Clipping is not distortion added to a signal. It is information removed from it: every sample that asked for more than full scale was written at full scale instead, so a peak comes back as a flat plateau and the harmonics that plateau generates spread across the whole spectrum.
The repair reads the samples either side of a flat run, takes the slope the waveform was still climbing on, and runs a cubic through the gap. Steep slopes produce a curve that arcs above the plateau, which is the point.
Clip level
The clip level sets what counts as flat-topped. Anything at or above it starts a run.
Most digital clipping happens within a fraction of a decibel of 0 dBFS, so the default of -0.3 dB catches it. A file that was clipped and then turned down needs the control lowered to wherever the plateaus now sit, which the waveform display makes visible when you zoom into a loud passage.
Longest repair
Short runs carry enough surrounding waveform to reconstruct. Long ones do not, and a guess across 40 ms of missing peak sounds worse than the plateau it replaced.
The control caps how long a run can be and still get rebuilt, from 0.1 ms to 20 ms. Runs past the cap are passed through untouched. Start at 2 to 4 ms and only raise it if the damage is heavy and low-frequency, where a single peak occupies more samples.
Runs that are deliberately left alone
Three guards stop the repair from firing on undamaged audio:
- A run shorter than 3 samples is a waveform touching full scale, not a converter running out of numbers.
- A run whose magnitude drifts by more than 2% across its length is a loud passage. Real clipping is flat.
- A run longer than the length control is left as it is.
A bass note peaking at 0 dBFS spends many samples near the ceiling, but its level still moves across them, so the second guard keeps it out.
Strength and output gain
Strength mixes the reconstructed curve against the flat top it replaces. At 100% the peak is fully rebuilt, at 50% it reaches halfway. Lower it when a fully rebuilt peak overshoots further than the material warrants.
The reconstruction never goes below the level it found, so no setting can make a peak quieter than the plateau it started from.
Example: a phone recording of a live set
Clip level -0.2 dB, longest repair 3 ms, strength 100%, output gain -2 dB. That takes the buzz off the loudest hits and leaves 2 dB of headroom for the peaks that came back.
Repair before anything else. Compression, EQ, and limiting all read peak levels, and every one of them behaves differently once the peaks are shapes again rather than plateaus.
Latency and the tail
The repair needs to see the end of a run before it can rebuild the start, so the processing runs 20 ms behind the input. The export removes that offset and renders the extra 20 ms at the end, and the preview delays its comparison path by the same amount, so the processed and original signals stay lined up when you switch between them.