HMM-based speech recognition using dynamic spectral feature

A description is presented of a fenonic Markov-model-based speech recognizer that can evaluate not only instantaneous spectral features but also dynamic spectral features, without requiring so many parameters and training data as some conventional models representing these features. The author first shows that the correlation between the features is very small. On the basis of this result, the features are vector-quantized separately and then independently evaluated in a multiple-feature-based fenonic Markov model. This recognizer was tested in speaker-dependent and speaker-adaptation-based isolated word recognition tasks using 150 confusable Japanese words. For each task, the recognition error rate was 45-75% lower than that of the conventional fenonic Markov-model-based recognizer, which evaluates only instantaneous features