We first analyze the distribution of three prosodic cues for continuous speech: segmental fundamental frequency, intensity and duration. Second, we propose a statistical prosodic model and show how it can be included in a Markov source-based recognizer. Finally, we present the performance results of different prosodic models in a very large vocabulary continuous speech recognizer.
Keywords: prosody, suprasegmental features, Markov source-based continuous speech recognizer, very large vocabulary