Study On Voiceprint Recognition Based On Deep Recurrent Neural Network

Voiceprint Recognition,also known as Speaker Recognition,is one of the most popular biometric-based identification technologies.Voiceprint recognition is a technique for identifying the speaker's identity of the voice to be tested based on the speech parameters that reflect the physiological and behavioral characteristics of the speaker in the speech waveform.Voiceprint recognition can be used as a supplement to multi-factor recognition,and the sound only needs a microphone during the sampling process.The data collection of voiceprint recognition is more convenient,the acquisition equipment is cheap,and in the scene with only sound data,the voiceprint identification is especially important.It is an economic,reliable,convenient and secure way of identification.Voiceprint recognition has begun to be applied to a variety of smart devices for the identification of user identities,such as the use of voiceprints to enable social software login,language assistants in smart audio,voice assistants in smart in-vehicle systems,and so on.Thesis studies and analyzes the development history and research status of voiceprint recognition technology,improves the endpoint detection of existing voiceprint recognition,and proposes a combination of Convolutional Neural Network(CNN)and Deep Circulation Neural Network(DRNN)is a voiceprint recognition scheme called CDRNN.CDRNN combines the advantages of CNN and RNN for mobile terminal voiceprint recognition applications.The main research contents of this paper are as follows:(1)Aiming at the endpoint detection problem of speech signals in noisy environment,an improved algorithm combining multi-window spectrum spectroscopy,energy entropy ratio and double threshold is proposed.The algorithm firstly performs noise-free signals by multi-window spectrum subtraction.The noise reduction process obtains a relatively pure effective speech signal,and then uses the energy entropy ratio algorithm to calculate the speech energy.Finally,the dual threshold method is used for subsequent endpoint detection.(2)Building a CNN network.The advantage of CNN is longer than that of processing images.The speech signal is transformed into a spectrogram for processing,and the personality characteristics of the speech signal are extracted from the spectrogram.The recognition rate of different layer CNN networks is explored.(3)Building a DRNN network,and the output of the CNN network after the spectrum processing is used as the input of the DRNN to complete further time series modeling.And explore the number of nodes per layer and the impact of the number of layers on the recognition rate.(4)Compare and analyze the CDRNN scheme with other commonly used voiceprint recognition schemes.The experimental results show that the CDRNN scheme can obtain better recognition accuracy than other schemes such as GMM-UBM and GMM-DNN.