The Research Of Speech Recognition Based On Deep Learning In Controller System

In recent years,the popularity of intelligent terminals has promoted the development of intelligent home industry.At the same time,along with the maturity of artificial neural network theory,the depth learning based on artificial neural network has provided a new theoretical reference for the research and application of speech recognition.In this condition,the use of speech recognition technology in the intelligent home system to make it more intelligent become possible.The traditional speech recognition technology mainly adopts the method of template matching.This method can obtain certain effect in the identification of isolated words,but it is difficult to obtain better recognition result in the recognition of large vocabulary continuous speech.Because the neural network can simulate the activity principle of human neurons,and has strong learning,association and reasoning ability,the neural network based on depth learning has become the main direction of speech recognition research.This paper introduces the basic principles of speech recognition and depth learning in detail,and then explains how depth learning can be applied to speech recognition to improve the performance of recognition system.1.Improved noise reduction stack encoder based on deep encoderThe traditional encoder uses a three-layer network structure which includes the input layer,the hidden layer,the output layer,and the hidden layer is used as a feature output.On the basis of the theory of the depth network model,the encoder is improved to increase the number of hidden layers to 5,we name it SAE.At the same time to make it work better in the real environment cited application,we add noise to the first hidden layer in the SAE,so that SDAE can enhance the system's robustness.And then we builds the noise reduction stack model on the MATLAB platform and adopts the layer-by-layer greedy algorithm to train the noise reduction stacker,at last adopts the HTK speech open platform to extract the speech feature of the noise reduction stack encoder model compared with the characteristics of traditional phoneme and MFCC speech,the experimental results show that the extracted eigenvalues of the noise reduction stack can improve the recognition rate of the system and improve the performance of the system.2.Speech recognizer based on Deep learning.Firstly,we analyze the command type characteristics of the control system in the design of the control system,then propose the ATO syntax model based on action,target,operation mode.Second,we study and design each module of the speech recognition system.In the designer of acoustic primitives,the system is designed and modeled according to the vowel model of the vowel with the Chinese pronunciation.In the selection of the acoustic model,we compare the DNN-HMM model with the GMM-HMM model,and find that for the recognition of continuous speech,the data volume of voice is large,and DNN-HMM model has better fitting ability for complex data,at the same time,The DNN-HMM is depth network,so we choose DNN-HMM as the acoustic model,Finally,we take the three-layer structure model of Token Passing Model as the decoder,At last,the speech recognizer based on the HTK platform is established through data preparation and DNN-HMM acoustic model training.Finally,the performance of the recognizer is evaluated and compared to the traditional GMM-HMM model.In the recognition rate,the system can reach 66%of recognition rate,which proves the feasibility of the system.3.Intelligent home voice control systemOn the basis of speech recognition system,we combine the intelligent home control system and speech recognition system to design the voice control system.Then we introduce the voice control system construction and process.At last we test the system performance and prove that the voice control system has feasibility.