【摘要】 为了提高非特定人汉语数码串连续语音识别系统的识别速度 ,对系统进行了细致的研究 ,有针对性地提出了一种新的快速识别算法 .通过对隐含马尔可夫模型输出概率密度函数运算的简化 ,以及采用结合段长信息的束搜索剪枝算法 ,在保证识别精度的情况下 ,使识别系统对不定长数码串平均识别时间从 0 .92 s减少为 0 .11s,而串识别正确率仅从 94.3%下降到 94.0 % ,大大地提高了系统的整体性能更多还原

【Abstract】 In this paper, a new efficient recognition algorithm is proposed based on speaker independent mandarin connected digit speech recognition system to enhance the integrated performance of this system. The average time of the unknown length connected digit recognition decreases from 0.92 second to 0.11 second and the recognition accuracy just degrades from 94 3% to 94 0% by simplifying observation probability computation of hidden Markov model and adopting improved beam search pruning algorithm combined with duration information.更多还原