Noise Robust Speech Recognition Using Prosodic Information

This paper proposes a noise robust speech recognition method for Japanese utterances using prosodic information. In Japanese, the fundamental frequency (F0) contour conveys phrase intonation and word accent information. Consequently, it also conveys information about prosodic phrase and word boundaries. This paper first proposes a noise robust F0 extraction method using the Hough transform, which achieves high extraction accuracy under various noise environments. Then it proposes a robust speech recognition method using syllable HMMs which model both segmental spectral features and F0 contours. We use two prosodic features combined with ordinary cepstral parameters: a derivative of the time function of log F0 (∆ log F0) and a maximum accumulated voting value of the Hough transform representing a measure of F0 continuity. Speaker-independent experiments were conducted using connected digits uttered by 11 male speakers in various kinds of noise and SNR conditions. It was confirmed that both prosodic features improve the recognition accuracy in all noise conditions, and the effects are additive. When using both prosodic features, the best absolute improvement of digit accuracy is about 4.5%. This improvement was achieved by improving the digit boundary detection by using the robust prosodic information.