Feature compensation algorithm based on vector Taylor series for speaker recognition

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 maximally 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.