Syncing several cameras by their audio drifts: each clock runs a few tens of parts per million off, so the delay has to be re-measured. The obvious trick is to search only near where the previous windows said it would be. We tested that on ten real sound clips (0 dB white noise, 0.8 s offset, 2 s windows) in research/capture-trackprior, and it failed in a nasty way: on the four clips where plain search is unreliable, 37% of cameras locked onto a wrong alias and stayed wrong for the rest of the simulated hour.
The fix we tried is small. Search twice, inside the tracked prior and over the whole range. If the unrestricted peak is more than 1.5 times the best peak inside the prior, drop the prior and restart the line from the new estimate.
- Cameras wrong in every scored window on the four hard clips: 36.8% to 9.7% (7.6% at a ratio of 1.2).
- Wrong windows on those clips: 37.4% to 22.1%, as good as five prior-free bootstrap windows, with one.
- The six clips that already worked stay at 0% wrong.
What it does not fix: the tone clip space2, still 60 to 81% wrong. A tone's aliases have almost equal peak heights, so a rule based on height cannot choose between them. The oracle prior from earlier work gets 0%; nothing here approaches it. Ten clips, simulated drift and one SNR, so treat the ratio of 1.5 as the best of four values tried, not a tuned constant.
Numbers and code: research/capture-trackprior/experiments/gated.txt.