Speech Recognition Based on Discriminative Classifiers

This work focuses on techniques for using discriminative classifiers such as the support vector machine (SVM) in speech recognition. We review previously proposed architectures and perform experiments with one of them. The results show that using SVMs instead of Gaussian mixtures can bring significant improvements in accuracy, but the computational cost is very high. To circumvent this problem, we propose a new architecture that achieves higher accuracy with a reasonable cost. Preliminary results for a spelling task indicate a decrease of 24% in word error rate over a generative HMM-based system.