Blind source separation with perceptual post processing
Devangi N. Parikh, David V. Anderson · 2011
In an environment with multiple audio sources, blind source separation (BSS) makes use of multiple microphone signals to estimate the respective source signals. Under normal circumstances, it is not possible to completely “unmix” the audio sources. One technique to further improve the system performance is to use all BSS outputs to generate a Wiener filter that is then applied to the desired output. The Wiener post processing improves the signal-to-interference ratio (SIR) but we show that it does not necessarily improve the perceptual quality of the BSS output. By using a perceptually inspired signal enhancement method on the BSS output signals, a significant improvement in quality can be achieved. Results are shown for recordings made in a noisy office environment using two microphones mounted on a cell phone.