Pitch-Synchronous Biomarkers for Voice-Based Diagnostics

Background: Voice analysis combined with artificial intelligence (AI) is rapidly becoming a vital tool for disease diagnosis and monitoring. A key issue is the identification of vocal biomarkers, that are quantifiable features extracted from voice to assess a person’s health status or predict the likelihood of certain diseases. Currently, the biomarkers are extracted using pitch-asynchronized methods that need improvement. Methods: Based on the tim- bron theory of voice production, a pitch-synchronous method of vocal biomarker extraction is proposed and tested. Results: As examples, the methods are applied on the ARCTIC speech databases published by Carnegie Melon University, and the Saarbrucken voice database for voice diagnostics. From the ARCTIC database, a complete set of formant parameters and timbre vectors for all US English monophthong vowels are presented. The timbre distances among all US English monophthong vowels are presented, showing the richness and accuracy of information contained in those biomarkers. By applying pitch-synchronous methods on the voice recordings in the Saarbrucken voice database, accurate and reproducible measure- ments of timbre vectors, jitter, shimmer, and spectral irregularity are shown. Furthermore, methods of detecting glottal closing instants from voice signals are discussed and demon- strated. Conclusions: The biomarkers extracted using pitch-synchronous analysis contain abundant, accurate, objective, and reproducible information from the voice signals that could improve the usability and reliability of voice-based diagnostics.