摘要
从整体上介绍了汉语连续语音识别系统的基本原理 ,并重点对声学和语言两个层面的建模与搜索策略进行了分析 .在对传统帧同步搜索算法进行研究的基础上 ,提出了基于统计知识的帧同步搜索算法 SKB- FSS.它包含了三个主要的功能层次 :基于归并的音节切分自动机产生确定的搜索边界点 ,由统计得到的差分状态驻留信息控制搜索过程中的状态转移 ,利用词搜索树控制音节候选的扩展规模并根据动态前向预测的方法进行合理而及时的路径剪枝 .实验结果验证了该搜索策略的有效性 .
In this paper, the principle of a Chinese continuous speech recognition system is introduced, the modeling and search strategies of its acoustic layer and language layer are also discussed in detail. On the basis of the research on the traditional frame synchronous search algorithm, the statistical knowledge based frame synchronous search (SKB FSS) algorithm is proposed. It contains three principal functional modules, generating definite search boundaries by a merging based syllable detection automaton, controlling state transitions by the statistical differential state dwell information, and restricting the syllable expansions by a word search tree and pruning unpromising paths by the dynamic forward prediction. The experimental results show the validity of the novel search strategies.
出处
《自动化学报》
EI
CSCD
北大核心
2000年第4期470-477,共8页
Acta Automatica Sinica
关键词
汉语连续语音识别系统
知识导引
语言处理
Continuous speech recognition, statistical knowledge based frame synchronous search, differential state dwell distribution.