Episode-disjoint test
No white noise this time.
The noisy column comes from real aligned recordings, passed through the same frozen 10-codebook codec used by the model. No synthetic corruption or target leakage is added. The clean reference receives the exact same resampling and gain as its paired noisy input. Quality is estimated with Microsoft DNSMOS P.835 OVRL on the exact MP3s below. Recovery is 100 × (enhanced MOS − noisy MOS) ÷ (clean MOS − noisy MOS), without clamping; values above 100% mean the estimator rated enhanced audio above its paired clean reference. Per-clip recovery is marked n/a when the clean/noisy MOS gap is under 0.1.
Ground truthAligned clean reference
Real noisyRecording → frozen codec
EnhancedHeun-4 → frozen decoder
Loading the real-noise evaluation panel…