Conditional histogram vector quantization for spellmode recognizer

Abstract

In speech recognition, vector quantizers have traditionally been used as a pre-processor for sophisticated algorithms such as hidden Markov modelling (HMM) or dynamic time warping (DTW). Recently, simpler systems based more directly on vector quantization (VQ) have been proposed for recognizing isolated words with small vocabularies. The major problem with these simple algorithms is the lack of temporal information. This paper describes a conditional histogram technique which incorporates temporal information by considering the relative likelihoods that certain codewords follow others. Simulation results show that this approach produces better decoding results than the simple VQ algorithm with similar complexity.