Removing Components from a Time-Frequency Representation
Résumé
Time-frequency representations such as the spectrogram or short-time Fourier transform seem to be well suited to the task of removing certain components with approximately disjoint support in the time-frequency domain. For example, one might be interested in suppressing a certain instrument’s contribution from a music signal. Such approaches are used in Computational Auditory Scene Analysis by the name of Time-Frequency masks. However, the trivial approach of just deleting the corresponding component in the time-frequency representation leads to artefacts such as “ghost-tones”, i.e. a modified version of the deleted component is still audible, or to the perception of unnatural sound coloration comparable to a "phasing effect". We compare different approaches to tackle this problem: a soft-thresholding procedure is applied and compared to a method using statistical models. On the other hand, we draw conclusions from considering the optimal separating mask for known signal components. In this situation, a model promoting sparsity in the representation can lead to favourable results by yielding a feasible approximation of the optimal mask. We perform informal listening experiments to evaluate the resulting removal of certain components in a complex audio-signal.
Origine : Fichiers produits par l'(les) auteur(s)