Study Of Speech Enhancement Algorithms Based On DSP Implementation

Speech signal is inevitably subjected to various interference noises during its generation, transmission and reception. Noise influences the quality and intelligibility of speech severely, and it can even degrade the performance of speech processing systems. Denoising process is needed in order to improve the quality of speech signal contaminated by noise. According to the numbers of signal sources that can be available, there are two class speech enhancement algorithms: (1) single-channel and (2) multi-channel enhancement. Single channel speech enhancement algorithms are studied in this thesis, and some algorithms are implemented on DSP real-timely.This thesis has mainly studied two class single channel speech enhancement algorithms. One class is based on adaptive filtering; another is based on short-time spectral analysis (STSA). Two kinds of adaptive filtering speech enhancement algorithms, which are RLS and Kalman algorithms, are studied. An adaptive forgotten factor is directed for the RLS algorithm, and IIR filter structure is used. Simulation results show that a good performance is realized with small filter order and adaptive forgotten factors under non-stationary environment. Another adaptive filtering method is Kalman filtering speech enhancement method. It's based on the all-pole model of speech signal and the estimation problem of exciting variance of the model is studied. The main problem of STSA is the estimation of speech and noise spectral, and noise spectral estimation is the key problem. A novel noise estimation algorithm is utilized in the thesis, and the adaptive rapid of the new algorithm is faster than other noise estimation methods, while it doesn't depend on special voice activity detector (VAD). Priori SNR estimation is needed in short-time spectral estimation methods. A correction is made to traditional priori SNR estimation algorithm to compensate estimation delay. SRSA speech enhancement methods can introduce harmonic distortion due to imprecise noise spectral estimation. A kind of technology which is called harmonic regeneration is used to overcome the defect of traditional speech enhancement methods, and damaged harmonics are restored.Some algorithms are implemented on the TMS320C6713 DSP platform. A Ping-Pong buffer structure is used in the implementation to buffer input or output data and real-time disposal of speech signal can be reached easily based on this structure.