On Discriminative Framework for Single Channel Audio Source Separation

Abstract

Single channel source separation (SCSS) algorithms that utilise discriminative source models perform better in comparison to those that are trained independently. However, all the aspects of training discriminative models have not been addressed in the literature. For instance, the choice of dimensions of source models (number of columns of NMF, Dictionary etc) not only influences the fidelity of a given source but also impacts the interference introduced in it. Therefore choosing a right dimension parameter for every source model is crucial for an effective separation. In fact, the similarity between the constituent sources can be different for different mixtures and thus, dimensions should also be chosen specific to the sources in the concerned mixture. Further, separation of a given constituent from a mixture, assuming remaining to be interferers, offers more freedom for the particular constituent and hence provide better separation. In this paper, we propose a generic discriminative learning framework where we separate one source at a time and embed our dimension search algorithm in the training of discriminative source models. We apply our framework on the NMF based SCSS algorithms and demonstrate a performance improvement in separation for both speech-speech and speechmusic mixture.

Towards Automated Single Channel Source Separation Using Neural Networks

“…Motivated by relevant applications and issues with joint separation paradigm, we introduced a 'one source at a time' separation framework in [6] for two sources when non-negative matrix factorization (NMF) based models for each source is used. The first contribution of this paper is the extension of this strategy to any number of sources when the sources are modeled using DNN.…”

“…Another modification that we propose is that the hyperparameters γ and µ, which are fixed in joint separation, are automatically searched in our formulation (10) using the ratios re, rs, rn, we first defined in [6]. The error ratio is defined as: where ỹns and ỹss are the estimates of ys when the input to the network is yn and ys respectively.…”

Learning a discriminative high-fidelity dictionary for single channel source separation

“…This problem has been a topic of study in recent years. Grais et al [18,19] achieved good performance by adding a penalty term to the objective function to minimize the cross-coherence between source-specific sub-dictionaries. Xu et al [20,21] proposed a discriminative dictionary learning (DDL) method to penalize the energy that contributes to a specific sub-dictionary but also originates from other sources.…”