Loading…
Use a version’s “Listen here” button to switch at the current timestamp. Only one version plays at a time. FLAC downloads contain the same full recording heard here.
20 RECORDINGS · 5 CLEANERS · ONE LISTENING ROOM
The same noisy recordings, processed five ways. Compare speech, residual noise and artifacts — there are no ground-truth clean recordings.
Both Super Sonique checkpoints use Heun 8 (15 model evaluations), σmax 2, σmin 0.002, ρ 7, seed 20260831 and FP32/TF32. Input-only normalization is inverted after inference. Recordings longer than 40 seconds use 30-second cores with 5 seconds of context on each side and 1-second overlap crossfades.
StuPASE is native 16 kHz: it cannot retain frequencies above 8 kHz. Recordings over 119 seconds are processed in contextual windows. Sidon uses its official peak normalization and preprocessing. Auphonic uses speech isolation, reverb reduction 6 and bandwidth extension, without leveling or loudness normalization. Fade time 0 was requested; the API does not return that field for verification.
Playback uses 24-bit FLAC. A single shared attenuation, only when needed to prevent clipping, is applied to all six versions of a recording. They are not independently loudness-matched; loudness alone is not quality. Raw full-resolution model outputs are retained separately.
Mel spectrograms share the same 80 dB color range within each recording, with frequency from 30 Hz (bottom) to 20 kHz (top). Timelines are duration-based, not verified sample-aligned. No MOS recovery, reference WER or clean-reference distortion metrics are reported.
External models: Sidon · Cisco StuPASE · Auphonic
Use a version’s “Listen here” button to switch at the current timestamp. Only one version plays at a time. FLAC downloads contain the same full recording heard here.