论文部分内容阅读
提高汉语合成语音的自然度的关键是要建立一个完善的汉语韵律模型.本文以连续的广播语言为研究对象,对汉语中语句重音对韵律特征参数的影响进行了初步探讨,分析了不同语句重音条件下音长和音高的变化及其相互关系,指出:(1)音高是语句重音的基本表达手段,随着语句重音级别的提高,音高分布曲线向高频方向推移。(2)在连续语流中词被‘重读”和“轻读”的情况下,音长分布出现双峰,表示它们的音长有的受语句重音的影响,有的不受语句重音的影响。(3)在“正常”、“重读”和“轻读”王种情况下,音高和音长的相互关系分别是:不相关、负相关和正相关,证实了汉语语句重音音高和音长之间的互补关系。这些研究结果为汉语会成系统中韵律模型的建立提供了基础。在此基础上,本文又用神经网络对连续语音的语句重音进行了部分标注,开集中的分类结果正确率为63%,对语音数据中重音等级的自动标注方法作了探索。
The key to improve the naturalness of Chinese synthesized speech is to establish a complete Chinese prosody model. In this paper, the continuous broadcast language is taken as the research object, and the influence of the sentence stress on the prosodic feature parameters in Chinese is discussed. The changes of pitch length and pitch under different sentence stress conditions are analyzed, and the relationship between them is pointed out: (1) High is the basic expression of sentence accent means, with the increase of sentence accent level, pitch distribution curve to the high-frequency direction. (2) In the case of continuous speech, words are “re-read” and “light-read”, there are bimodal distribution of the sound length, which indicates that their sound length is affected by the sentence stress, and some are not influenced by the sentence stress . (3) In the “normal”, “reread” and “light read” king case, the relationship between pitch and sound length are: irrelevant, negative correlation and positive correlation, confirms the Chinese sentence stress pitch and length of sound The results of these studies provide the foundation for the establishment of the prosodic model in the Chinese system.On this basis, this paper also partially annotates the sentence accent of continuous speech with neural network, the correctness of the classification results in the open set Was 63%. The method of automatic annotation of accent level in speech data was explored.