Improving the Robustness of Persian Large Vocabulary Continuous Speech Recognition System for Real Applications

In this paper vocal track length normalization (VTLN) with adaptation methods, MLLR and MAP were investigated to making robust Persian HMM-based speaker independent large vocabulary continuous speech recognition system. The robustness for speaker and environmental noises were achieved in real world applications in this system. In VTLN method, a line-search based approach was used in order to find speakers relative warping factors. The factors were applied to signal´s spectrum to normalize the variations in vocal track length between speakers. In the MLLR method, Gaussian mean and variance transformations in full adaptation were experienced. In this method regression tree-based adaptation in supervised fashion was used. Also the standard MAP was experienced as an adaptation method for compensate speaker and environment variations. Combinations of these approaches were evaluated on 4 different noisy tasks. We could achieve the significant improvement in the recognition performance in noisy environments as it makes our system operational in real applications