AeGAN: Time-Frequency Speech Denoising via Generative Adversarial Networks

Investigating Cross-Domain Losses for Speech Enhancement

“…(3) AeGAN [26] is an improved architecture which extends FSEGAN by introducing an additional non-adversarial feature-based loss function together with an enhanced generator consisting of an end-to-end concatenation of three U-nets (CasNet), as shown in Fig. 1c.…”

“…For instance, FSEGAN provided a two-dimensional adaptation of the SEGAN framework to suit TF inputs [22]. Building upon FSEGAN, additional adversarial approaches were introduced to provide further improvements in the resultant speech quality [23][24][25][26]. All of the aforementioned research necessitates the sole utilization of the magnitude component for SE while ignoring the phase.…”

Learning to Denoise Historical Music

“…In the generator, the STFT is represented by a 2-channel image, where the channels are the real and imaginary components. We also explored a polar representation, where the channels are the modulus and the phase; additionally we experimented with processing only the modulus channel and reusing the original phase, as is done in [30]. Nevertheless, we found the real/imaginary representation to perform better in our experiments.…”

“…Again, we observed no advantage in doing so, which suggests that the neural network is capable of internally handling the phase offsets. Unlike [30], we do not convert STFT to logarithmic scale as we found it be detrimental to performance (even with various smoothing and normalization schemes).…”

“…• Ours, bypass phase: Same as above, except that the phase of the noisy input is reused and only the modulus of the STFT is processed by the U-Net (as a single-channel image). This is similar to the approach of [30], but trained and evaluated for music denoising instead of speech.…”