汉语大词汇量连续语音识别中混淆网络算法的研究

在汉语大词汇量连续语音识别中,以往基于最大后验概率准则解码得到的识别结果具有最小的句子错误率,为了得到字错误率最小的识别结果,可以采用最小贝叶斯风险解码策略,通过将识别输出的word lattice转换成为混淆网络以得到最小字错误率的识别结果。在以往混淆网络算法的基础上,根据汉语语言的特点,提出一种改进的构造混淆网络的算法。基于863测试语音库进行的实验表明,与最大后验概率识别结果和以前的两种混淆网络算法的识别结果相比,改进的混淆网络算法有效地降低汉语大词汇量连续语音识别结果的字错误率。

Research on Confusion Network Algorithm Based on Minimum Bayes Risk Decision Rule

In mandarin large vocabulary continuous speech recognition,the recognition result with minimum word error rate(WER) can be obtained by using minimum hayes risk(MBR) decoding strategy.One method of MBR de- coding is that the word lattice can be transformed into confusion network in order to achieve the recognition result with minimum WER.According to the characteristic of Chinese linguistics,we proposed an improved algorithm of constructing confusion network for mandarin large vocabulary continuous speech recognition.Evaluated on the Chi- nese 863 speech corpus,experimental results show that compared with the MAP one-best decoding and previously proposed two confusion network algorithms,our improved algorithm effectively reduces the WER of recognition out- put hypothesis.