Chinese Speech Keyword Detection Technology Research

Deep neural network(DNN)and recurrent neural network(RNN)have been successfully applied to English speech recognition and spoken term detection systems.In this paper,we use the deep neural network-hidden markov model(DNN-HMM)and the recurrent neural network with long-short-term memory(LSTM)to model the initials and finals to improve the performance of existing Mandarin spoken term detection system.In this paper,the framework of continuous speech recognition is introduced,including feature extraction of speech signal,acoustics modeling,pronunciation dictionary and language model,the speech decoding network based on weighted finite state transducers.There are four kinds of aoustic feature are introduced,including perceptual linear prediction coefficient,Mel frequency cepstral coefficient,filter bank characteristic,and pitch.Then,the spoken term detection technology based on continuous speech recognizer is studied,including lattice structer,index building,search method,confidence level and evaluation of keywords spotting system.In this paper,an improved Mandarin keywords spotting system is proposed.The system uses initial and final with low error recognition rate to perform acoustic modeling and retrieval.The key words in the form of Chinese characters are converted into initial and final by look-up table method.The acoustic model based on DNN-HMM and the acoustic model of LSTM-RNN were trained respectively,separately reached the recall rate of 73.32%and 79.84%.The result shows using LSTM-RNN acoustic modeling has imprived the spoken term detection system.A set of experiments based on different acoustic features is set.The results show that pitch can improve the detection performance.Then fusion confidence level are applied in the system.Based on the experimental results,this paper discusses the application range of the two kinds of Mandarin keywords spotting systems.Finally,a new Mandarin spoken term detection system based on system fusion is proposed.