Automatic detection of Chinese accent-index based on approximation-ratio

For a TTS system, to synthesize speech with better prosody, accent information is expected to be involved. Therefore, we defined a set of accent indexes (AI) to represent the variances of accent in Chinese speech, and proposed a novel method to automatically annotate Chinese speech with the AI. In the method, a parameter, named approximation-ratio, was used to numerically indicate the accent of a prosodic unit. And the value of AI was the discretization of the approximation-ratio. One corpus was annotated with AI by the method. And with the corpus, a refined prosody parameter prediction model was built. The experimental results showed that prosody parameters predicted by the refined model were closer to those of real speech than the former model without AI. Further, a perceptual evaluation showed that the accent manifestation generated by the AI-ready synthesizer was distinguishable and acceptable.