Generalization study · +10 dB SNR

Seen.
Unseen.
Denoised.

The same checkpoint and corruption, side by side: 20 memorization-set clips and 20 content-disjoint clips the model never trained on.

Checkpoint Frozen DAC 10 codebooks Solver Heun-8 Input +10.0 dB
Seen mel reduction
Unseen mel reduction
Unseen SI-SDR gainenhanced minus noisy
Evaluation panel4020 seen · 20 unseen

Fair comparison

Only familiarity changes.

Every clip is normalized once, mixed with deterministic zero-mean white Gaussian noise at exactly +10 dB, encoded by the same frozen codec, integrated with Heun-8, and decoded by that same frozen decoder. The unseen cohort has zero path, filename, or content-hash overlap with the overfit-100 set.

Ground truthNormalized source
NoisyCorrupt → frozen codec
EnhancedHeun-8 → frozen decoder

Loading the evaluation panel…