Research On Voiceprint Recognition Based On Deep Learning

With the continuous development and application of statistical machine learning and deep learning technology in the field of artificial intelligence,biometrics has gradually become one of the important research objects in the field of artificial intelligence.Voiceprint recognition is a kind of identity verification in biometrics.It refers to a technology that achieves identity recognition by modeling the speaker's acoustic characteristics.In recent years,thanks to the rapid development of statistical machine learning,deep learning,and information technology,voiceprint recognition has made great breakthroughs in basic theoretical research and algorithm models.The main research objects in voiceprint recognition can be divided into acoustic feature extraction and acoustic feature modeling.This paper introduces deep learning methods to voiceprint recognition.It is hoped that deep identity information extraction will be performed on the acoustic features with the help of the nonlinear structure of deep learning.This improves the representation of acoustic features in the model,and provides a better data foundation for closed-collection and open-collection voiceprint recognition.The main work results and innovations are as follows:1.The deep learning model uses a deep belief network that is a stack of three-layer restricted Boltzmann machines,uses Mel cepstrum coefficients as low-dimensional acoustic features,and maps low-dimensional acoustic features into high-dimensional acoustic spaces through the deep belief network.And use this network as a "feature extractor" for deep acoustics.Specifically,based on a 24-dimensional MFCC,a 256-dimensional deep acoustic feature is obtained through the deep confidence network.From the visual and comparative analysis of deep acoustic features and Mel cepstrum coefficients,we can see that high-dimensional acoustic features are more distinguishable than low-dimensional features.2.Relevant researches are carried out using " Deep Belief Network + Support Vector Machine" as a closed-set voiceprint recognition model.Compared with traditional voiceprint features and recognition models,the speaker recognition model designed in this paper has improved recognition accuracy and recognition time efficiency.At the same time,the generalization capabilities of small sample voiceprint recognition and deep acoustic "feature extractor" are studied.When training a deep belief network,each speaker selects small sample data composed of short-term speech signals as the training set of the network.The trained network extracts high-dimensional acoustic features from the speech signals of untrained speakers.The experimental results show that the small sample voiceprint recognition model under this framework has high recognition accuracy,and the deep belief network has good generalization ability.3.In open set voiceprint recognition,an adaptive threshold algorithm based on signal similarity distribution and cumulative probability is proposed.The advantage of this algorithm is that the signal similarity distribution of each speaker is considered.The model first establishes a Gaussian mixture model for the deep acoustic characteristics of each speaker,and uses the result calculated by the Gaussian mixture model as the signal similarity value.Find the approximate distribution of signal similarity values in two subsets,and give a probability to allow false rejection and false reception.Finally,this probability value is used as the cumulative probability in the two distributions,and the distribution point corresponding to the cumulative probability in the two distributions is taken as the candidate value of the threshold.The speaker can completely refuse to know.