Onset-based segregation of stop consonants

Summary form only given. Speech segregation from acoustic interference is a challenging task. Previous systems have successfully dealt with voiced speech, but cannot handle unvoiced speech. We study the segregation of stop consonants, which contain significant unvoiced signals. We propose a novel method that employs onset as a major cue to segregate stop consonants. Our system first detects stops through onset detection and Bayesian classification of acoustic-phonetic features, and then performs grouping based on onset coincidence. The system has been tested and performs well on utterances mixed with various types of interference.