针对实际环境中语音信号的时频分量普遍存在部分缺失或严重失真的问题,在已知语音先验知识的条件下,提出了一种利用可靠时频分量对缺失数据进行补偿的方法。利用贝叶斯准则,将最优补偿转化为求解后验概率最大化的问题,并利用缺失数据自身的能量信息,给出了一种局部最优补偿的方法。实验表明,该方法在各种噪声、信噪比环境下,综合性能优于传统的鲁棒语音识别技术;采用缺失信息对补偿进行限定,在低信噪比下鲁棒性能有了明显的提高。
Data missing is a natural occurrence in the real environment. According to the prior speech distribution, a missing data imputation method is proposed using the reliable data compo-nent. By the Bayesian rule, solving optimal imputation comes down to finding the value which maxi-mizes the posterior probability, and a suboptimal data imputation is proposed according to energy of the missing data. The results of the experiment show that the proposed method outperforms the state-of-the-art robust speech in different SNR environments; especially when the SNR is low, the energy bounded imputation exhibits an obvious improvement in robustness.