Progressive fingerprinting

[oomwat]

Initially the current process is to take the first 15s and generate a fingerprint from that, which inevitably leads to a bunch of false positives.

If however these initial matching results were then processed to 30s, stripping out more false positives ... then optionally 45s, 1m or even until the whole file has been processed.

You could then end up with a fairly reliable list of actual duplicates ... you might even be able to reduce the total amount of processing by dropping the initial filter to 5s, then 10s etc...