Voiceprint Recognition Algorithm Based On Telephone Channels

Voiceprint Identification is a technology which can determine the identity of the speaker based on the voice. Now in the lab environment we can obtain a good recognition performance, but once in the real environment, it is susceptible to the interference of ambient noise and a variety of channels, and it declines the recognition rate. Based on improving the recognition rate of the telephone voiceprint identification system, the article conducts in-depth discussion and research.Generally the interference of environmental noise and variety of channels influence the recognition performance by three spaces:signal space, feature space and model space. This paper studies the signal space and feature space in order to control the impact of interference, at the same time the article studies Gaussian mixture model-universal background model (GMM-UBM) in-depth to improve the recognition performance. The main work is as follows:1) This article begins with the signal space, the traditional voice de-noising algorithms. And through the experiments we know that when with strong noise they play a minor role, so this paper focuses on the de-noising method based on independent component analysis (ICA), especially notes the fast-ICA base on negative entropy, and proposes relevant improvement, then the experiments show that this de-noising method significantly improves SNR.2) In order to determine the start and the end of the speech effectively, the article studies the traditional dual-threshold endpoint detection algorithm, and proposes an improved dual-threshold algorithm based on the increment of the magnitude and zero crossing rate, then through experiments the article proves that the improved method can reduce the computation time of the feature extraction.3) Various characteristic parameters are studied, including:LPC, LPCC, MFCC and their delta coefficients, and the article studies the recognition performance by the static and dynamic feature combination. And through the study of pitch, we know that the pitch change is very small by channels interference. In the different channels, we can improve the robustness of the system by combining pitch and other characteristic parameters.4) Common recognition models are studied:VQ model, GMM and based on these GMM-UBM is deeply studied, through the experiments we analysis the effects of the number of codebook, the length of training, the number of Gaussian mixture on recognition performance. Finally, the results show that the GMM-UBM can effectively improve the recognition performance.5) On the background of attendant reconstruction project, using the related algorithms, the article designs and achieves a complete set of the voiceprint system based on telephone channel.