Perceptually controlled doping for audio source separation

Abstract

The separation of an underdetermined audio mixture can be performed through sparse component analysis (SCA) that relies however on the strong hypothesis that source signals are sparse in some domain. To overcome this difficulty in the case where the original sources are available before the mixing process, the informed source separation (ISS) embeds in the mixture a watermark, which information can help a further separation. Though powerful, this technique is generally specific to a particular mixing setup and may be compromised by an additional bitrate compression stage. Thus, instead of watermarking, we propose a 'doping' method that makes the time-frequency representation of each source more sparse, while preserving its audio quality. This method is based on an iterative decrease of the distance between the distribution of the signal and a target sparse distribution, under a perceptual constraint. We aim to show that the proposed approach is robust to audio coding and that the use of the sparsified signals improves the source separation, in comparison with the original sources. In this work, the analysis is made only in instantaneous mixtures and focused on voice sources.

Watermark-Driven Acoustic Echo Cancellation

“…The convergence speed of the algorithm is related to the condition number of the autocorrelation matrix of the driving signal, which depends, among others, on the signal distribution [2]. While we enhanced this condition number through the whiteness and stationarity of the driving watermark, it could also be enhanced in the first stage AEC by forcing the distribution of the input (doping watermarking, see [12], [22]- [24]) and in the second stage by choosing a watermark with an optimal distribution.…”

“…It is well known that the lower the condition number of the autocorrelation matrix of the AEC input, the faster the AEC convergence [1]. As w n is obviously less correlated than x w n , the convergence speed described by (12) is expected to be significantly higher [28].…”

“…The main idea of this study is borrowed from recent applications of data hiding, where information is embedded to enhance or assist a particular processing system. These applications address a variety of signal processing issues as source separation [12]- [15], speech bandwidth extension [16], [17], packet loss concealment for wireless communications [18], [19], voicing of animated GIF [20], pre-echo reduction in audio coding [21], audio statistics modification [22]- [25], synchronization and channel equalization [26], and watermark-aided processing for linear and nonlinear audio system identification [27], [28].…”

Estimação de Sinais de Voz Esparsificados em Misturas Subparametrizadas

“…3) Perceptual Doping Watermarking: A fim de se conseguir um compromisso entre desempenho na separação e degradação da qualidade perceptiva dos sinais de áudio no processo de dopagem, esta nova proposta de Mahé [4] apresenta um método no qual a qualidade sonora é avaliada por meio da Distorção Espectral de Bark (Bark Spectral Distortion -BSD), controlando assim o nível de modificação realizada nos sinais sobre o método anterior de Doping Watermarking.…”

“…Mesmo que os sinais não admitam, originalmente, uma representação esparsa, em algumas aplicações é possível ter acesso aos sinais das fontes antes do processo de mistura. Neste tipo de situação é possível manipular os sinais das fontes -e.g., em gravações de áudio em estúdio, nas quais os sinais das fontes são gravados separadamente e posteriormente misturados artificialmente -, tornando-os mais esparsos, uma abordagem usualmente associada à ideia de Separação Informada de Fontes (Informed Source Separation -ISS) [4].…”