Spellmode recognition based on vector quantization

Abstract

Vector quantizers have traditionally been used in speech recognition systems as a pre-processor for sophisticated algorithmes 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 such VQ-based systems is the lack of temporal information in the recognition algorithm. Recent and new variations of the VQ-based systems incorporating temporal information are described. The principal new variation introduced here is a conditional histogram technique which incorporates relative likelihoods of successive codewords into the distortion measure used in the VQ recognition algorithm. Several VQ-based recognition algorithms are applied to the recognition of spoken letters of the English alphabet, a subset of the IBM Spellmode vocabulary. Simulation results highlight the relative merits of the algorithms.