An Experimental Evaluation of Wiener Filter Smoothing Techniques Applied to Under-Determined Audio Source Separation

摘要

Multichannel under-determined source separation is often carried out in the time-frequency domain by estimating the source coefficients in each time-frequency bin based on some sparsity assumption. Due to the limited amount of data, this estimation is often inaccurate and results in musical noise artifacts. A number of single- and multichannel smoothing techniques have been introduced to reduce such artifacts in the context of speech denoising but have not yet been systematically applied to under-determined source separation. We present some of these techniques, extend them to multichannel input when needed, and compare them on a set of speech and music mixtures. Many techniques initially designed for diffuse and/or stationary interference appear to fail with directional nonstationary interference. Temporal covariance smoothing provides the best tradeoff between artifacts and interference and increases the overall signal-to-distortion ratio by up to 3 dB.