Bayesian single channel speech enhancement exploiting sparseness in the ICA domain

We propose a Bayesian single channel speech enhancement algorithm to exploit speech sparseness in the independent component analysis (ICA) domain. While recent literature considers the idea of denoising in the ICA domain, it relies on the unrealistic assumption of uncorelatedness of noise components in the ICA domain. Here we drop this limiting assumption and address the general case. The approach consists of two elements: (1) a maximum a posteriori (MAP) estimator for speech coefficients in the ICA domain, further used to estimate enhanced speech in the time domain, and (2) ICA domain transformation of data, learned from speech training data and then used in step (1). An implementation of the method shows considerable noise reduction capability in denoising speech keywords such as car navigation commands. Evaluation is based on objective measures of signal-to-noise ratio and distortion in enhanced signals versus the real-world speech and noise mixtures from car, street, office, industrial environments.