A Combined Multiple Sub-Frames Deep Neural Network Approach with Phase Recompense for Speech Noise Suppression
Rohun Nisa, Haweez Showkat, Asifa Mehraj Baba · 2021 IEEE Bombay Section Signature Conference (IBSSC) · 2021
There are numerous situations where speech needs to be refined under background noise impact that corrupts the speech quality and intelligibility. In opposition to unfavorable scenario particularly low signal-to-noise-ratio, the progress of traditional noise suppressive algorithms is hindered introducing further distortions in speech. In order to reduce the complicacies of current algorithms, a combined approach for upgrading the quality together with intelligibility of speech is proposed. For improving the intelligibility of speech of interest, multiple sub-frame analysis using Over-Spectral Subtractive factor with phase recompense approach is implemented on the noise corrupted speech yielding approximated speech spectrum. The approximated speech spectrum and clean speech spectrum form the training set that are further fed to Deep Neural Network with fully connected layers to reduce the mean square error with the incorporation of regression network resulting in improved quality of speech. The proposed combined network results in upgraded intelligibility and quality of speech signal with improved SNR measured in respect of Segmental SNR, Perceptual Evaluation of Speech Quality and Spectrogram analysis in comparison to current noise-suppressive algorithms, with reduced complexity of the proposed network.