A robust speaker recognition approach based on model compensation

Abstract

Model compensation is an important means to improve the robustness of speaker recognition in noise environment. The robust speaker recognition approach based on model compensation is proposed in this paper. The proposed method combines data reliability estimation and feature components effectiveness estimation, so the errors of the second kind due to the estimation error are reduced greatly. The proposed method estimates the reliability of data in time-frequency domain based on the fuzzy reasoning, and estimates the effectiveness of feature components in the current environment. Then according to the result of estimation, the model compensation in the noise environment is determined based on fuzzy reasoning to improve the robustness of speaker recognition. The experiments compared the proposed method to the missing data and feature selection combined method. The results showed that the proposed method is more effective in different noise environment.