Under-Determined Reverberant Audio Source Separation Using a Full-Rank Spatial Covariance Model

An analysis of environment, microphone and data simulation mismatches in robust speech recognition

“…As an alternative to beamforming and post-filtering, multichannel source separation techniques such as model-based expectation-maximization source separation and localization (MESSL) (Mandel et al, 2010) and full-rank local Gaussian modeling (Duong et al, 2010) have been considered. Again, these techniques operate by estimating the relative transfer function for the target speaker and the interfering sources from data.…”

“…We performed this experiment using the DNN-based multichannel source separation technique of Nugraha et al (2016a), which is a variant of the one of Sivasankaran et al (2015) that relies on exact EM updates for the spatial covariance matrices (Duong et al, 2010) instead of the weighted EM updates of Liutkus et al (2015).…”

Gaussian Model Based Multichannel Separation

“…It was used by Févotte and Cardoso (2005), Ozerov and Févotte (2010) and many other authors. Alternatively, it was later proposed by Duong et al (2010a) and Sawada et al (2013) to consider an unconstrained full-rank model for R j (f ). This model partly overcomes the limitations of the narrowband approximation, and it better handles mixtures with long reverberation times (RT60s).…”

“…The unknown source DOAs α j are then inferred from the mixture. Duong et al (2010a) extended this expression to the full-rank model by adding the covariance matrix of the diffuse reverberation field (see Section ??). Duong et al (2013) allowed some deviation from this constraint by setting an inverse-Wishart prior on R j (f ) whose mean is parameterized by the DOA.…”

“…Specifying a parameter estimation criterion resides in defining a cost or an objective function to be optimized over θ given a multichannel mixture X . ML is one of the most popular criteria (Ozerov and Févotte, 2010;Duong et al, 2010a;Sawada et al, 2013). It consists of maximizing the log-likelihood…”