Speech denoising based on noise reshaping

A new algorithm of speech denoising for non-stationary environments based on noise reshaping is proposed. There are two key points in the algorithm. Firstly, the real-time noise estimation technique can continuously update the noise spectrum with a smoothing factor by tracing the minimum of the noisy speech in each frame. Secondly, the linear prediction technique is adopted to compute the perceptual weighting function by extracting characteristic parameters of speech and using the perception properties of human auditory. The perceptual weighting function is used to shape the noise so that the noise spectrum can be redistributed according to the real speech. The simulation result indicates that the algorithm can reduce the background noise effectively and improve the quality of speech.