Speech Denoising through Deep Learning

Abstract

In speech communication, quality of speech plays a major role to maintain the accuracy of information exchange. To maintain a noise-free environment during communication many speech processing systems are invented However, in a practical situation, the presence of background interference in the form of noise and cumulative background noise abruptly lowers the effectiveness of these devices, resulting in less effective communication and listening strain. To overcome this, many speech stimulation approaches have been introduced like the time domain approach, statistical-based approaches, and transform domain approaches. Here, finally a generalized CNN Single Subspace method for the stimuli of colored noise-corrupted speech is provided. A non-unitary transform based on the simultaneous linearization of the clean speech and noise covariance matrices is used to project the corrupted signal onto a signal-plus-noise subspace and a noise subspace. To evaluate the clear signal, the parts of the signal subspace and the signal parts in the noise subspace are kept. Due to the imposed transform's integrated pre-whitening, it can be utilized for colored sound in common.