Performance evaluation and improvement of speaker recognition over GSM environment

Abstract

Performance evaluation and improvement of speaker recognition over real GSM environment are investigated. A text-independent speaker recognition system based on Gaussian mixture model (GMM) is implemented for performance evaluation. To match the real-world conditions, an NTUT-LAB416 speech corpus is collected over GSM telecommunication network from in-car environment of various driving speeds. An approach employing multistyle training model is proposed to alleviate the adverse effects due to environmental mismatch. Also, a post-processing scheme using auto-regression and moving-average (ARMA) filter is suggested to overcome the varying noise conditions. Experimental results indicate that the proposed approaches can effectively improve the performance of speaker recognition over GSM environment.