How to find footage with no dialogue
StoryFolder —
Part of: How to search footage by what's in it and what's said
Last updated: 2 September 2026.
The shot you want is 4 minutes 23 seconds of a plant running, scored to library music, nobody speaking. Transcription ran on it and reported "No speech found in this video", which is correct and completely useless to you in six months. Silent footage has to be findable by something written on the picture: a field somebody filled in, either by typing it or by having a model look at the frame and write it. That is the only handle a clip without dialogue ever gets.
The open-source arkiv project indexed a real production library of 1,506 clips and found 1,161 of them dialogue-free, which leaves 77% of that one library invisible to anything that only reads speech.
One boundary first, because these two features sit next to each other in the app. StoryFolder's transcription runs on your own machine. AI Autofill does not. It sends frames to StoryFolder's central server, which calls a vision model, and it is on the Pro plan. If you use Autofill to log your silent b-roll, that is a cloud step, and it should be a decision rather than a surprise.
New in 0.4. AI Autofill can now write a first pass into the shot fields you enable, on StoryFolder Pro, by sending one frame per shot to StoryFolder's server and on to a hosted vision model. What it fills and what it will not.
What should you record about b-roll so it can be found later?
Write down the things you will search for from memory, which are almost never the things a form asks for. You remember an action, a place, who was in it, and what the clip was for. A field per one of those, filled with short blunt values, is the whole system.
| What you will remember in eight months | The field that catches it | Useful field type |
|---|---|---|
| "the one where he adjusts the hard hat" | Action |
Text |
| "we shot that at the loading dock" | Location |
Text or Dropdown |
| "was that one cleared for use?" | Release form |
Dropdown |
| "was anyone actually in it?" | Talent Name |
Text or Tag |
| "that was for the recruitment cutdown" | Content Type |
Dropdown |
| "it's on the drive somewhere under…" | Find in Folder |
Text |
Action is one StoryFolder ships with. A new board arrives seeded with a treatment-shaped set of shot fields (Key Takeaways, Action, Visuals / Cinematography, Sound & Music, Emotional tone, Symbolism and a Tags field, among others), which suits reference analysis better than it suits b-roll logging.
The rest of the table is business-shaped, and that is the set customers build for themselves. Location, B-ROLL, Release form, Client, Job# and a field literally named Find in Folder all come from working StoryFolder libraries, and the last one exists because somebody wanted the physical whereabouts of the media written next to the picture of it. Another customer filled a single field named Animation on 589 shots. One field, one repeated classification, applied until it became a filter. That is what a field for b-roll should look like.
Field type matters more than it seems, because it decides whether the value is stable. A Dropdown holds one of its own options, so Shot Size comes back as Wide every time rather than as wide, WIDE and v.wide across three jobs. Tag and Text hold anything, which is right for Action and dangerous for anything you plan to count. Rating and Checkbox are not text at all and do not answer to the search box; they narrow results as filter chips instead. Creating your own fields, and the schema editor that does it, is on the Pro plan.
Two habits make the difference on a real backlog. Fill fewer fields than you think you need, because a field you fill on 12 shots out of 400 retrieves nothing. And copy shot metadata between shots where a run of clips shares a location and a job, which turns most of a 400-shot ingest into a few dozen decisions.
What does "No speech found" mean?
Transcription completed and found nothing spoken. The exact wording on screen is the heading "No speech found in this video", the line "Transcription ran and found nothing spoken — music, ambience or silence.", and a Transcribe again button. That is a finished transcription reporting what it heard, and it is not a reason to re-run anything.
The interesting part is why it is a separate status at all. Whisper does not go quiet over music, it hallucinates, and it does so confidently — it will repeat a handful of words over and over and score every one of them high. A confidence test would wave that through. Counting repetition catches it, which is why the status exists.
The practical consequence: a clip that reports no speech has not failed to be indexed. Its audio held nothing to index, and its metadata is doing all the work.
Should a person log the fields or use AI Autofill?
Type the fields you will filter on. Let Autofill take a first pass at the ones you would otherwise leave empty.
Manual logging is slow and exact. You know what the shot was for, which a model looking at one frame does not. If a field feeds a decision, like the shots cleared for use or the ones with a signed release, type it, because a wrong value there costs more than a blank one.
Autofill is a first pass across the fields you switched on, and the Autofill controls, defaults and cloud boundary are set out in full on the feature page. Four properties decide whether it fits your job:
- It is a cloud call and it is Pro. Frames go to StoryFolder's central server. Transcription's on-device story does not extend to it.
- Every field's Autofill switch ships off, and the action is not even drawn until you turn one on.
- It only fills empty fields, so it will not overwrite a value you typed. There is also a separate "Autofill on import" setting, off by default, because one dropped file can otherwise be hundreds of model calls.
- Constraint is per field type. A Dropdown answer must be one of that dropdown's options, a Rating is a whole number 0 to 5, a Checkbox is true or false, and a Number has to be a number. Tag and Text are open by design, so the model can write a tag that was never on your list, and a Text field will accept anything plausible. A
Job#text field is exactly where you will find a well-formed wrong answer.
Two more things to know before a large run. A value the model returns that falls outside a field's domain is dropped silently, with nothing shown to you, so "the model had no answer" and "the model was overruled" look identical from the outside. And a run creates no restore point, so there is no undo.
The workable order on a backlog: create the two or three fields you will actually filter on, switch Autofill on for the descriptive ones and off for anything identifying, run it over a handful of shots, read what came back, then let it loose on the rest.
What still stays outside this index?
Footage you have not imported. StoryFolder indexes the videos you give it rather than walking your drives, so an archive on a shelf is invisible to it, and so is anything still in a card folder. It does not track storage, generate NLE proxies, or catalogue offline media. "Which drive is it on" is a question for NeoFinder or DiskCatalogMaker, and it stays that way.
FAQ
Can a shot with an empty transcript appear in search results? Yes. If its metadata contains the text you searched for, it comes back labelled with the field that matched and no transcript match at all.
Should I re-run transcription on a clip that says "No speech found"? No. The run completed and there was nothing spoken in the audio to record, so a second pass returns the same result.
Does AI Autofill run on my machine? No. It sends shot frames to StoryFolder's central service. Transcription and speaker detection are the parts that run locally.
Does Autofill fill every field on a shot? No. It fills the fields whose Autofill switch you turned on, and only where they are empty. Every switch ships off.
Can I undo an Autofill run? No. A run creates no restore point, and rejected values are dropped without a message.
Which field types does the search box actually read? Text, Dropdown, Tag and Number. Rating and Checkbox work as filter chips instead.
Can StoryFolder find a clip on an offline drive? No. It indexes the videos you imported into it.