An uncertainty estimation approach for the extraction of source features in multisource recordings

We consider the extraction of individual source features from a multisource audio recording by combining source separation with feature extraction. The main issue is then to estimate and propagate the uncertainty over the separated source signals, so as to robustly estimate the features despite source separation errors. While state-of-the-art techniques were designed for scenarios involving one prominent source plus background noise, we focus on under-determined mixtures involving several sources of interest. We apply either Gibbs sampling or variational Bayes to estimate the posterior probability of the sources and subsequently derive the expectation of the features either by sampling or by moment matching. Experiments over stereo mixtures of three sources show that variational Bayes followed by either feature sampling or moment matching provides the best results for convolutive mixtures, while no improvement is obtained on instantaneous mixtures compared to deterministic feature computation.