Baum-welch Training for Segment-based Speech Recognition

The use of segment-based features and segmentation networks in a segment-based speech recognizer complicates the probabilistic modeling because it alters the sample space of all possible segmentation paths and the feature observation space. This paper describes a novel Baum-Welch training algorithm for segment-based speech recognition which addresses these issues by an innovative use of finite-state transducers. This procedure has the desirable property of not requiring initial seed models that were needed by the Viterbi training procedure we have used previously. On the PhoneBook telephone-based corpus of read, isolated words, the Baum-Welch training algorithm obtained a relative error reduction of 37% on the training set and a relative error reduction of 5% on the test set, compared to Viterbi trained models. When combined with a duration model, and more flexible segmentation network, the Baum-Welch trained models obtain an overall word error rate of 7.6%, which is the best result we have seen published for the 8,000 word task.