采用上下文相关音素HMM的连续语音识别  

Continuous Speech Recognition Based on the Context-Dependent Phonetic HMM

作　　者：谢锦辉[1] 

机构地区：[1]华中理工大学

出　　处：《通信学报》1994年第2期83-87,共5页Journal on Communications

摘　　要：本文简要讨论了在基于ＨＭＭ的连续语音识别系统中怎样选取基本语音单元的问题，介绍了在欧洲Ｐｏｌｙｇｌｏｔ（Ｅｓｐｒｉｔ２１０４）课题下在法国ＬＩＭＳＩ－ＣＮＲＳ建立的基于上下文无关（ｃｏｎｔｅｘｔｉｎｄｅｐｅｎｄｅｎｔ）音素ＨＭＭ的一个基本连续语音识别系统。然后，本文详细给出了利用左或右上下文相关（ｌｅｆｔｏｒｒｉｇｈｔｃｏｎｔｅｘｔ－ｄｅｐｅｎｄｅｎｔ）音素ＨＭＭ，作者对上述系统改进后进行的连续语音识别。采用美国语音库ＤＡＲＰＡ－ＲＭ１，在不考虑句法（ｓｙｎｔａｘ）信息时，我们获得了连续语音（句子）识别时词识别率大约３～１０个百分点的明显提高。实验是在法国ＬＩＭＳＩ－ＣＮＲＳ进行的。最后，本文还给出了进一步改进系统性能的努力方向。This paper briefly discusses the problem of how to choose basic speech units in HMM-based contiruous speech recognition systems, and introduces a baseline contintous speech recognition system, based on the context-independent phonetic HMM, established in LIMSI-CLRS, France, under the Polyglot (Esprit project 2104), Then, the paper details the continuots speech recognition on an improved system, made by the author, for the above-mentioned system by using the left or right context-dependent phonetic HMM.Using the American speech data base DARPA-RM1 and considering no syntax information, we obtain the distinct improvements of about 3-10 percent of word recognition accuracy in continuous speech (sentences) recognition. The experiments are done in LIMSI-CNRS. France. Finally, some aspects for improving further the performance of the system are pointed out.

关 键 词：音素HMM 连续语言 元音识别 

分 类 号：TN912.34[电子电信—通信与信息系统]