. This thesis examines three interrelated applications of voice analysis for health monitoring. The first application investigates acoustic voice features as a biomarker for acute decompensated heart failure (ADHF), a serious escalation of heart failure symptoms including breathlessness and fatigue. ADHF-related systemic fluid accumulation in the lungs and laryngeal tissues is hypothesized to affect voice acoustics. A set of daily voice samples from 52 patients undergoing inpatient ADHF treatment is analyzed to identify vocal biomarkers for ADHF and examine the trajectory of voice change during treatment. Data from an audio microphone and from a neck-surface vibration sensor are also compared. Results indicate that speakers have more stable phonation, more creaky voice, faster speech rates, and longer phrases after ADHF treatment compared to pre-treatment. These findings motivate work on two additional acoustic features: irregular pitch periods (IPPs), which contribute to the perception of creaky voice, and cepstral peak prominence (CPP), a measure of dysphonia and phonatory stability. To that end, the second application uses voice recordings from healthy speakers and compares the output of an existing algorithm for creaky voice detection to hand labels of IPPs. A perceptually relevant creak probability threshold is determined. These results are useful for voice monitoring of ADHF, since speakers produced more IPPs after ADHF treatment. In the third application, CPP thresholds that distinguish speakers with and without voice disorders are determined separately for continuous speech and sustained vowels using two widely-used voice analysis programs. These normative CPP values provide an objective dysphonia indicator to aid evaluation of voice and other disorders. For example, CPP tended to improve with ADHF treatment for patients whose pre-treatment CPP was relatively low. Together, these projects present a novel method of monitoring ADHF using vocal biomarkers and develop a more-detailed understanding of relevant voice features. Proposed future work includes prospective at-home monitoring of patients at risk for ADHF.