汉语连续语音识别结果评价算法研究

A Research on Mandrin Speech Recognition Result Evaluation Algorithm

在汉语语音识别中,由于汉语构词的特点,使得基于词的汉语语音识别结果评价不准确。论文对于传统连续语音识别结果评价算法进行了改进,提出了一种基于字词混合的汉语连续语音识别结果评价算法,可以有效完成基于词的识别结果评价,同时也将识别结果评价由四种情况(正确、替代、插入、删除)扩展到六种情况(增加了插入式替代和删除式替代),可以为语音识别的后处理提供更多有用的信息。实验表明,本文所提算法可以有效降低传统评价算法带来的虚假错误。

In mandarin speech recognition, traditional word based recognition result evaluation algo- rithm is not accurate because of the unique characteristics of Chinese word-formation. In this paper, in order to evaluate word based recognition result more efficiently, we propose an improved mandarin speech recognition evaluation algorithm based on mixed words. Two new situations of recognition re- sult evaluation, which are inserted substitution and deleted substitution, are added to the four traditional situations of evaluation including hit, substitution, insertion and deletion. Our experiments show that our proposed algorithm can provide more useful information for post-processing of speech recognition, and effectively reduce false error caused by traditional evaluation algorithm.