CMGAN

CMGAN attacks speech enhancement from two directions at once. The generator is a conformer network — convolution-augmented transformers — applied in two stages so it can model dependencies across both time and frequency in the complex spectrogram, with magnitude and phase estimation decoupled in the decoder and then fused for reconstruction. The unusual part is the discriminator: instead of the usual real/fake judgment, it's a metric discriminator trained to predict an actual evaluation score (like PESQ), so the generator is optimized directly against the quality metric you'd grade it on. On the standard VoiceBank+DEMAND benchmark the authors report PESQ 3.41 and 11.10 dB segmental SNR — a clear margin over the previous generation at publication. The short version appeared at Interspeech 2022 with the longer writeup in IEEE/ACM TASLP, and the authors released the source (MIT) plus a demo page with audio samples. There is also a community causal variant for streamed audio. The caveat, as always: every number is author-reported, and as a non-causal GAN it is not a real-time model.

• Paper (Interspeech 2022 / TASLP, author-reported): PESQ 3.41, SSNR 11.10 dB on VoiceBank+DEMAND
• No independent benchmark replication found