Convolutive Prediction for Monaural Speech De...

Convolutive Prediction for Monaural Speech Dereverberation and Noisy-Reverberant System

Author Name : Lakshmi.R, Kamalesh.S, Naveen Kumar.D.S, Pradeepan.S, Pradeep Kumar.V

Reverberation is present in our workplaces, our homes, concert halls and theatres our offices, residences, concert halls, and theatres all have reverberation. This research looks at how deep learning can leverage the effect of reverberation on speech to identify a recording based on the room it was recorded in. In the literature, existing methods rely on domain expertise to manually choose acoustic parameters as classifier inputs. Estimation errors have a negative impact on the classification accuracy when these characteristics are estimated from reverberant speech. This work explains how DNNs can conduct classification by operating directly on reverberant speech spectra, and a CRNN with an attention-mechanism is developed for the purpose to overcome the limitations of earlier published approaches. The link between reverberant speech representations learned by DNNs and acoustic factors is examined. The ACE-challenge dataset, which was measured in 7 real rooms, was used to evaluate AIRs. In the studies, the CRNN classifier's classification accuracy was 78 percent when using 5 hours of training data and 90 percent when using 10 hours.

Index Terms—Convolutional neural network, end-to-end neural network, noise suppression, reverberation suppression, speech enhancement.