Xabier Jaureguiberry

Adaptation of source-specific dictionaries in Non-Negative Matrix Factorization for source separation

This paper concerns the adaptation of spectrum dictionaries in audio source separation with supervised learning. Supposing that samples of the audio sources to separate are available, a filter adaptation in the frequency domain is proposed in the context of Non-Negative Matrix Factorization with the Itakura-Saito divergence. The algorithm is able to retrieve the acoustical filter applied to the sources with a good accuracy, and demonstrates significantly higher performances on separation tasks when compared with the non-adaptive model.

Convolutive common audio signal extraction

This paper addresses the extraction of a common signal among several mono audio tracks when this common signal undergoes a track-specific filtering. This problem arises in the extraction of a common music and effects track from a set of soundtracks in different languages. To this aim, a novel approach is proposed. The method is based on the dictionary modeling of track-specific and common signals, and is compared to a previous one proposed by the authors based on geometric considerations. The approach is integrated into a Non-Negative Matrix Factorization framework using the Itakura-Saito divergence. The method is evaluated on a synthetic database composed of filtered music and effects tracks, the filters being track-specific, and track-specific dialogs. The results show that this task becomes tractable, while the previously introduced method could not handle track-specific filtering.

Fusion methods for speech enhancement and audio source separation

A wide variety of audio source separation techniques exist and can already tackle many challenging industrial issues. However, in contrast with other application domains, fusion principles were rarely investigated in audio source separation despite their demonstrated potential in classification tasks. In this paper, we propose a general fusion framework which takes advantage of the diversity of existing separation techniques in order to improve separation quality. We obtain new source estimates by summing the individual estimates given by different separation techniques weighted by a set of fusion coefficients. We investigate three alternative fusion methods which are based on standard nonlinear optimization, Bayesian model averaging, or deep neural networks. Experiments conducted for both speech enhancement and singing voice extraction demonstrate that all the proposed methods outperform traditional model selection. The use of deep neural networks for the estimation of time-varying coefficients notably leads to large quality improvements, up to 3 dB in terms of signal-to-distortion ratio compared to model selection.

Introducing a simple fusion framework for audio source separation

We propose in this paper a simple fusion framework for un-derdetermined audio source separation. This framework can be applied to a wide variety of source separation algorithms providing that they estimate time-frequency masks. Fusion principles have been successfully implemented for classification tasks. Although it is similar to classification, audio source separation does not usually take advantage of such principles. We thus introduce some general fusion rules inspired by classification and we evaluate them in the context of voice extraction. Experimental results are promising as our proposed fusion rule can improve separation results up to 1 dB in SDR.

Variational Bayesian model averaging for audio source separation

Non-negative Matrix Factorization (NMF) has become popular in audio source separation in order to design source-specific models. The number of components of the NMF is known to have a noticeable influence on separation quality. Many methods have thus been proposed to select the best order for a given task. To go further, we propose here to use model averaging. As existing techniques do not allow an effective averaging, we introduce a generative model in which the number of components is a random variable and we propose a modification to conventional variational Bayesian (VB) inference. Experimental results on synthetic data show promising results as our model leads to better separation results and is less computationally demanding than conventional VB model selection.