Abstract: Voiceprint recognition determines the identification of the given speaker by voice,using the individual differences of biological characteristics.It has a wide range of use,with the characteristics of non-contact,simple acquisition,feature stability and so on.The existing statistical methods of voiceprint recognition have the limitations of single-source extracted feature and weak generalization ability.In recent years,with the rapid development of artificial intelligence and deep learning,neural networks are emerging in the field of voiceprint recognition.In this paper,a method based on Long Short-Term Memory (Long Short-Term Memory,LSTM) neural network was proposed to realize text-independent voiceprint recognition,using spectrograms to extract voiceprint features as the model input.Spectrograms can represent the frequency and energy information of voice signal in time direction comprehensively,and express more abundant voiceprint features.LSTM neural network is good at capturing temporal features,focusing on the information in time dimension,which is more consistent with the characteristics of voice data compared with other neural network models.The method in this paper combined the long-term learning of LSTM neural network with the sequential feature of voiceprint spectrograms effectively.The experimental results show that 84.31% accuracy is achieved on THCHS-30 voice data set.For three seconds short voice in natural environment,the accuracy of this method is 96.67%,which is better than the existing methods such as Gaussian Mixture Model and Convolutional Neural Network.

Key words: Deep learning, Long Short-Term Memory, Neural network, Spectrogram, Voiceprint recognition