| Speaker Recognition Technology has already made a great advance in the past many years' development, but it's still in the early class stage comparing with the sense of hearing ability of mankind and can't satisfy the vast application need. Now many developed countries are researching speaker recognition technology one after another, the 863 plan includes the technology too in our country. In one word, speaker recognition has become one of the most development potential technologies of information realm in the 21st century.Paper's research contents concentrate on the strategy of speaker recognition. By reviewing and analyzing the actuality of speaker recognition technology, as far as it goes, there are two methods to improve speaker recognition technology is thought in this paper. One is to improve the skilled technology to increase the ratio of recognition; the other is to combine the skilled technology with the advanced technology to come into being a new recognition technology. So paper places emphasis on the way of combining ANN with HMM and improving vector quantization of HMM, Then puts forward a recognition way of combining VQ with HMM and uses SOFM to replace the normal LBG of VQ for vector quantization and codebook design, adaptive technology is used too. To test this strategy, paper builds a small-scale system of speaker recognition.By theoretic analysis and experimentation, we known using SOFM to replace LBG may realize general optimum training codebook and controlling the training time, may increase the computation speed by using parallel arithmetic. Comparing with the signal HMM recognition, this strategy may decrease the error of vector quantization and increase recognition precision of HMM. The recognition result is credible for decided by VQ and HMM. For such as speaker's age, the speaker mode's parameter is adjusted by adaptive technology along with the change of speaker's voice.Keywords:SOFM;VQ;HMM;EndPoint-Detect;Pitch;Cepstrum-Coefficient...
|
PDF Full Text Request