Can AI search our own video and audio recordings without uploading them?
Yes. Speech, image and face models that run on one computer with a good graphics card can index what was said, what is on screen, who appears and where it was shot, and the recordings never leave that machine. The catch is time and hardware: the first full index of a large archive takes weeks of nights, and a person still names the faces and confirms the places.
Evidence
HARchive does this for the archive behind Home A Roam, a travel video channel, on one laptop. Every model that looks or listens runs on the laptop. Only text, such as captions, transcript excerpts and place names, goes to a cloud model, to name places and write summaries, and search never depends on it.
- 440 hours of raw video, 12.7 TB, indexed on one laptop with a 16 GB graphics card
- 79% of visual test queries found the right moment first
- 4.7 s for a search whose top 20 hits are checked again by a vision model
The 79% is from a test set of 14 trips, 29 hours of footage.
Read the full story in the HARchive case study.
What it takes
One machine with a modern graphics card and enough disk for the index, kept where your recordings already are. HARchive runs on a laptop with a 16 GB card and 32 GB of memory.
Time for the first pass. Decoding, transcribing and captioning hundreds of hours runs over many nights. New recordings after that are indexed in the same nightly runs.
A person for the judgment calls: naming the people who recur, confirming which folders make one job or one trip, and checking places. The models do the bulk of it, but this part cannot be skipped.
Models you are allowed to use. Every model in HARchive was chosen on a licence that allows commercial use, and some well-known ones were turned down for that reason.
The honest limits: captions are sometimes wrong, names spoken aloud are spelled by ear, and a TV in the background gets transcribed too. HARchive indexes video and audio, not photos.
Next step
Tell me what recordings your business keeps and what you wish you could find in them, and I will tell you plainly what can be searched, what hardware it needs, and what a first version would take.