基于声调核参数及DNN建模的韵律边界检测研究AutomaticMandarin.PDF

文章编号：1003-0077 （2011）00-0000-00 基于声调核参数及 DNN 建模的韵律边界检测研究 1 1 1 1 林举 ，解焱陆 ，张微 ，张劲松 （1.北京语言大学 信息科学学院，北京市 邮编 100083） 摘要：韵律边界对言语表达的自然度和可理解度有着重要作用。韵律建模也是语音合成、语音理解中的重 要方面。本文从相邻声调的相互作用角度出发，提出基于深度神经网络 （DNN）及声调核声学特征的汉语韵 律边界检测方法。该方法首先采用声调核部分的声学特征来计算边界检测相关参数。然后，利用深度神经 网络进行建模。作为对比，实验中采用了以整个音节的声学特征为输入特征的基线系统。结果表明，只使 用调核部分声学特征的系统优于使用整个音节的系统，韵律边界检测正确率相对提高了 4%，这表明本文提 出的汉语韵律边界检测方法的有效性。 关键词：韵律边界建模；声调核；深度神经网络 中图分类号：H01 ，TN912.3 文献标识码：A Automatic Mandarin Prosody Boundary Detecting Based on Tone Nucleus and DNN Model 1 1 1 1 LIN Ju , XIE Yanlu , ZHANG Wei , ZHANG Jinsong (1. College of Information Sciences, Beijing Language and Culture University, Beijing 100083 , China) Abstract: Prosody boundary plays an important role in naturalness and intelligibility of speech of verbal expressions. Thus, prosody modeling is important aspect of speech synthesis and understanding. In this paper, from the angle of interaction of adjacent tones, we propose a method of prosody boundary detecting based on tone nucleus and DNN model. This method firstly calculated the boundary -related parameters by applying the tone nucleus features. Then, the parameters were modeled by deep neural network. For comparison, the baseline system was used the acoustic features of syllable. The experimental results showed that the proposed method used tone nucleus features outperformed the baseline system, with a relative 4% improvement. It demonstrated the efficiency of the proposed method. Key words: Prosody Boundary Modeling; Tone Nucleus; Deep Neural Network 1 引言 人们在进行言语交际的时候，除了字面的文字信息之外，话语韵律变化也是相互传递的 一个重要信息。合理有效地组织话语韵律结构不仅有助于说话者更清楚地表达，而且听话者 也能够更清楚、准确地理解说话人的意图。从语音