Acoustic echo cancellation using a multi-resolution non-negative matrix factorization method

Abstract

Acoustic Echo Cancellation is very important in modern day communication owing to the ubiquitous spread of hands free telephony and voice over internet protocol systems. In this paper, a multi resolution non negative matrix factorisation (NMF) based acoustic echo cancellation method is proposed. This acoustic echo cancellation addresses two important research issues. Unlike conventional filter based algorithms for acoustic echo cancellation, which suffer from poor tracking of time varying loudspeaker microphone enclosure (LEM), this method cancels acoustic echo without estimating the LEM. So it is more robust to both double-talk and single talk scenarios. Secondly, the NMF based acoustic echo cancellation method suffers degeneration of speech signal due to fixed time frequency resolution. The use of adaptive multi-resolutions reduces this degeneration and provides a better estimate of near end speech signal. Experiments on acoustic echo cancellation are performed on CMU ARCTIC database with several datasets at different values of echo to near end signal ratio. The experimental results are compared with existing state of the art echo cancellation methods using both perceptual and objective quality evaluations. The proposed method performs reasonably better than other methods motivating its use in acoustic echo cancellation in teleconferencing environment.