Vector quantization based upon bandpass filtering applied to speech recognition

A method based on vector quantization and bandpass filtering for fast speaker-independent discrete utterance recognition is presented. The time alignment based on voice excitation of bandpass filtering was proposed to compress the speech data from bandpass filter and reduce the amount of computation in vector quantization. Two alternatives for distortion measure were tested for generating codebooks and classification. The approach proposed is faster than that of vector quantization based on linear prediction code in generating codebooks and classification, and uses approximately 2500 bits to represent each utterance in the recognition vocabulary. Preliminary limited speaker-independent testing on 10 digits and 10 Chinese city names has demonstrated excellent performance.<
>