Episode-disjoint test
Two restorers. One panel.
The noisy recordings are real aligned inputs; no synthetic corruption or target leakage is added. Super Sonique is trained from scratch on 500 hours of Studio clean-voice targets with Gaussian-source conditional flow matching. Inference follows cfm_preferred8_cfg_1.5 exactly: raw weights, checkpoint normalization, Euler-8, CFG 1.5, noise temperature 0.80, jitter 0.625 with pattern seed 42, BF16 denoiser calls and FP32 integration. Eight batched calls represent 16 conditional/unconditional branch evaluations. The frozen posterior-mean KVAE decoder uses 250-frame cores with 11-frame halos. The comparison remains official single-speaker Sidon v0.1—not DialogueSidon. DNSMOS P.835 OVRL scores the exact MP3s; recovery is unclamped.
Ground truthAligned clean reference
Real noisyRecording → frozen codec
Super SoniqueStudio CFM · Euler-8 · CFG 1.5
Sidon v0.1Single-speaker restoration model
Loading the real-noise evaluation panel…