论文部分内容阅读
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。
In order to reduce the influence of noise in speech recognition system, a model adaptive method using hidden space projection algorithm is proposed. The method uses the correlation between states to extract the base vectors that reflect the common characteristics of the codebook and the speech to be recognized. Since speech and noise are independent of each other, when there is noise in the speech recognition system, the part of the residual that can not be represented by the basis vector is noise. Compared with the eigen-tone method, this method can effectively reduce the impact of noise on the speech recognition system. This method uses adaptive data when extracting base vectors and saves storage space. The experimental results show that the proposed method has a 4 ~ 9% improvement over the maximum likelihood linear regression adaptive method under noisy environment, which is greatly improved compared with the maximum a posteriori probability and eigen-tone method.