A comparative study of speech segmentation and feature extraction on the recognition of different dialects

Speech is the most intuitive way of communication, except for mutes and deaf-mutes. Hong Kong, a multicultural society, is an ideal place to develop a multi-lingual (Cantonese, Mandarin, and English) ASRVR system. Once this happened, numerous techniques were explored of the three major stages of speech data: segmentation, preprocessing, and recognition. The speech segmentation process includes the convex hull method for finding S/sub min/, SVF for S/sub max/, and normal decomposition for S/sub opt/. The sub-word boundaries are considered by the LBDP-based method. The three proposed feature extraction methods for speech preprocessing are MFCCs, RASTA, and FBDCCs. Speech recognition is the final phase of our system. Approaches including Navie Bayesian classification, HMM with Viterbi algorithm, and backpropagation with a two-hidden-layer structure are studied in this paper. The best performance for multi-lingual recognition in our system can be found as applying MFCCs into HMM, after the segmentation procedure.