Research On Algorithm Of Speech Enhancement-Based DFRCT And Adaptive Filter Banks

The voice signal plays a very important role in our daily life. It is often difficult to avoid a variety of noise interference, which affect the information receive, in severe cases may even cause the failure of the voice processing system. Therefore, it is necessary to denoising the speech signal with noise, and improves communication quality. This paper mainly study on adaptive speech enhancement algorithm based on discrete fractional cosine transform filter banks, which introduces orthogonal transformation better performance of the discrete fractional cosine transform into the common transform domain adaptive LMS filter algorithm, and divides the input voice signal into sub-band, then processes each sub-band separately, The aim is to reduce the computational complexity of the algorithm. The main work of the paper is as follows:(1) This paper first introduces several common algorithms for speech enhancement and adaptive filtering algorithm development, which lay the theoretical foundation for this paper to study adaptive filtering algorithm, then the voice characteristics and noise characteristics, and the methods of evaluating enhanced voice quality, selects the evaluation criteria for the speech enhancement algorithm.(2) This paper presents filter banks DFRCT-LMS for speech enhancement, it main researches the DFRCT-LMS algorithm, and introduces the definition and basic properties of the discrete fractional cosine transform, and the time domain and transform domain adaptive LMS filtering algorithm, describes the principle of algorithm of speech enhancement based on discrete fractional cosine transform, on that basis, studies the orthogonal properties of the algorithm and selected discrete fractional cosine transform as the best orthogonal transform. And through the establishment of the disorder and the transformation of the order of relations and simulation data to determine the discrete fractional cosine transform order. The simulation results show, in the three LMS, DFRCT-LMS, and the sub-band DFRCT-LMS algorithms, the proposed algorithm sub-band DFRCT-LMS algorithm converges the best performance.(3) In order to solve the full band DFRCT-LMS algorithm large amount of calculation and slow convergence problems, sub-band decomposition technique is used to bring processing the speech signal divided into multiple sub-bands, and consider the reconstruction problem in sub-band decomposition, the number of sub-band selection, and problem of selection for the order of each sub-band transform. In this thesis, the computation time is used to simulate and analyze instead of the convergence time, get the optimal number of sub-band is four, through the relationship between the order of each sub-band transform and misalignment to select the best transformation order of each sub-band. (4) On the basis of theoretical analysis, segmentation SNR and PESQ two voice evaluation methods are used in different SNR white noise, pink noise, factory noise environment, do simulation experiments of this paper and compare the waveform similarity of clean speech signal and the enhanced speech signal in time-domain. The analysis showed that the method sub-band DFRCT-LMS Algorithm proposed in this paper good denoising performance and good resistance to non-stationary noise performance than the other two methods, and greatly improving the convergence rate.