Minimum classification error optimization of word recognizers using evolution strategies

This paper introduces a new approach to speech recognizer design which comprehensively optimizes the front-end analyzer and the back-end classifier. The algorithm is based on the principle of minimum classification error rate (MCE) in which a discriminative training procedure is performed for directly minimizing the recognition error rate. Due to the fact that the error rate of a given finite set of data is a piecewise-constant function of the recognizer parameters, the proposed algorithm uses evolution strategies (ES) for successive modification of recognizer parameters. So we need not derive a smooth and differentiable loss function as misclassification measure as in other well known MCE-optimization approaches. In addition, the ES-based method can be applied to any existing recognition system and does not require modifications of the current algorithms. The MCE/ES optimization was tested on a speaker-independent whole word recognition task in which the front-end feature extraction transformation matrix was optimized. A recognition rate of 99.3% (98.0%) was achieved using a discrete HMM- (DTW-) recognizer compared to 87.7% (76.3%) without optimization