基于后验概率解码段模型的汉语语音数字串识别

Mandarin Digit String Recognition Based on Segment Model Using Posterior Probability Decoding

通过对语音解码的分析指出了基于似然概率解码的连续语音识别的局限性,并给出了三种基于后验概率段模型(Segment Model,SM)的语音解码方法.这三种方法成功地运用于随机段模型(Stochastic Segment Model,SSM),使误识率比基线系统下降了11%;与此同时还给出了段模型的快速算法,使算法的计算复杂度降到了与隐马尔可夫模型(Hidden Markov Model,HMM)相同的数量级,满足了实用要求.

The decoding algorithms of most continuous speech recognition systems are based on the likelihood score now. However, the likelihood score is only an approximate of the posterior probability and will lead to a suboptimal solution in continuous speech recognition task. In this paper, three Segment Model（SM） decoding methods based on posterior probability are introduced and successfully implemented on a Stochastic Segment Model（SSM） based system. SSM is one kind of segment models. The new decoding methods achieve 11% error rate reduction compared with the baseline system. In the meantime, a fast algorithm for SM is also proposed, which can reduce the computation complexity of the above algorithms to the same level as that of HMM and meet the requirement of real-time applications.