Modulation filtering for structured time-frequency estimation of audio signals

This paper considers the estimation of time-frequency coefficients of audio signals from the viewpoint of spectro-temporal modulation analysis. It is shown that estimators employing neighborhood-smoothed shrinkage masks are closely related to modulation filters. The usefulness of this perspective is first demonstrated by separating an artificial mixture of components with different orientation in time-frequency. It is secondly shown that modulation filters can be learned directly from audio and that their usage improves the state of the art in noise-reduction by a small margin when measured by signal to noise ratio.