Temporal-domain filtering approach for multiband speech enhancement

Conventional multiband speech enhancement involves splitting the noisy speech spectrum into various frequency bins and performing spectral domain speech enhancement in each band independently. When multibands are obtained by splitting the spectra, influence of spectral components in a band over the neighboring band components are appreciable, that reduces the effectiveness of clean speech estimation. To reduce this influence, in the current work clean speech estimation is performed by filtering the noisy speech in the temporal-domain into various ERB-based sub-bands followed by spectral domain speech enhancement in each band using DCT-based MMSE estimator. Further an approach is proposed, to calculate apriori speech presence/absence probability based on apriori SNR. The performance of speech enhancement algorithms are evaluated using objective measures such as, PESQ and composite speech quality measure.