The Research Of Speaker Recognition Under Noisy Environment

As an important technology for speech signal processing, speaker recognition is widely researched at present. Speaker recognition identify the speaker by the specific information related to the speaker in speech signal. In this paper, speaker recognition is researched from different viewpoints, a speaker recognition system is implemented which include speech enhancement, feature extraction and model selection. Work achieved in this paper are as follows:(1) Analyzes the reason for poor real-time of STSA-MMSE algorithm. Parameter (ω|-) is introduced to improves the real-time of STSA-MMSE, human auditory masking properties is then combined with parameter (ω|-) to reduce the voice distortion and improve the signal-to-noise ratio (SNR). MATLAB simulation shows that the algorithm improved is better than the original one both in real-time and voice distortion reducing.(2)Analyze features of different bands for Mel Frequency Cepstral Coefficient(MFCC). As the stability of high-level parameters to noise, add feature parameters for sinusoidal window to improve the robustness of feature parameters. It is proved by experiment that the robustness of the new feature parameters is better than of the former.(3) Application of Vector Quantization (VQ) theory in speaker recognition is researched. As different Parameters contribute differently to recognition rate, different weights are added to each parameter to improve the recognition rate. Experimental results show that the vector quantization algorithm with weight proposed in this paper performs better that the original algorithm in noisy environments.