Optimization Of Speaker Recognition For Noisy Speech

Voiceprint recognition,also called speaker recognition,as one of artificial intelligence technologies,is being used in people's lives,and brings convenience to human beings.However,the performance of speaker recognition can be interfered by noise.Therefore,in practical application,its performance cannot be as good as the recognition accuracyofpure speech.According to previous studies,noise hasdifferent effects on different recognition models.Noise can increase the equal error rate of speaker recognition to 1.2?4 times,so it is one of the key factors affecting the recognition accuracy.Therefore,it is of great significance to solve the problem of interference caused by noise.At present,there are three kinds of research methodsto improve the performanceof speaker recognition under the condition of noise.The first one is reducing noise in speech.The second one is to study anti-noise features.The third one is data enhancement and to optimize the model structure of speaker recognition.In this paper,we will study the first and the third methods to reduce noise and optimize anappropriate model.In terms of reducing noise,this paper extracts the basic structure of the PIT system for separating multi-speaker speech,that is,processing noisy speech to extract pure speech.This paper tries three methods for data enhancement to improve the anti-noise ability of the model.As for the model of speaker recognition,after trying a variety of models,the LSTM model that can take the context information into account is selected as our recognition model,and the attention mechanism is added to the model to improve the robustness of the model.We use the GE2E loss function in our model.Experimental results show that our model has improved the performance of speaker recognition with noisy speech,and the EER has decreased by 60%?70%.