连续汉语语音识别中基于归并的音节切分自动机  

Merging-based Syllable Detection Automaton in Continuous Chinese Speech Recognition

作　　者：张继勇[1] sp.cs.tsinghua.edu.cn 郑方[1] sp.cs.tsinghua.edu.cn 杜术 sp.cs.tsinghua.edu.cn[1,1,1,1] 宋战江 sp.cs.tsinghua.edu.cn[1,1,1,1] 徐明星[1] sp.cs.tsinghua.edu.cn[1,1,1,1] 

机构地区：[1]清华大学计算机科学与技术系语音实验室

出　　处：《软件学报》1999年第11期1212-1215,共4页Journal of Software

摘　　要：文章研究并实现了汉语连续语音中的音节自动切分算法─—基于归并的音节切分自动机（merging－basedSyllabledeteCtionautomaton,简称MBSDA）算法．MBSDA算法利用了包括语音的短时能量、过零率和基音周期在内的多种特征参数，把特征参数高度相似的相邻帧（1帧或若干帧）的语音信号进行“归并（merging）”，形成“归并类似段（mergedsimilarsegment，简称MSS）”，它们被认定属于同一音节的相同状态．这些MSS经过一个包含若干状态的“音节切分自动机（syllabledetectionautomaton，简称SDA）”后，输出音节的切分点．每个确定的切分段中所包合音节个数的范围（rangeofsyllablenumber，简称RSN）也由MBSDA算法给出．In this paper, an automatic syllable detection method namely merging-based syllable detectionautomaton (MBSDA) is studied and implemented. The MBSDA uses a variety of features including the frameenergy, the zero crossing rate and the fundamental frequency to merge similar consecutive frames (one or sever-al frames) into one merged similar segment (MSS). The frames in the same MSS are treated as frames of thesame state of a phonetic. These MSSs are passed into a syllable detection automaton (SDA) to give the syllabledetection results. In addition, the MBSDA gives the range of syllable number (RNS) of each definite detectionsegment.

关 键 词：语音识别 归并 音节切分自动机 连续汉语语音 

分 类 号：TN912.34[电子电信—通信与信息系统]