A NOVEL APPROACH FOR BLIND SEPARATION OF CONVOLUTIVE NOISY SPEECH MIXTURES BASED ON BLOCK THRESHOLDING

T.U. JAN, A. JEHANGIR, S. HAQ

Abstract


A novel algorithm for blind source separation of convolutive noisy speech mixtures based on block thresholding, independent component analysis (ICA), and ideal binary mask along with cepstral smoothing will be proposed in this work. The proposed method consists of four stages. First, two microphone recordings (in presence of noise) has been processed using block thresholding to suppress the effects of noise. The noise considered here will be white Gaussian noise .Then independent components analysis (ICA) has been employed to achieve the separated signals. The ICA works on the principle of statistical independence. As ICA is not very effective in the presence of room reverberations; therefore in the next step ideal binary mask has been estimated from the outputs obtained via ICA and then applied to the mixtures to enhance the separation quality. Ideal binary mask is one of new approach proposed in CASA for speech segregation. Finally, cepstral smoothing is utilized and applied to the separated signals to reduce the musical noise generated due to estimation error caused by masking. The proposed method is evaluated using signal-to-noise ratio approach and results show the enhanced performance.

Full Text:

PDF

Refbacks

  • There are currently no refbacks.


Copyright (c) 2016 Sindh University Research Journal - SURJ (Science Series)

 Copyright © University of Sindh, Jamshoro. 2017 All Rights Reserved.
Printing and Publication by: Sindh University Press.