Speech Keyword Retrieval Based On Deep Neural Network

Deep neural networks have developed rapidly in recent years.Deep neural networks are quite mature in English speech recognition and keyword retrieval.This article first uses the DNN-HMM(Deep Neural Network Hidden Markov Model)model to model the Chinese and Uyghur acoustic units respectively,and then LSTM(Long Short Term Memory)is used on the original basis,and GRU(Gated Recurrent Units)network replaces DNN(Deep Neural Network)network for improvement and training.Finally,a grid fusion is proposed.The method performs keyword retrieval of Chinese and Uyghur speech respectively to improve the performance of the keyword detection system,and uses kaldi and pytorch-kaldi to conduct experiments;the experimental results show that the grid generated by LSTM and GRU decoding is a fusion of Chinese-Uighur speech Keyword detection system performance,compared with GMM-HMM,DNN-HMM,LSTM-HMM,GRU-HMM acoustic model retrieval performance,the performance of Chinese-Uyghur voice keyword retrieval is the best,with accuracy rates of 95.25% and 92.67% respectively,recall the rates reached95.12% and 92.33% respectively.This article first introduces the principle and framework of continuous speech recognition in keyword retrieval.Next is the feature extraction process of the speech signal,the Mel frequency cepstrum coefficient and the filter bank feature,and then the cepstrum mean variance is normalized.Then comprehensively use the acoustic model,pronunciation dictionary,language model,modeling technology,WFST to generate the decoding network Lattice.Finally,the research is to capture keywords from the grid Lattice generated by continuous speech recognition,and use the evaluation index of the voice keyword detection system to measure the performance of the continuous voice keyword retrieval system.