Noise and speaker robustness in a Persian continuous speech recognition system

In this paper VTLN speaker normalization, MLLR and MAP adaptation methods are investigated in a Persian HMM-based speaker independent large vocabulary continuous speech recognition system. Speaker and environmental noise robustness are achieved in real world applications for this system. 4 search-based method is used in VTLN to find speaker relative warping factors. The warping factors are applied to signal's spectrum to normalize the variation effect of VTL between speakers. In the MLLR framework, Gaussian mean and covariance transformations in global and full adaptation are experienced. In this method, regression tree based adaptation in batch-supervised fashion is used. Also the standard MAP is experienced as an adaptation method. Combinations of these approaches with CAN robust feature method are evaluated on 4 different tasks. Significant improvement is achieved in the recognition performance in noisy environments such that it makes the system operational in real applications.