Speech Recognition of Phones Using Feature Streams

Artificial neural networks (ANNs) have been used to classify phonetic features in speech. The feature streams from the ANNs are used here as the observations for Hidden Markov Models (HMMs). Using such observations allows us to build a competitive speech recogniser. This recogniser is compared to a similar recogniser that was trained on melfrequency cepstral coefficients (MFCCs). While the cepstral HMMs perform better, the feature HMMs do exhibit their potential for further development. 1 Introduction HTK (Young et al., 1996) is a popular tool for constructing speech recognisers. HTK is used to train HMMs to recognise speech according to a specified dictionary. The normal input for training the HMMs is MFCCs (Rabiner and Juang, 1993). These have proven to be one of the best methods for training speech recognisers. A new area of research involves the possibility of training the HMMs on feature streams. That is, instead of giving the cepstral coefficients for each frame of speech, give ..