Speech/nonspeech discrimination for speech recognition system under real life noise environments

A method to identify speech segments in noisy environments is presented. It consists of two subsystems: one is for segmentation of a nonstationary signal corrupted by a stationary background noise, and the other is for discrimination of the nonstationary segment, whether it is speech or not. The proposed segmentation method is sensitive enough to detect weak consonants at word initials correctly. Feature parameters effective for speech/nonspeech discrimination are presented. >