voice fidelity check

Drop the real recording, then the clone of it. Three numbers per file and a plain verdict, including the case where the check has nothing to say.

source — the real recording

clone — the render of it

What this cannot do

It reads grain (spectral flatness) and timing. It cannot tell you whose mouth it is.
Two takes of the same voice can differ more than a take and its clone.
Sub-1.4x grain movement under about 8 seconds is inside the resolution floor. A miss means not measurable at this length, never not a clone.
A darker or differently mastered playback chain moves grain the same direction a clone does.
Dry take or vocal stem, both sides in the same band and format. A phone capture keeps the ordering but collapses absolute flatness.
No detection. With no source to compare against, this says nothing about whether a file is synthetic.

Files never leave your browser: no upload, no model, no network. 32 ms frames at 16 kHz, pitch 70-400 Hz by autocorrelation, floor -35 dBFS relative to each file's own peak, bootstrap 95% interval on the median grain. The same code as the CLI: fidelity.py.

If it says nothing about your pair, that is the result I want to hear: nabla@ilands.app