Research Of Speaker Recognition In Low-SNR Environment

The voice information of different people contain different voice features.Through speaker recognition technology extracting the differences of feature,the identities of different speakers can be verified.However,due to the impact of different noise environments in practical applications,the system performance is significantly reduced compared to the pure environment.This kind of difference seriously hinders the application and development of speaker recognition technology.Therefore,improving the speaker recognition performance in noisy environments has gradually become a new research direction.In the low signal-to-noise ratio environment,the robustness of the existing speaker feature parameters is interfered by noise.The performance of system is degraded and can`t satisfied the recognition requirements.At the same time,the I-vector speaker recognition model is affected by channel factors.Channel factors make the model poorly distinguish between speakers and low recognition accuracy in noisy environments.This is difficult to satisfied large-scale multi-scene applications.This thesis studies the problems of speaker recognition in noisy environments.Aiming at the degradation of speaker system recognition performance in low-SNR environments,this thesis respectively study the methods of feature parameter extraction and channel compensation.The main work of the thesis is as follows:(1)Use preprocessing,feature extraction and other steps to obtain the I-vector speaker recognition model.This model is used as the basis in this thesis.The baseline feature extraction methods and improvement methods are based on this model.(2)Research for the feature extraction method in the low-SNR environment,use the more robust CFCC for feature parameter extraction,and improve the CFCC to obtain a new feature extraction method.Use an improved Wiener filter as the front-end processing method.Compare with traditional feature extraction methods such as MFCC,the new feature extraction method is more robust and improves the accuracy of system recognition in noisy environments.(3)Research for the feature space channel compensation technology,use feature bending technology to process the front-end voice signal,and use the improve LDA method with WCCN to perform channel compensation.Compared with LDA+WCCN and the baseline,under the interference of external environmental noise,the improved method improves the discrimination between the speakers and the recognition accuracy rate is improved to a certain extent.