| With the development of computer technology and pattern recognition, speaker recognition technology has achieved excellent results. Nowadays, acquiring speech from the Internet has become a mainstream method. And most of speech from Internet is in the form of speech coding. However, there are still a lot of shortcomings for the performance of speaker recognition system using compressed speech.Experiments show that the performance of traditional MFCC based speaker recognition system has a great degradation in the case of speech codec mismatch, which affects the generality of speaker recognition system.This paper is mainly focused on the problem of speech coding mismatch in speaker recognition system. I proposed a new I-vector based speaker recognition method by using the PNCC feature as a main system and the residual feature excitation from bit-stream, the modified SCF feature as the complementary systems for G.723.1 codec speech.Compared with the performance of traditional speaker recognition system,the new system improved the performance and robustness of the proposed system.Towards the speaker recognition technology for compressed speech,this paper mainly finished the following research:Firstly, I finished an I-vector based speaker recognition system. To study the configuration of the speaker recognition system, I evaluated the performance of speaker recognition system under different datasets, and we also perfected the performance of the system.Secondly, towards the G. 723.1 low bit-rate speech coding, I studied the robustness of speech feature under codec mismatch from three aspects:acoustic features, residual feature extraction from bit-stream and SCF features. In this paper, we proposed a modified SCF feature to improve the robustness.Finally, a robust I-vector based speaker system is proposed. I used the PNCC acoustic feature as main system, combining with the residual feature extraction from speech coding bit-stream and the modified SCF feature as complementary systems. Both I-vector fusion method and score fusion method showed the improvement of the system. We found the LR fusion can best result compared with different fusion method of our system. |
