多预测子融合实时连续语音识别输出词正误判别

Combination of Multiple Predictors for Correct-Incorrect Classification of Output Words in Real Time Continuous Speech Recognition

本文在采用堆栈译码词网重估输出作为识别最终输出的连续语音识别实时解码条件下,利用决策树方法将多个预测子融合,对识别输出词进行正确和错误的判别。本文首先构造了词后验概率、词长、相邻词的后验概率、词的声学和语言得分等共13个预测子,然后利用决策树方法,通过选择不同的预测子组合方式和适当的决策树建树参数,筛选出预测子的最佳组合,建立优化的决策树进行输出词的正误判别。实验结果表明:利用局域词图计算的词后验概率与词长、相邻词的后验概率等几种实时预测子融合后,对识别输出词的正误判别能力得到提高,并且在实时性和分类效果两个方面优于n-best输出的相应结果,相对于基线系统,则分类错误率下降41.4%。实验结果也表明本文提出的相邻词的后验概率是相对重要的预测子。

Under the decoding strategy of using stack decoding to rescore the word trellis to generate final output, this paper uses decision tree to combine multiple predictors to identify each of recognition output words as correct or incorrect. A series of predictors are constructed, including word posterior probability, word length, word posterior probabihty of neighboring words, 13 in all. Optimal combination of predictors is found and best decision tree is constructed for correct-incorrect classification of output words by testing different combination of predictors and choosing appropriate tree parameters. The experimental results show that the combination of local word posterior probabilities （LWPP） with some of other predictors constructed by this paper, including mainly word length and LWPPs of neighboring words, can give a significant improvement in classifieation performance, and is better in time consumption and quality than the corresponding results from n-best list. Compared with baseline system, the classification error rate getsan improvement of 41.4%. The experimental results also show that posterior probabilities of neighboring words proposed by this paper are among relatively important predictors.