Combining stochastic trajectory model and discriminative feature in speech recognizer

Our new approach described in this paper is to embed the stochastic trajectory model into the HMM with a discriminative feature extractor as the front-end which is realized by neural networks followed by a transformation of the MLP outputs. This kind of feature is not only powerful in discrimination against out-of-class confusion data but also it can make it easier to apply the stochastic trajectory model with the HMM. The stochastic trajectory model is an extension of the “trend HMM”. Instead of using one deterministic function for each state to model the temporal evolution of the mean of observation vector, we propose to determine the function parameters statistically. Experiments showed that improvement has been achieved over the original “trend HMM” and the CDHMM based on the cepstral coefficient feature vector