The Improvement Of Text Processing Technology For Speech Synthesis

In addition to resolve the currently widespread, affecting synthesized speech intelligibility and naturalness of sound and rhythm structure problems in chinese speech synthesis system, and improve the performance of synthesized speech, especially make it with accurate, vivid semantic expressiveness, need text analysis module can output more rich linguistic information, and use this information to synthesize a more accurate and vivid voice. Therefore, this article focus on:speech database and its rhythm annotation, and the synthesis of the front-end module, to start the study.On the one hand, this paper will raise text object of the speech synthesis from the statement level up to chapter level, and build the chapter level broadcast news reporting database; On the other hand, improve the front-end speech synthesis:text processing module to some extent. The main work as follows:1. Select the news broadcast corpus for research/processing material, considering the demand of calculation modele and the characteristics of the sample, based on the previous work, develop a set of prosody annotation scheme of chapter-level, which make the new rhythm description include:prosodic hierarchy, stress, mood and tone, enrich and improve the existing rhythm description; and implement the annotation work according to the scheme, and then build the chapter level broadcast news reporting database.2. Based on the work flow of original text processing module and the part of the text processing module, compare its word segmention module with other modules', then replace the word segmention module based on the comparison, and then train a new text-processing model based on the binary grammar, the core is training a new rhythm structure prediction module.Then, through the within and out of test experiments, prove the new model to achieve a better effect of rhythm structure prediction, and text-processing effects have been improved to some extent.