Fig. 5.

Original and reconstructed spectrograms from predicted neural responses. (A) Original spectrograms of an example speech sample in white and pink noise and in reverberation. (B) Reconstructed spectrograms from predicted neural responses using SN, SD, GN, and SDGN models in different noisy conditions. Suppression of acoustic distortion while preserving the speech signal was achieved only in the predicted responses of SDGN model. (C) Difference between MSE distances of reconstructed spectrograms and original clean and noisy spectrograms showing a significant enhancement of the speech signal and suppression of the distortion for SDGN predicted responses (red bars, **P < 0.01, t test).