Chinese prosodic phrasing with the source-channel model

The prosodic phrasing is a classic problem in nature language process, which is not only useful for text-to-speech(TTS), but for speech recognition, statistic machine learning etc.. This paper introduces and discusses the source-channel model for Chinese prosodic phrasing. Based on the basic idea, the hidden Markov model (HMM) and the improved source-channel model are both used to describe the phrasing problem. In the improved source-channel model, maximum entropy model is used, and the discriminative training is introduced. And the rhythm model is proposed to describe the property of the utterance. The phrase-length model and the foot-pattern model both are used to describe the rhythm model, respectively. The experiments show that this approach achieved a good performance for prosodic phrasing. The improved source-channel model achieve a better performance than the hidden Markov model. And the foot-pattern model is the better one as a rhythm model.