Research On Convolutional Neural Network Based On Variable-Length Speech Data Voiceprint Recognition Technology

Voiceprint recognition has been widely used in biometrics because of its simple and popular speech acquisition equipment,non-contact recognition method and effective recognition efficiency.For example,voiceprint recognition technology in the field of forensic science is used to provide authentication services and increase the basis of trials.The application of voiceprint recognition technology in criminal investigation can reduce the scope of investigation and even directly determine the identity of criminals.The application of voiceprint recognition technology in military field can quickly and effectively determine the target position.Therefore,the study of voiceprint recognition technology has important practical application requirements.Most of the voices in real life are variable-length voices or continuous voices.The feature parameter dimension of speech is often closely related to the duration of speech after the feature extraction of Mel Frequency Cepstrum Coefficient.Since the input of the convolutional neural network needs to match the fixed length data,voiceprint recognition using convolutional neural network for variable length speech will have the problem of mismatch between the feature parameters and the network input.At the same time,diversity of speech acquisition methods in modern society.Voice Transmission Channel also from a single use of radio transmission into a variety of channels,such as the use of millimeter wave,light wave,terahertz communications,etc.When the transmission channel is changed,the model of voiceprint recognition should consider how to remove the influence of the channel.In particular,the voice data transmitted in the near-earth wireless optical communication channel is different from the noise signal transmitted in other channels due to the particularity of the channel,so it is necessary to improve the pre-processing method of voiceprint recognition.The paper focuses on the research of how to improve the recognition rate of voiceprint recognition in the variable length speech signal transmitted through the near-earth wireless optical communication channel.In this paper,a feature clustering based convolutional neural network voiceprint recognition scheme is designed,in which convolutional neural network is used as the core.which effectively improves the recognition rate of voiceprint recognition in wireless optical communication.At the same time,this paper improves the preprocessing method of noisy speech,whichimproves the robustness of voiceprint recognition in speech transmitted through near-earth Optical Communication Channel.This paper introduces an effective clustering method by comparing different application conditions and different recognition effects of various clustering methods.This clustering method solves the problem of the mismatch between the feature parameters and the network input,and realizes the convolutional neural network recognition in variable length speech.The paper use acoustic-phoneme continuous speech database,self-built noise-free speech database(voice transmitted through near-earth Optical Communication Channel),self-built noise-free speech database as speech sample to build the model.Three different recognition schemes,Gauss general background model,identity vector model and convolution neural network model,are tested respectively.The experimental results show that the performance of convolution neural network scheme based on feature clustering is better than Gauss general background model scheme and identity vector model scheme in self built noisy speech database.