汉语连续语音识别系统中三音子模型的优化

Refining triphone model in mandarin continuous speech recognition

为了更准确地估计状态聚类前有调三音子的模型参数,从而提高聚类后捆绑状态的精度及系统的识别性能,针对汉语连续语音识别中,有些有调三音子的训练样本数非常少,而其对应的无调三音子的训练样本数相对较多的情况,提出用其对应的无调三音子的模型参数进行初始化,并用最大后验概率准则训练模型。汉语大词汇量连续语音识别实验表明,该方法可以提高训练语料中稀疏三音子聚类前的模型精度,从而提高系统的识别性能。

In order to estimate toned triphone＇ s model parameters accurately before state clustering and increase recognition rate, this paper used the corresponding toneless triphone model parameters for the initialization of toned triphones, and applied the maximum a posteriori criterion for model estimation. In the experiment of mandarin large vocabulary continuous speech recognition, this method can improve the accuracy of sparse triphone＇ s model in the training corpus before clustering and a- chieve certain increase of recognition rate.