Comparing and Combining Generative and Posterior Probability Models: Some Advances in Sentence Boundary Detection in Speech

We compare and contrast two different models for detecting sentence-like units in continuous speech, using both acoustic and lexical information. The first approach is based on hidden Markov sequence models based on N-grams, uses maximum likelihood estimation, and model interpolation to combine different representations of the data. The second approach models the posterior probabilities of the target classes, is therefore discriminative, and integrates multiple knowledge sources in the maximum entropy (maxent) framework. Both models combine lexical, syntactic, and prosodic information. We develop a technique for integrating pretrained probability models into the maxent framework, and show that this approach can improve, if only slightly, on an HMM-based state-of-the-art system for the sentence-boundary detection task. A much more substantial improvement is obtained by combining the posterior probabilities of the two systems.