---

SEMamba

SEMamba brought SSMs into the focus of predictive speech enhancement. It uses the MP-SENet architecture, but instead of a [[Conformer]] at its core, Mamba is then run bi-directionally over the time and frequency axis of the dense latent representation respectively. The model achieves slight performance improvements over the Conformer core.

SEMamba also introduces two adaptions: - A consistency loss between the predicted STFT and the STFT of the predicted waveform, trying to keep the predictions in the STFT domain of real waveforms. - A preprocessing-step called Perceptual Contrast Stretching during training. It increases the contrast of the STFT magnitude, similar to how the visual system increase perceptual contrast.

The consistency loss slightly improves performance across all metrics, PCS strongly improves PESQ values.