Study On Speech Enhancement Algorithms In Speaker Recognition And System Implementation

Speaker recognition, a technology based on biologic features, automatically recognizes the speaker according to the tested speech. It is of great application value and wide application prospect. Speaker recognition system can obtain high recognition rate for clean speech, but the performance will degrade rapidly in noisy environments. Speech enhancement is the chief method for solving problems. If the noisy speech has been processed by speech enhancement before sent to the speaker recognition system, the system could achieve higher recognition rate. Therefore, it is of great importance to further study on this technology.In general, speech enhancement works as the pre-processing or the front-end processing module of the system. Its purpose is to extract the clean speech from noisy speech as far as possible. For the various source and individual characteristics of noise, specific method should be chosen to cope with it. Aimed at the additive noise, the dissertation studied on the speech enhancement algorithms based on short time spectral amplitude, and did some test and improvements as follows:①This dissertation compared the common energy features, and improved the voice activity detection (VAD) algorithms based on short-time energy and short-time zero-crossing rate by introducing the log energy and adjusting the threshold, which will be convenient of VAD in low SNR conditions.②This dissertation did some researches on the speech enhancement technologies which are based on short time spectral amplitude estimation, including spectral subtraction, wiener filtering and minimum mean square error estimation. Fundamentals and implementations were stated as well as some improvements. Advantages and disadvantages of the mentioned methods were obtained by system analysis, so did the generic form of the speech enhancement technologies based on short time spectral amplitude estimation. Besides, the methods above were implemented in computer simulation.③This dissertation improved the technology for noise spectral estimation. Furthermore, a modified Wiener filtering speech enhancement algorithm based on a priori SNR estimation is proposed. Derivation process was stated theoretically. Later, experiments results verified the effectiveness of the proposed algorithm.④After studied the speech enhancement algorithms, this dissertation took the speech enhancement as the front-end processing module of the speaker recognition system and developed the anti-noise speaker verification system.