語音識(shí)別綜述數(shù)字語音識(shí)別

1、Speech Recognition 語音識(shí)別By TerrySpeech RecognitionnSpeech recognition is a high technology of processing voice signal into corresponding texts and commands by machine recognition and understanding.nSpeech recognition technology has involved signal processing, pattern recognition, probability theory a

2、nd information theory, vocal mechanism, hearing mechanism and artificial intelligence. nSpeech recognition technology is mainly consist of three module , including feature extraction, pattern matching technology and model training. Speech RecognitionnThe History of Speech Recognition Development 195

3、9Ten phoneme recognition system Audry System,Bell Labs20th,50slate 60s to early 70sLPC,DTWVQ,HMMSphinx System,Carnegie Mellon University,ANN,HMM80s90sIBM，Apple，ATT and NTT A Hot Area in AI,More processing Method,NowadaysSpeech RecognitionnCategory of method:nIsolated word recognition nConnected word

4、 recognitionnContinuous speech RecognitionnSpecific person recognition nNon-specific person recognitionnSmall vocabulary nMedian vocabularynLarge vocabularynInfinite vocabulary Speech RecognitionnMainly Methods:nTemplate Matching DTW(Dynamic Time Warping ) VQ(Vector Quantization)nHMM DHMM(Discrete H

5、idden Markov Model) CHMM(Continuous Hidden Markov Model) SCHMM(Semi-Continuous Hidden Markov Model)nANN(Artificial Neural Net)Speech RecognitionnSignal Pre-processingnFramming-5ms to 50msnEndpoint detection-detect the starting point and terminal pointnSpeech Enhancement-inhibit noise and improve spe

6、ech quality nICA-Independent Component AnalysisSpeech RecognitionnFeature ExtractionnLPC-Linear Prediction coefficientnLPCC-Linear Prediction Cepstrum Coefficient nMFCC-Mel Frequency Cepstrum Coefficient Cepstrum:njnjwenxeX)()(njnjwenxeX)()(njnjwenxeX)()(deeXmcjmjw| )(|ln21)(Speech RecognitionnSpeec

7、h RecognitionnTemplate Matching DTW(Dynamic Time Warping ) VQ(Vector Quantization)nHMM DHMM(Discrete Hidden Markov Model) CHMM(Continuous Hidden Markov Model) SCHMM(Semi-Continuous Hidden Markov Model)nANN(Artificial Neural Net)CRS Introduction Matlab GUICRS IntroductionnPeocedurenPre-ProcessingnFea

8、ture ExtractionnDTW+VQCRS IntroductionnPre-ProcessingnPre-emphasis nWindowing-Non-stationary signal Rectangle Window Hanning Window Haiming Window 1( )1H ZuZ 20.540.46cos()( )10nw nNCRS IntroductionnFeature ExtractionnEndPoint Detection Short-time energy Zero crossing rate(Double Gates) nMFCC-Based

9、on Auditory ModelDFT DFT逆DFT 信號(hào) 頻譜 對(duì)數(shù) 倒譜 CRS IntroductionnTemplate MatchingnTemplate sets selectingSingle Optimal Selection MethodSFS(Sequence Forward Selecting)SBS(Sequence Backward Selecting)GRNN(General Regression Neural Network)n Template subsets (our own)Classifying according to the size of fra

10、me A.30 C.ElseCRS IntroductionnTemplate MatchingnDTW AlgorithmNnnNnnjiCWWnynxdD11),(),(min) 1()(, 2 , 1) 1()(, 2 , 1 , 0)() 1(:)(, 1) 1 (:nwnwnwnwnwnwMNww連續(xù)條件邊界條件CRS IntroductionnTemplate MatchingnDTW Algorithm.)(,),() 1()(,) 1()(, 1),(:)2,(),1,(),(),(min, 1), 1(的約束條件取值滿足就是其中nwmnmngnwnwnwnwmngmnDmnD

11、mngmnDmndmnDNnnNnnjiCWWnynxdD11),(),(minCRS IntroductionnTemplate MatchingnDTW AlgorithmDP(Dynamic Programming)123fori21(1. );1?(1,1):Re;2?(1,2):Re;( , )( , )min( 1,2,3);to n dofor jtom doDd ijDjD ijalMaxDjD ijalMaxD i jd i jD DDendendCRS IntroductionnClassic K-NNnSort the Distance-Sequnce By Small to largenFind the first K distance elementsnThe best match(result) is the number with the largest proportion in the K elementsnOur Own:Weighted K-NNCRS IntroductionnVQnTraining CodeBook),(1)(1TtittixxdTCDiitCx ),(minargijCiyityxtdxij)(min)(iikCDCDCRS IntroductionnRe