Discriminative training based on frame level likelihood normalization and its application for speech and speaker recognition

We present a method for discriminative estimation of parameters of Gaussian distribution based classifiers and its application to speech and speaker recognition. The objective of this method is to maximize the normalized likelihood of the design samples. In contrast to other discriminative algorithms such as minimum classification error/generalized probabilistic descent (MCE/GPD) and maximum mutual information (MMI), the objective function is optimized using a modified expectation-maximization (EM) algorithm. The evaluation experiments using both clean and telephone speech showed improvement of the recognition rates compared to the maximum likelihood estimation (MLE) training method, especially when the mismatch between the training and testing conditions is significant. Compared with the MCE/GPD discriminative method, our algorithm showed better performance in both the speech and speaker recognition tasks