EI / SCOPUS / CSCD 收录

中文核心期刊

基于回归分析的语音识别快速自适应算法

Rapid adaptation algorithm based regression analysis for speech recognition

  • 摘要: 从回归分析的角度推导出最大似然线性回归算法的等价算法——最小二乘线性回归算法,以及相应的多元线性回归模型。该模型中回归因子间存在着多重共线性,它导致了算法在自适应数据很少时失效。为减轻多重共线性的影响,提出改进算法:伪自适应数据算法。实验表明,当仅有1s~3s自适应数据时,新算法使得系统误识率相对下降2%~6%,随着自适应数据增多,其性能与最大似然线性回归(或最小二乘线性同归)算法趋于一致。

     

    Abstract: Least Square Linear Regression (LSLR) algorithm was given by regression analysis method. LSLR was the equivalent algorithm of traditional Maximum Likelihood Linear Regression (MLLR). The corresponding multilinear regression model was found. The multicollinearity of LSLR caused the performance degradation when adaptation data was limited. Pseudo Adaptation Data (PAD) method was proposed for decreasing the degree of multicollinearity. Experimental results showed that PAD outperforms MLLR when adaptation data was sparse and converges to the MLLR/LSLR performance when more adaptation data was available.

     

/

返回文章
返回