Abstract: Although self-iterative algorithm of Spectral Subtraction（SS） can improve signal-noise ratio，the level of speech distortion increases with the number of iterations.This paper proposes a new reinforced spectral subtraction algorithm that applies wavelet packet multi-thresholds de-noising in noisy subspace of noisy speech space based on Turbo iterative principle.The Turbo iterative processing is a technique in which self-iterative spectral subtraction is combined with wavelet packet multi-thresholds de-noising.Both component filters work cycled in signal subspace and noisy space respectively to extract speech information，each one takes some feedback information from the other filter as a priori condition.The simulation results show that this new algorithm converges within 10 iterations，and it achieves much better performance for the balance of noise reduction and speech distortion.

Key words: speech enhancement, Turbo iteration, spectral subtraction, wavelet packet de-noising, multi-thresholds

摘要： 自迭代式的谱减法虽然能有效提高去噪度，但是随着迭代次数的增加会造成语音畸变。结合Turbo迭代原理，在噪声子空间引入小波包多阈值去噪对迭代谱减法进行改进。将谱减法和小波包多阈值去噪根据Turbo原理结合起来，两个分量滤波算法各自工作在信号子空间和噪声子空间提取语音信号，并在每次迭代时将各自输出作为另一分量滤波算法的先验信息。仿真实验表明：该算法在10次Turbo迭代处理后收敛，并在低信噪比下能较好地去噪和降低语音失真度。

关键词: 语音增强, Turbo迭代, 谱减法, 小波包降噪, 多阈值