HMM自适应法在应力变异语音识别系统的应用

Adaptation approach based on HMM and its application in G-Force stress speech recognition

为了使应力变异在顽健语音识别系统中能够达到较好的识别效果,研究了基于隐马尔可夫模型(HMM)的自适应技术,提出了将最大后验概率(MAP)和最大似然回归方法(MLLR)用于应力变异语音的自适应中。实验结果表明,与基本系统相比,两种方法均有效地提高系统识别率。以SD为初始模型的最大后验概率方法在150个训练样本时识别效果最好,可以达到90.4%。

In order to achieve the better effect on G-Force stress in robust speech recognition system, this paper, with an emphasis on the self-adaptation techniques based on HMM model, introduces the application of the maximum a posteriori and maximum likelihood linear regression algorithms in G-Force stressful speech adaptation. Compared with baseline system, the two approaches enable such an effective improvement in the performance of system that the best recognition rate is up to 90.4% with 150 training tokens on SD initial model by maximum a posteriori.