利用段长信息的改进连续语音识别算法

Improved algorithm with duration information for continuous speech recognition

以基于段长分布的隐含Ｍａｒｋｏｖ模型为基础，提出了一种改进的帧同步束搜索连续语音识别算法，使段长信息在连续语音识别中得到充分有效的利用，并与原算法作了比较。在该算法的具体实现上，建立了状态的局部路径组和全局活跃路径两种数据结构，并采用了帧同步快速实时算法来处理局部路径的剪枝和跳转。说明了段长参数的估计方法。介绍了非特定人大词汇量连续语音识别的实验，实验结果表明，利用段长信息改进识别算法比原识别算法字的误识率降低了６％。

Based on the duration distribution based hidden Markov model (DDBHMM), an improved frame synchro nous beam search algorithm for continuous speech recognition is presented, which efficiently utilizes the duration information. The algorithm is compared with the old one. On realizing the algorithm, two data structures are built and the frame synchronous realtime algorithm is used to deal with the pruning and jumping of the local path. The approach on estimating the duration parameter is discussed. The experiment on speaker independent large vocabulary continuous speech recognition is introduced, the word error rate is reduced by 6%, which shows that the duration information improves the performance of the system.