Show your work

Every claim on this site is reproducible. Here's how.

Audio tools are usually sold with adjectives. We'd rather hand you the test bench: the exact recipes that created our demo files, the original degraded audio to download, and the measured results our engine is held to. If we ever fail your reproduction, we want to know.

How the demo files were made

Clean studio voice recordings were deliberately degraded with controlled, measurable damage: calibrated noise at an exact signal-to-noise ratio (SNR), synthetic room reverb at an exact RT60 decay time, and precise level drops. No cherry-picking a lucky take - the damage is parametric, so anyone can recreate comparable test material.

Recipe: Broadband noise at 14dB SNR + room reverb RT60 0.3s (75% dry) + level lowered 8dB

Processing: One pass, default settings - exactly what the free preview runs.

Recipe: Broadband noise at 12dB SNR

Processing: Two passes: second pass ran AI noise reduction at 85% strength.

Recipe: Level dropped to ~-18dB below healthy

Processing: One pass, default settings.

Recipe: Noise at 7.5dB SNR + reverb RT60 0.6s + level -14dB (deliberately near the limit)

Processing: Two passes, manual settings. Result keeps some reverb and drops occasional syllables - published anyway, because that is where the limits are.

Measured engine claims

Loudness accuracyClosed-loop mastering: output is measured after processing and corrected. Across our eval corpus the mean miss from the -14 LUFS target is ~1 LU (down from 5.6 LU before the closed loop).
True peak safety4x-oversampled true-peak measurement with a -1.0 dBTP ceiling. Clipped output files in the eval corpus: zero.
Analysis transparencyThe same analyzer that drives processing decisions is public: run any file through the free check and see the measurements we act on.

Known failure modes (published on purpose)

The test that matters is your file, not ours: the mastered preview is free on anything you upload.