A new HMM/LVQ hybrid algorithm for speech recognition

Abstract

It is shown that by combining the discriminative power of learning vector quantization (LVQ) training algorithms and the capability of modeling temporal variations of a hidden Markov model (HMM) into a hybrid algorithm, the performance of an HMM-based recognition algorithm is significantly improved. The hybrid algorithm was tested in a multispeaker, isolated word mode, using a highly confusable vocabulary consisting of the nine English E-set words. The average word accuracy for the original HMM-based system was 62%. When the LVQ classifier was incorporated, the word accuracy increased to 81%.