New Approach in Transform-Based Speaker Adaptation Using Minimum Classification Error

Automatic speech recognition (ASR) systems work well when trained for a number of specific speakers. However, in most applications there are multiple speakers and they are unknown to the system; performance of ASR system may be degraded because of such speaker variations. This paper examines the use of minimum classification error (MCE) as a preprocessing operation to improve the performance of conventional MLLR (Maximum Likelihood Linear Regression) adaptation. MCE applies its effect by providing better classified components for regression tree in the case of making regression tree on the basis of acoustic space. In this case, distribution of Gaussians will be more smoothing in regression classes. Experimental results on TIMIT database show that 0.42%-0.58% relative improvement is achieved in phoneme recognition rate using our proposed method.