Segment specific enhancement of speech characteristics for improving speech intelligibility under adverse listening conditions

Degradation of speech intelligibility in the presence of noise is a serious problem in speech communication devices. We propose an offline speech enhancement technique based on speech recognition and segment-specific modification of speech characteristics. Hidden Markov model based speech recognition is used for locating fricative segments in vowel-consonant-vowel utterances. Speech characteristic modification involved increasing the intensity of the fricative segment with respect to the nearby vowel segment by 12 dB. Intelligibility of modified speech was evaluated by conducting listening tests using 6 normal hearing subjects with speech-spectrum shaped noise as masker. The results of listening tests indicated statistically significant improvements in fricative recognition scores by 4%, 5%, 6%, and 14%, at SNRs of 12, 6, 0, and -6 dB, respectively.