Improving Mandarin Prosody Boundary Detection by Using Phonetic Information and Deep LSTM Model

Automatic prosodic boundary detection is useful for automatic speech processing, such as automatic speech recognition (ASR) and speech synthesis. In this paper, we propose two techniques to improve the boundary detection performance. First, in addition to prosody features (e.g, pitch, duration and energy), phonetic information (word/articulatory information) is integrated into the framework of prosodic boundary detection. We compared two forms of phonetic information: word form and articulatory form. Moreover, boundary detection can be regarded as a sequence labeling task. A deep Long Short-Term Memory (LSTM) is adopted for this task, which replaces the traditional Deep Neural Networks (DNN) model. The experimental results showed that the boundary detection performance can be improved by the additional phonetic information, with relative 5.9% (word form) and 9.8% (articulatory form) improvements respectively in contrast with the system that only used prosody features modeled. The articulatory information and prosody features with deep LSTM achieved the best result, with further performance enhancement from 76.35% to 77.85% (relative 6.3%) compared with that modeled by DNN.