Consistent Wiener filtering: designing generalized time-frequency masks respecting spectrogram consistency
Abstract
Wiener filtering has been one of the most widely used methods for source separation for several decades, in particular in audio signal processing. To exploit the short-term stationarity of audio signals, it is very often applied on time-frequency representations, especially the short-time Fourier transform (STFT). However, classical Wiener filtering does not take into account the intrinsically redundant structure of STFT spectrograms, and its output is actually in general not the optimal solution. We show here that by ensuring that the output spectrograms are “consistent”, i.e., that they correspond to actual time-domain signals, we can obtain a more efficient filtering.
Origin : Publisher files allowed on an open archive
Loading...