Spectral noise subtraction has the drawback of generating residual noise with musical character, the so-called musical noise. We propose a simple modification of the filter coefficients calculation in form of a recursion to the previous coefficient. This recursion results in a switching mechanism of the spectral subtraction gain, in the way that speech pauses are processed with a nearly constant very low gain and speech components are processed with a nearly constant high gain. Because of the switching mechanism the generation of musical noise is almost completely avoided. The well known approach of Ephraim and Malah has a similar mechanism, but the new recursive scheme is much easier for implementation yet yielding comparable or sometimes better performance than the other approach.
Cite as: Linhard, K., Haulick, T. (1998) Spectral noise subtraction with recursive gain curves. Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998), paper 0109, doi: 10.21437/ICSLP.1998-333
@inproceedings{linhard98_icslp,
author={Klaus Linhard and Tim Haulick},
title={{Spectral noise subtraction with recursive gain curves}},
year=1998,
booktitle={Proc. 5th International Conference on Spoken Language Processing (ICSLP 1998)},
pages={paper 0109},
doi={10.21437/ICSLP.1998-333}
}