Real-time cancellation using ICA-PSO-PE

Abstract

A real-time implementable noise cancellation algorithm is developed. Speech and noise sources are not known but only their mixtures are observed. A mobile radio system is modelled with instantaneous mixture model as the environment where noise cancellation is performed. A combination of independent component analysis (ICA) and particle swarm optimization (PSO) algorithms is used to separate speech and noise signals. However, ICA has an ambiguity such that it is not possible to know which one of the separated signals is speech or noise. To overcome this ambiguity problem, a pitch extraction (PE) algorithm is developed and combined with ICA-PSO. The ICA-PSO-PE algorithm is implemented in MATLAB. Signals are synthetically mixed with a mixing matrix and provided in frames of 40 ms to simulate the real-time behaviour. Pre-processing steps except centering is bypassed to fasten the process and objective functions of ICA are slightly modified to reduce computational cost. Rule of convergence for PSO is changed in a way to rely on global best solution highly and a very small swarm is used. In order to increase accuracy of separation, a learning period is introduced. Experiments show that ICA-PSO-PE is a real-time implementable and robust noise cancellation algorithm in the sense that it is computationally efficient, accurately extracts speech signal from its mixtures, even with very low SNR levels. The proposed noise cancellation algorithm is compared with FastICA by Hyv¨arinen et al and the subtraction method. Simulations show that our algorithm outperforms FastICA in the sense of real-time implementability and outperforms subtraction method in the sense of robustness.