| 摘要: |
| 文章从声学基元和词法树两个方面对连续语音识别和汉语语音听写机中声学层面的搜索策略进行了分析,提出了基于统计知识的帧同步搜索算法和基于词法约束的词搜索树结构,构成了声学层面的双层搜索网络.算法中利用了统计知识,包括声学层面的差分状态驻留信息和特征变化量信息等.实验结果表明,基于知识的搜索策略使连续语音识别的性能提高了36.6%.文章还介绍了N-Gram统计语言模型的修正退化频度估计算法和搜索算法原理.通过对多年研究成果的分析,实现了一个汉语语音听写机的引擎,并在PC机上构建了两个系统：非特定人汉语语音听写机 |
| 关键词: 连续语音识别,汉语语音听写机,搜索策略,基于统计知识的帧同步搜索算法,差分状态驻留,特征变化量,词搜索树,双层搜索网络,汉语语音听写机引擎,语音命令. |
| DOI： |
| 分类号: |
| 基金项目:本文研究得到国家863高科技项目基金资助. |
|
| Research and Implementation of the Techniques for Chinese Dictation Machines |
|
ZHENG Fang,MU Xiao-long,XU Ming-xing,WU Jian,SONG Zhan-jiang
|
| Abstract: |
| In this paper, the search strategies in the acoustic layer of the CSR (continuous speech recognition) and the CDM (Chinese dictation machine) are addressed in two aspects, the acoustic recognition unit and the syntax-constrained word search tree. The SKB-FSS (statistical knowledge based frame synchronous search) algorithm and the syntax-constrained WST (word search tree) structure are proposed, they form the TLSN (two-level search network) in the acoustic layer. The statistical knowledge used by the algorithm includes differential state dwell distribution, the feature difference sum and so on, which result in an improvement of 36.6% in CSR. The principles of a modified back-off estimation algorithm and the search algorithms for the N-gram based language models are also introduced. Finally, by integrating the authors' techniques, a Chinese dictation machine engine (CDME) is implemented. A speaker-independent CDM text editor named ST97 and a voice command system named CMD97 are established for personal computers (PCs) based on the CDME. |
| Key words: CSR (Continuous speech recognition), CDM (Chinese dictation machine), search strategy, SKB-FSS (statistical knowledge based frame synchronous search) algorithm, differential state dwell, feature difference sum, WST (word search tree), TLSN (two-level sear |
