Research And Improment Of Speech Recoginition In Noisy Environment

With the development of science and technology and the proliferation of computers, higherrequirement has been put forward about computer exchanges, which promote the development ofspeech recognition technology, making it an important research direction in the field of speechprocessing. A lot of progress has been made in speech recognition so far from research.Inlaboratory with non-noise environment, the recognition rate of the speech recognition system hasreached a very high level, but the recognition performance in noisy environments will besubstantially reduced. The main reason is that noise interference makes training template andrecognition template does not match, resulting in the decline of the identification rate. Therefore,noise becomes the biggest obstacle of a wide range of practical in voice recognitiontechnology,and becomes particularly important in the study of a speech recognition system in thenoise environment.The technology under Gaussian white noise is studied in this paper,include the basicprinciple and every component. There are three main anti-noise technology: speech enhancement,anti-noisy speech feature extraction, anti-noisy speech model parameter adjustment. With theresearch and development of anti-noise technology, how to combine the above technologytogether makes speech recognition system achieve high performance in different noiseenvironments becomes an important direction.Good ability is owned by wavelet transform both in the time and frequency domain,which isa good method of signal analysis, are widely used in many fields such as signal de-noising,and areintroduced into this paper.The theory and characteristics of wavelet transform has beenanalysis.The wavelet threshold de-noised method is paid emphasis on this paper.Speech enhancement,anti-noisy speech feature extraction is paid emphasis on in this paper.Anti-noise method about the combination of the above two methods is given. The basic principleis belove:Firstly, voice enhancement section is increased to the front of the speech recognitionsystem, a new threshold method is given for a problem of traditional threshold,and the denoisingvalidity has been verified by Matlab simulation. Secondly, in the feature extraction stage of thispaper, a new anti-noise feature parameter based on wavelet transform-TDWTMFCC, combinedwith DWTMFCC parameters is used to feature extraction. The two anti-noise improvementmethods (speech enhancement method and characteristic parameter extraction method) is combined and used in a non-specific person, small vocabulary speech recognition system, theeffectiveness of which is verified by Matlab simulation and comparison.