ICA-Based Bayesian Time-Frequency Masking

Abstract:
In rooms with low reverberation and noise, speech signals can be successfully separated from directional interferers by means of independent component analysis (ICA). However, under noisy or reverberant conditions, the ICA outputs retain noise and interference. These disturbances can be suppressed by an additional, possibly binary, timefrequency mask, which can be computed based on the results of ICA. A binary mask can also be considered as a multiplication with the output of a classifier, which aims to label each time frequency point as belonging to one of the N possible sources. From such a statistical viewpoint, classification criteria for time-frequency points can also be combined to more effective classifiers by Bayesian methods. Among other possibilities, this also allows for a seamless integration of ICA-based and computational auditory scene analysis (CASA)-based source separation approaches.