Study And Design Of Speech Recognition System Based On Phonetic Element

With the progress of society, people have much stronger willing to pursuitintelligence. Speech recognition is a bridge between the communication of machine andhuman, and obtaines more and more people’s attention. It requires that machines couldunderstand large amount vocabulary and continuous statements, and could resist thecolloquial, noise and other interference factors, so that people can communicate withmachines normally. However, large amount vocabulary and continuous speechrecognition still encounters many difficulties at present; its identification efficiency isnot satisfied. Additionally, in the process of globalization, the speech recognition systemis unable to deal with many languages-mix situations which make speech recognitionmore difficult, such as in the sentence with English and Chinese.Based on the thinking of difficulties of amount of vocabulary and voice continuity,this thesis studies the speech recognition system based on the phonetic sound element,which has further refined recognition unit of the acoustic vowel speech recognitionsystem. Many mature voice recognition technology is used and combined with theproposed phonetic element segmentation method and separation of phoneticidentification and semantic identification to form a new structure of voice recognitionsystem framework, and is studied in the MATLAB. The main content of this thesisincludes three parts below.1) Speech technology, related to the speech recognition system based on phoneticsound element, is discussed in this thesis. Many related technology are introduced indetail, such as voice model principle, Chinese phonetic knowledge, signalpre-processing technology, peech feature extraction, and template matching technique.The LPCC and MFCC parameter are compared, and MFCC is more suitable.2) A speech recognition system based on the phonetic sound element is constructed,and the solutions of each technology used in all parts of the system are introduced indetail, including the double threshold method for speech endpoint, FE algorithm forsegmentation of consonant and vowel, phonetic element segmentation method, MFCCcoefficients as the initial and final characteristic coefficient, DTW algorithm realize thefeature template matching, character building and search technology. Several phoneticelement segmentation methods are analyzed and compared. Two element segmentationmethods, which are based on speech envelope and speech extremum, have good effect. The element segmentation method based on phonetic extremum has the advantages oflow computation amount and high characteristics rate is used as the pitch segmentationtechnology in the speech recognition system based on phonetic element segmentation.The speech recognition system uses different feature template libraries for consonantand vowel; has more detailed segmentation of vowel. Detailed segmentations of vowelare used to extract the speech features, and that makes the feature template lengthshorter. The system obtain the consonant and vowel sequence of letters after speechfeature template matching, which is the pronunciation recognition; followlly, intelligentphonetic methods are used to convert letter sequence to concrete words,which issemantic recognition. Phonetic and semantic recognition separated can reduce thematching difficulty and the amount of search, and is helpful for variety of mixedlanguage recognition; it can use other more mature technology, such as intelligentphonetic system, and makes the speech recognition system more intelligent.3) Multiple audio element segmentation experiments are accomplished inMatlab2007, and the effects of these experiments are compared based on many noisycases. The element segmentation method based on phonetic extremum has bettersegmentation performance. Sound element segmentation is very important for speechrecognition system based on voice element, and it affects the vowel template featureextraction and the final result of speech recognition. The element segmentation accuracyof the phonetic element segmentation method can be as high as90.2%; and it can bebetter when the speech recognition system use more efficient Chinese Pinyin inputmethod which can recorrect the wrong spelling oth letter sequence.This thesis studies the speech recognition system based on tone element and someexperimental analysis about key technologies. Later this thesis proposes a new solutionfor large vovabulory and continuous real time speech recognition and establishing avoice recognition and semantic conversion separation architecture.