Research On Algorithm Of Speech Enhancement Based On Sub-band Decomposition Fraction Fourier Transform

The voice signal plays a very important role in all aspects of daily life and works. It often being interfered by varieties of noises, which affect the information receive badly, even cause the voice processing systems failure. Therefore, the noisy-voice signal must be denoised, suppressing noise and improving the communication quality. In this paper, mainly research on voice enhancement algorithm which based on sub-band decomposition fractional Fourier transform. In the system of sub-band decomposition, firstly using Fractional Spectral Subtraction to estimate voice model parameter, then using the Kalman filter further to eliminate the musical noise. Fractional domain voice enhancement algorithm mainly contains Spectral Subtraction model improving, the best transformation order choosing etc.Firstly, introducing the characteristics of voice and noise, then expounding the voice enhancement basic algorithm respectively and briefly, there are the voice-production-based model algorithm, non-parametric algorithm, the statistical-based model algorithm and other method.Secondly, researching on the basic definition of Fractional Fourier Transform, and briefly introducing the basic characteristics of Fractional Fourier Transform and four kinds of discrete Fractional Fourier transform algorithms. On this background, researching on the Fractional domain optimal filter operator principle and choosing the best transform order times. Comparing with the improvement of traditional spectral subtraction model, should choose the best transform degree of fractional spectral subtraction at first, at the same time using the best transform degree of fractional domain to process the noisy-voice signal, finally inserting phase coefficient, and separating the voice and noise optimally. However, given the noise estimation experiential mode assumptions, although increasing the detection efficiency on the basis of the self-correlation, but it is unstable. The estimated value of power spectrum will be lost some original voice information as well as producing musical noise, impacting on audition effect.Kalman filtering algorithm using the parameters method which depends on the voice production model, suppressed musical noise well, but the algorithm complexity is higher by using the iteration,.This paper using technique of sub-band decomposition, decomposed the noisy voice signal into M sub-bands, determine the best transform degree as premise, using the fractional spectral subtraction to improve the power spectrum estimation of the voice parameters model before the Kalman filter, extracted amplitude characteristics of noisy-voice signal which in this transformation degree as the input of sub-band decomposition Kalman filter, resulting that on the basis of fractional spectral subtraction to achieve suppression of musical noise. Finally, on the basis of the theoretical analysis, using segSNR and PESQ voice evaluation measure, focus on different SNR additive Gaussian white noise, pink noise, factory noise environment, to the method of this paper simulate and analysis, and comparing the time-domain waveform similarity between clean voice signal and the enhanced voice signal. Results of analysis show that using sub-band decomposition reduces the kalman filter model degree, on the basis of Fractional spectrum reduction can estimate kalman filtering voice parameter model better, and it can further suppress generation of musical noise, has better enhance effect.