Model for Speech Recognition Based on Multiple Time Scale Features

The model explicitly models the correlation among successive frames of speech signals in segment scale by using segmental features representing contours of spectral parameters. By using the proposed segmental feature dependent non-stationary time series model, the new model not only achieves the modeling of correlation between different scale features but also implicitly models the correlation among neighboring frames in frame scale via parametric mean trajectory function. A modified Viterbi algorithm based on joint statistical distance of multiple time scale features is proposed, and a algorithm based on the maximum likelihood criteria for estimating the model parameters is also proposed in the training. Experimental results show that the new model achieves better performance than the standard HMM and the trended HMM.