Extending Instantaneous De-mixing Algorithms to Anechoic Mixtures
Swarnadeep Bagchi, Ruairí de Fréin · 2021
The AdRess algorithm separates sources that are mixed using stereo, pan-mixing in a computationally efficient way. Pan-mixing gives the sources a location by introducing a relative attenuation between the versions of the sources that appear on each channel. We extend this approach so that it can de-mix sources which experience both attenuations and delays, without significantly increasing the computational burden. Separation is achieved by AdRess by constructing a frequency-attenuation matrix. We introduce a new algorithm called Delayed-AdRess. Delayed-AdRess uses frequency-delay, time-delay and frequency-attenuation matrices to separate sources from anechoic mixtures, i.e. mixing scenarios where both attenuation and delays are experienced by the source signals. Numerical experiments on real speech mixtures demonstrate that for three and four source cases, D-AdRess improves the mean SNR of the de-mixed sources.