矢量泰勒级数特征补偿的说话人识别

Feature compensation algorithm based on vector Taylor series for speaker recognition

将矢量泰勒级数(Vector Taylor Series,VTS)特征补偿算法应用于说话人识别,给出了卷积噪声方差的近似闭式解,构建了联合快速估计卷积噪声和加性噪声均值和方差的框架。该算法可在无需失配环境先验信息的前提下,直接从失配语音中估计出卷积噪声和加性噪声的均值和方差,实现对环境失配的补偿。实验结果表明,在信道变化较大的无线信道下,卷积噪声方差的补偿最高可降低误识率3.24%.提升了系统的识别性能。在存在加性噪声的无线信道下,与基于线性失真模型的特征映射算法和倒谱均值减算法相比,本文算法可分别最大降低49.65%和68.06%的误识率,适合于信道变化较大的失配环境补偿。

A feature compensation algorithm based on vector Taylor series （VTS） is applied to the speaker recognition system. The approximate closed-form solution of the channel variance is derived, and a joint rapid estimation framework for convolution noise and additive noise is proposed. The mean and variance of convolution noise and additive noise can be estimated from the mismatch speech to compensate the environment mismatch without any other prior information about the mismatch environment. The experimental results show that the compensation of the channel variance max- imally reduces the error rate by 3.24%, which improve the system performance in the wireless environment with large channel variations. Compared with feature mapping algorithm and cepstrum mean subtraction algorithm based on the linear distortion model, the proposed algorithm decreases the system error rate by 49.65% and 68.06% at maximum in the wireless environment with additional noise, respectively, especially for the mismatch environment of large channel variations.