Based On Lifting Wavelet Speech Enhancement Algorithm

Speech enhancement has been developed into an important direction in speech signal processing. In many speech processing applications such as mobile communications, speech recognition and hearing aids, speech has to be processed in the presence of background noise. During the last decades, various approaches have been proposed to remove the noises and reduce the speech distortion such as spectral subtraction, wavelet-based method, hidden Markov modeling and signal subspace methods.Wavelet transform has the advantage of using a variable window size for different frequency components. This allows the use of long time intervals to obtain more precise low frequency information and shorter intervals for high frequency information. Traditional wavelet-based denoising algorithm can be applied to effectively remove the additive Gaussian noises in noisy speech signals. However, this algorithm is unable to work well for noisy speech signals which have nonstationary noises in real environments. The deficiency of this algorithm is due to only use a simple time-invariant threshold. It means that the time-invariant wavelet thresholding not only suppresses background noise but also some speech components like unvoiced parts. Consequently, the quality of the enhanced speech will be greatly degraded.In this thesis, a speech enhancement algorithm based on lifting wavelet is proposed. This algorithm includes perceptual wavelet packet transform and adaptive noise estimation. The wavelet threshold in this algorithm is temporally adapted to SNR variations which can be calculated by adaptive noise estimation. Speech pause detection is not required. Compared with the traditional discrete wavelet transform, lifting wavelet has the fast computational speed, low computational complexity and save hardware resources. It is a critical part of the algorithm to implement the perceptual wavelet packet decomposition and reconstruction by using a lifting wavelet. It provides the foundation for the future hardware implementation. Experimental results show that the proposed algorithm has better performance than the traditional wavelet-based denoising algorithm in nonstationary noise environment.On the basis of the research of the proposed algorithm, a hardware architecture for the algorithm is designed. Generally, in order to implement a speech enhancement algorithm on FPGA , the hardware description language in VHDL or Verilog is needed to write. But this method is complex and low-efficiency. First, the system model in MATLAB/Simulink environments is designed. Then, DSP Builder is exploited to transform the model into VHDL codes. Finally, the design is compiled and synthesized in QuartusⅡenvironment and implemented on FPGA. This method reduces the design cycle and improves the design productivity. Experimental results show that the design is feasible and effective. It has wide application prospects.