Voice Frequency And Nonlinear Enhancement Processing Method

It has been an important studying task in our country and overseas that the technologies studying in noisy environment and the speech processing systems developing in real situations, the noise controlling is becoming more important with progress of communications technologies. The quality of speech signal is affected to decline quickly by high noises, so controlling noise, speech distortion and improving speech intelligibility have become central aims. Transmission characteristic, signal feature and the masked hearing feature are thought synthetically in this thesis, and time-frequency filter algorithm, non-linear speech processing enhancement algorithms are proposed. The main contributions are listed as follows:1,Proposed speech fractal enhancement algorithm in 2-dimensional, which aims at correlation and non-stationary of noise. The algorithm synthesizes 2-D Fourier transform and speech processing technology, The algorithm has reduced non-stationary noise and restrained musical noise effectively. The study shows fractal theory is non-linear theory, it can supply the gaps of linear analysis. 2-D adapting filter make original character of speech signal be kept.2,Proposed the concept and algorithm of adaptive network-based Neuro-fuzzy inference system (ANFIS).The algorithm can filter linear or non-linear noise , the filtering result is better than general linear filters. Experiments results show ANFIS method is better than that of neural network and high order neural network, it has high speed, little error and small data, effect is all right.3,Proposed a new controlling noise algorithm : Wavelet filtering and weighted function two-ends detection methods: random noise make signals variation , the measure of variation is measuresd wirh Lipschitz exponent. random Lipschitz exponent is different with effective signal's,and binary wavelet module maximal value also is not same in different scale. So we can obtain the effective signals with the feature. This method can be used in many SNR, also can avoid musical noise. On the other hand, weighted function provide a basis for two-ends detection4,Proposed speech enhancement based morphological Filter. Morphological Filter(MoF) is generally applied in image processing ,the paper uses it in speech processing, impulse noises is reduced by morphological Filter, Harmonics of various degrees is restrained with increscent structure elements . high-level Harmonics and impulse noises can be filtered by LPF and MoF. enhanced speech and clean speech is a perfect match. Speech recognition achieves the highest recognition rate in a large range of SNR. Signal-to-noise ratio tests .mean opinion score tests recognition accuracy comparison show good results than other methods.