サブワード単位に基づく音声認識

抄録

Large vocabulary, isolated word recognition requires a large amount of training data proportional to the vocabulary size to characterize each individual word model. A subword-unit-based approach is a more viable aternative than the word-based approach to overcome the problem of the training data size, since different words can share common segments in their representations in the former approach.
This paper deals with a word recognition system where the subword-unit-based approach is employed. Subword boundaries within a word are detected by finding peaks of a sequence of power spectral distance or a sequence of delta cepstra associated with the word and the resulting sequence of subword units is deciphered into the original word by means of concatenated hidden Markov models of isolated words. This system attains average word recognition rates over 98%.