Research On Text-Independent Speaker Recognition

Speaker recognition refers to recognizing persons from their voice, it owns the natural advantage compared to other authentication technologies. The acquisition process of the speaker's utterance is simple, speaker recognition technology has high user acceptance. It has been widely applied to military field, public security and judicial department as well as internet security.The performance of the speaker recognition system is vulnerable to background environment, the health and mood of the speaker and the channel mismatch problems, therefore, the key to improve the system performance is to enhance the ability of feature representation. In the view of feature selection and feature extraction, the common vector approach and deep learning were introduced to speaker recognition as the way of feature extraction in this paper. Two modified methods were also studied in this paper. In the view of feature compensation, the normalization and compensation technologies applied to GMM-SVM and Total Variability Factor analysis were studied in this paper. The introduced and modified methods can improve the capacity of different systems with the result of comparative experiments.(1) CVA extracts more effective acoustic feature by transforming the space of original spectrum feature to the indifference subspace. The research of speaker verification studied on three main models including GMM-UBM, GMM-SVM and Total Variability Factor analysis that it was used to demonstrate the effectiveness of CVA introduced to speaker recognition by this paper.(2) Deep learning extracts deep feature layer by layer which can characterize the identity information of the speaker. The research of speaker verification studied on two models including GMM-UBM and Total Variability Factor analysis that it was used to demonstrate the effectiveness of deep feature introduced to speaker recognition by this paper.The mainly research jobs are as follows:(1) Study on the acoustic feature analysis and extraction.(2) Study on speaker recognition based on common vector approach. CVA is proven to be effective on speaker verification based on three models by theoretical analysis and comparative experiments. Two modified methods and compensation technologies are also proven to be effective by the experiments.(3) Study on speaker recognition based on deep learning. The result of the experiments showed that deep feature learning from deep RBM structure applied to speaker recognition can indeed improve the performance of the systems.