Prototype-based minimum error training for speech recognition

A key concept in pattern recognition is that a pattern recognizer should be designed so as to minimize the errors it makes in classifying patterns. In this article, we review a recent, promising approach for minimizing the error rate of a classifier and describe a particular application to a simple, prototype-based speech recognizer. The key idea is to define a smooth, differentiable loss function that incorporates all adaptable classifier parameters and that approximates the actual performance error rate. Gradient descent can then be used to minimize this loss. This approach allows but does not require the use of explicitly probabilistic models. Furthermore, minimum error training does not involve the estimation of probability distributions that are difficult to obtain reliably. This new method has been applied to a variety of pattern recognition problems, with good results. Here we describe a particular application in which a relatively simple distance-based classifier is trained to minimize errors in speech recognition tasks. The loss function is defined so as to reflect errors at the level of the final, grammar-driven recognition output. Thus, minimization of this loss directly optimizes the overall system performance.