Hybrid combination of knowledge- and cepstral-based features for phoneme recognition

A new, general, mathematically sound technique is developed to integrate knowledge-based information with standard cepstral features into the formal HMM framework for phoneme recognition. By using these hybrid features, the maximum amount of information contained in the speech signal can be utilised. It is shown that a trivial extension of the statistical models used to model the cepstral features, cannot be used to model the hybrid feature vectors, as this results in a decrease in phoneme recognition accuracy. By using the proposed hybrid technique though, a statistically significant increase in phoneme recognition accuracy is achieved.