【摘要】 为了降低语音识别系统中噪声的影响,提出一种利用隐空间投影算法的模型自适应方法。该方法利用状态间的相关性提取出反映码本和待识别语音共同特性的基矢量。由于语音与噪声是相互独立的,因此,当语音识别系统中有噪声存在时,认为不能用基矢量表示的那部分余量就是噪声。与本征音方法相比,该方法可以有效地降低噪声对语音识别系统的影响。该方法在提取基矢量时利用了自适应数据,并且节省了存储空间。实验结果表明:该方法在噪声环境下相对于最大似然线性回归自适应方法有4~9百分点的提高,相对于最大后验概率和本征音方法有更大的提高。更多还原

【Abstract】 To reduce the effect of noise on speech recognition systems,this paper proposed a new model adaptation method using the projection to latent structure(PLS) algorithm.The algorithm uses correlation information between different states to extract a group of basis vectors reflecting the codebook and adaptation data characteristics.Since speech and noise are statistically independent,the noise can be subtracted from the speech.The algorithm use adaptation data when extracting basis vectors which differs from the eigenvoice method.Furthermore,it needs very little storage space.Experimental results show that the method is superior to maximum likelihood linear regression by 4%—9% and is better than maximum a posterior and the eigenvoice method.更多还原