Estimation of boundaries between speech units using Bayesian changepoint detectors

Abstract

This contribution addresses the application of Bayesian changepoint detectors (BCD) for the estimation of boundary location between speech units. A novel segmentation approach based on the family of Bayesian detectors using an instantaneous envelope and instantaneous frequency of speech rather than waveform itself is suggested. This approach does not rely on phonetic models, and therefore no supervised training is needed. No apriori information about speech is required, and thus the approach belongs to the class of blind segmentation methods. Due to the small percent of error in signal changepoint location, this method can be also used for tuning boundary location between phonetic categories estimated by other segmentation methods. The average bias between exact boundary location and its estimation is up to 7 ms for real speech.