Small Vocabulary Chinese Isolated Word Speech Recognition Theory And Technology Research

Speech recognition is a hotspot in intelligent human-machine interface. Thedissertation performs systemic study on the theories and technologies of smallvocabulary mandarin isolated word recognition, including five sections ofspeech corpus building, speech de-noising, end-pointing detection, recognitionmethods, and system implementations, with some effective improvements. Thecollection task of speech samples has been organized and completed, meanwhilea standard mini-type speech corpus was established, and the corpus settles thebase for algorithm tests. The theory and experiments show that waveletde-noising method based on soft thresholding operations in different scales issuperior than other existing methods. According to the comparative study ofend-pointing detection methods, some conclusions are obtained, that is, in thecase of high SNR, the energy and zero-crossing ratio method with lowcomputations is a better choice, while in the case of low SNR, theautocorrelation, linear prediction error or cepstrum distance method with highrobustness should be given priority. By constructing distortion measure tableand Hash search function, the traditional dynamic time warping algorithm isimproved based on search table and vector quantization, largely reducingcomplexity in the precondition of less degradation of the recognition ratio.Based on mandarin consonant-vowel structure, the traditional vectorquantization has been translated into segmental vector quantization, improvingquantization precision and the ability for describing feature vector space.Quantity analysis proves that no complexity is increased, and it is an idealmethod for isolated word recognition. The proposed method of segmental quasihidden markov models has beautifully solved the problems of data sparsenessand underflow in training process, and it can be seen as a facility form ofconventional hidden markov models in the case of few samples. At last, asoftware system based on MATLAB and a hardware system based on RSC-300chip have been separately developed. Now the hardware system can wellrecognize 50 addresses of hunan province, and it is in process of argumentationand testing as a part of the pre-study program named "Railway motorcycleinformation control system".