Research And Implementation Of Speaker Recognition Method Based On Short Utterance

As a part of speech signal processing,speaker recognition has broad application prospects and is one of the key technologies in the field of human-computer interaction.However,When applied in practical applications,the speaker's utterance extracted is so short that the recognition performance is unstable.Addressed on the improvement of human-computer friendliness interaction,this paper has done some research on short utterance recognition system,which has theoretical significance and practical application value.First of all,this paper has designed the overall structure of the speaker recognition system,and analyze the processing process of the system.Feature extraction and acoustic model are taken as the research focus.Secondly,under the short speech conditions,the biased distribution of speech makes the extracted I-Vector unreliable,so this paper proposes an I-Vector compensation method based on generative adversarial networks(GAN).This method trains the generator network to generate a compensated I-Vector from the phrase sound I-Vector,and at the same time trains the discriminator network to determine whether the input I-Vector is generated by the generator compensation or from the long speech I-Vector.subsequently,the I-Vectors extracted from the phrase sounds are compensated as much as possible to their corresponding long-voice I-Vectors by completing the confrontation training between the generator network and the discriminator network.The compensated I-Vector can recover the lost information and provide better voice feature information for subsequent speaker identificationThen,considering the possible shortage of short speech training samples,this article focuses on the acoustic model and deep learning model,and proposes a new dual-discrimination method based on I-Vector and deep neural network(DNN)hybrid model.This method uses Probabilistic Linear Discriminant Analysis(PLDA)to reduce channel interference in speech and improve the robustness of the speaker recognition model based on I-Vector under phrase sounds.At the same time,DNN is used to train speech features to extract deep features in phrase sounds,so as to effectively train and recognize a small amount of speech.Finally,a dual discriminant mechanism based on the hybrid model can effectively overcome the shortcomings of fewer training samplesand susceptibility to interference,thereby improving the recognition performance of the phrase sound speaker recognition system.Finally,this paper completes the integration of the speaker recognition system on the service robot.The experimental results show that the recognition performance of the speaker recognition system combined with the improved method of this article is improved compared to the original speaker recognition system under the phrase sound environment of different lengths.It shows that the speaker recognition system has certain application value in the actual application environment.