Research On Speech Recognition Method Based On Deep Learning

With the increasing depth of artificial intelligence technology in people's lives,speech recognition as one of the indispensable technical means has become the focus of attention.The Gaussian Mixed Model-Hidden Markov(GMM-HMM)acoustic model has dominated the speech recognition technology for decades.However,with the voice data becoming more and more large and complex,the shortcomings of the traditional model gradually revealed.As a deep network model,deep Neural Network(DNN)can better deal with the modeling of complex data.The paper details the basic theory of deep learning and experimental principle of speech recognition technology,effectively apply deep learning theory to speech recognition.The main work are follows.1.Start with the basic principle of depth automatic encoder,and study from the aspects of acoustic feature preprocessing,parameter adjustment and system optimization.Extracting stronger robust speech features from original MFCC features using deep auto-encoders,and experimental research on TIMIT voice database through HTK voice recognition framework.2.By analyzing the similarities and differences between deep neural network and Gaussian mixture model in structure and training methods,then clarifies the feasibility of DNN-HMM model which is used to give a more accurate description of output probability.Finally,an acoustic model modeling experiment of GMM-HMM model and DNN-HMM model was performed on the Kaldi speech recognition platform.3.A DNN-HMM acoustic model based on Mel-Frequency Cepstral Coefficients(MFCC)feature and Filter-bank(Fbank)features is established,and experimentally analyzed and compared the effects of the two characteristics on the experimental results.4.Proposed to use Fbank features instead of MFCC features,compare the effect of Fbank feature of different filter banks on recognition rate.Through the study of this paper,we can see that the study of speech recognition based on deep learning theory not only has theoretical significance,but also has application value at the application level.