Speaker-independent isolated-digit recognition based on hidden Markov models and multiple vocabulary specific vector quantization

A discrete hidden Markov model (HMM) system recognizer using word-specific vector quantization is described. The word-specific VQ approach is suggested as an alternative to universal codebook vector quantization. The set of word-specific VQ index sequences is connected to each of the word-specific HMM models. For speaker-independent isolated digit recognition with a studio recorded database, a performance of 99.5% was obtained for the word-specific codebook VQ-HMM recognizer, which is an improvement of 2% when compared to a universal codebook VQ-HMM recognizer tested on the same speech database.<>