N-best vector quantization for isolated word speech recognition

Speech recognition is performed by utilizing acoustic and linguistic model. The contribution of this paper is improvement of acoustic model. Acoustic model is constructed by hidden Markov model (HMM). HMM has two representations, that are discrete HMM and continuous HMM. The former uses vector quantization (VQ), whereas the latter uses functions such as (mixture) Gaussian distribution. In Viterbi algorithm, VQ has advantage that it only operates by addition. However VQ also has a problem of distortion. This paper attempts to improve recognition precision in discrete HMM with modified VQ that gives multiple outputs for an input.