An Environment Learning Mechanism for Robust Speaker Recognition
Jing Zhang, Yibiao Yu · 2019
When application environment is inconsistent with training, the performance of speaker recognition system will drop significantly. Moreover, the real application environment is not able to be predicted in training stage, and it varies all time. In this paper, an environment self-learning method for robust speaker recognition is proposed. An improved Vector Taylor Series (VTS) is used to characterize the statistical distribution relationship between environment model and pure speaker model, and applied to the feature domain and model domain to compensate additive noise. When the environment changes, the prior environment noise data between speech intervals is collected and be used to update the Gaussian Mixture Model (GMM) of environment for compensating mismatches, so then to flexibly make the pure speaker model to fit the current application environment. The speaker identification experiment results show the proposed method improves the system performance significantly under different kinds of noise at low SNR.