Research On The Speech Recognition System In Noise

With the development of the science and the computer's popularization, people has more demands on the means of communication with the computer, which hastens ? the development of speech recognition and make it be a very important research direction in the field of speech signal processing. At present, the correct recognition radio of speech recognition system is very high in the lab circumstance, and there are some products on speech recognition. But on the effect of the noise in surroundings, the correct recognition radio is much lower than in the lab. So the noise is the biggest obstacle in practical application extensively. It also becomes more import to research speech recognition in noisy environment.There are there kind of de-noising processing technology, including speech enhancement, speech parameter compensation and speech model compensation, introduced in the thesis, and a speech recognition system for isolated words is carried out in these technology. The thesis is mostly composed of four parts. There categories of anti-noise processing technology are introduced in foregoing three parts. In the last part, there are the speech recognition system and the emulational experiment.In the first part, after researching several methods of speech enhancement, the basic theories of wavelet and wavelet de-noising based on the technique of threshold, it is put forward that speech signal is enhanced by wavelet de-noising based on the technique of multi-level threshold firstly, then detecting endpoint and picking-up parameters.In the second part, on the basis of anatomiszing the negative effect of speech enhancement and CNM, the second time de-noising is brought forward, which is wiping off the noise in the speech after speech enhancement in order to eliminate the influence of noise like as convolution noise. Namely, the parameters are standardized.In the third part, after studying HMM decomposing and PMC, the idea that the signal is processed by PMC to wipe off noise the third time is put forward in the thesis. But in order to avoid the correct recognition radio of the system being fall when the signal have nearly reached to purity speech after twice de-noising, the signal must be disposed in SMN before in PMC .In the last part, the emulational experiments are done in Matlab. It proves that the performance of the system that unites three kind of de-noising methods is better, through comparing the correct recognition radio in the method put forward in this thesis with it in other popular methods.