Tying state-specified rotation using semi-tied covariance transform to model correlations between feature vectors

An acoustic modeling method called the tying State-Specified Rotation is proposed. The method incorporates the merits of State-specified Rotation (SSR) and Semi-Tied Covariance Transform (STC), and overcomes computation and memory problems which are incurred because each state has one full feature-space transform matrix. Based on more precision initial model of SSR, STC is used for tying the feature-space transform matrices among different states. The technique solved the problem that the parameters are overload after SSR, and decreased the number of transform matrices without reducing recognition accuracy. Experimental results on a large vocabulary continues speech recognition task of mandarin show that in comparison to the traditional diagonal modeling technique, the proposed method can get nearly 18.8% word error rate reduction without incurring much computation load during decoding.