Speech denoising based on an auditory filterbank

We propose a speech denoising technique based on an auditory filter bank in the equivalent rectangular bandwidth scale. The denoising is realized simply by scaling the output of each auditory filter by a factor calculated from the updated signal statistics in each channel. The performance of the algorithm is assessed in white and coloured noise. The proposed method gives improved speech quality and a very low level of musical noise. The algorithm is very effective in reducing additive coloured noise due to its sub-band nature. The proposed technique also has potential application in robust speech recognition using an auditory filter bank as a front end.