论文部分内容阅读
本文针对最大似然线性回归算法线性假设的缺点,将多项式回归方法用于模型自适应,构建了基于最大似然多项式回归的非线性模型自适应算法。该算法在对数谱域用多项式回归方法,逼近每个Mel子带上识别环境模型均值与训练环境模型均值之间的非线性关系。多项式系数通过EM算法和最大似然准则从识别环境下的少量自适应数据中估计。实验结果表明,二阶多项式就可以较好地逼近模型均值的非线性环境变换关系。在噪声补偿和说话人自适应实验中,最大似然多项式回归算法的误识率都明显低于最大似然线性回归算法。本文算法较好地克服了线性模型自适应算法线性假设的缺陷,可同时减小噪声,和说话人的改变或其它因素对语音识别系统的影响,尤其适合说话人和噪声的联合自适应。
In this paper, according to the shortcoming of the linear assumption of the maximum likelihood linear regression algorithm, the polynomial regression method is applied to model adaptation, and a nonlinear model adaptive algorithm based on maximum likelihood polynomial regression is constructed. The algorithm uses the polynomial regression method in the log spectrum domain to approximate the nonlinear relationship between the mean of the identified environmental model and the mean of the training environment model in each Mel subband. Polynomial coefficients are estimated from a small amount of adaptive data in the context of identification through EM algorithms and maximum likelihood criteria. Experimental results show that the second-order polynomial can approximate the nonlinear environmental transformation of the mean of the model. In noise compensation and speaker adaptive experiments, the maximum likelihood polynomial regression algorithm has a significantly lower false-positive rate than the maximum likelihood linear regression algorithm. The proposed algorithm overcomes the shortcomings of the linear assumption of linear model adaptive algorithm, and can reduce the influence of noise, speaker change or other factors on the speech recognition system at the same time, especially for joint adaptation of speaker and noise.