论文部分内容阅读
该文针对维吾尔语说话人之间的发音差异会在一定程度上影响维吾尔语语音识别系统的性能这一情况研究了说话人自适应技术,将目前较为常用的MLLR和MAP以及MLLR和MAP相结合的自适应方法应用于维吾尔语连续语音识别的声学模型训练中,并用这三种方法自适应后的声学模型分别在测试集上进行识别实验。实验结果表明MLLR、MAP以及MAP+MLLR自适应方法使基线识别系统的单词错误识别率分别降低了0.6%、2.34%和2.57%。
This article aims at the fact that Uyghur speaker’s pronunciation difference will affect the performance of Uyghur speech recognition system to a certain extent. The speaker adaptation technology is studied. The most commonly used MLLR and MAP and MLLR and MAP are combined The adaptive method is applied to the acoustic model training of Uyghur continuous speech recognition, and the acoustic models adaptively using these three methods are respectively tested on the test set. Experimental results show that MLLR, MAP and MAP + MLLR adaptive methods reduce the error rate of word recognition of baseline recognition system by 0.6%, 2.34% and 2.57% respectively.