Performance improvement of speech recognition system using microphone array

In this paper, we present some methods to improve the performance of microphone array speech recognition system based on Limabeam algorithm. For improving recognition accuracy, we proposed weighted Mahalanobis distance (WMD) based on traditional distance measure in a Gaussian classifier and is a modified method to give weights for different features in it according to their distances after the variance normalization. Experimental results showed that Limabeam adopted weighted Mahalanobis distance measure (WMD-Limabeam) improves recognition performance significantly than those by original Limabeam. In compared experiments with some other extended versions of Limabeam algorithm such as subband Limabeam and N-best parallel model for unsupervised Limabeam, we could see that the WMD-Limabeam show higher recognition accuracy. In cases of the system that adopted WMD, we obtained correct word recognition rate of 89.4% for calibrate Limabeam and 84.6% for unsupervised Limabeam, 3.0% and 5.0% higher than original Limabeam respectively. This rate also results in 9.0% higher than delay and sum algorithm.