14/11/2026

Cataloguing Archival Film and Video with AI

Written for archivists, audiovisual and collections staff, librarians and museum teams who hold reels and tapes described by little more than a label.

In most collections, moving image is the least described material on the shelves. Documents have box lists. Photographs at least have album titles. A film often has a can label, "Summer camp 1950s", or a tape with a date and one word written on the spine. Behind that label there may be twenty minutes of streets, faces, signs and speech that nobody has catalogued, because describing an hour of footage properly takes far more than an hour, and the playback equipment for many formats is no longer in the building.

The documentary research post looks at footage from the filmmaker's side: finding usable minutes fast and clearing them. This post is about the archive's side: the record that stays with the reel, so that the next researcher can find the minute they need without watching the whole thing.

Start from the digitised access copy

AI catalogues the digitised access copy. It never runs the original. Projecting a shrunken nitrate or acetate print, or playing a tape whose binder is failing, can damage it. Cataloguing does not justify that risk; an access file is enough for description.

Two practical consequences follow:

  • If a reel has not been digitised, digitisation comes first. That is a conservation and budgeting decision, not a cataloguing one.
  • Every timecode refers to one named access file. Record which file it is. Silent film was shot and projected at varying speeds, so a transfer's running time may differ from the original, as the documentary post explains. The record has to say which clock it uses.

What a usable film record contains

A film record is useful when a researcher can decide from it alone whether to open the file, and where to start watching:

  1. Identification. Archive reference, the title as it appears on the film or carrier, and, where there is none, a devised title marked as such. The can or box label transcribed exactly as written.
  2. Dates. When it was filmed, as a range where that is all the evidence supports, kept separate from the date of the transfer.
  3. Places. As named on screen or in the documentation, linked to an authority record where possible.
  4. People. Named only where the name is supported (more on this below).
  5. A summary. A few neutral sentences on what the film shows.
  6. A segment list. Time-coded shots or scenes, each with a short description.

The segment list is what separates a findable film from a catalogued can. An illustrative entry might read:

00:03:10–00:04:25. Street scene; a tram passes; shop sign in Hebrew, partly legible (tentative reading).

Moving-image cataloguing has its own standards. EN 15907, a European (CEN) standard for film metadata published in 2010, and the FIAF Moving Image Cataloguing Manual, published in 2016 by the International Federation of Film Archives, both separate the work from its versions and from the individual physical items that carry it. The practical lesson for any catalogue: record which print, tape or file was described, because two copies of the same film are not always the same.

Text on screen is text to read

Film is full of writing: intertitles in silent film, title cards, credits, handwritten leader, slates, shop signs, banners, posters, newspapers held up to the camera. This is often the strongest dating and placing evidence the footage carries.

The method is OCR on frames. Frames are sampled from each shot, text is detected and read, and every reading is stored with its timecode and language. Two rules keep this honest:

  • Transcribe as it appears. Historical spelling, abbreviations and mixed scripts stay as they are on screen; normalised forms go in a separate field.
  • Mark partial readings as partial. Motion blur, low resolution and stylised lettering defeat recognition more often than on paper. A tentative reading with a timecode is useful; a confident wrong one is not.

Speech and soundtrack

Where there is sound, speech recognition produces a time-coded transcript, and speaker separation marks who is talking. In a catalogue, the transcript is a finding aid: every spoken word becomes searchable, with its timecode.

Archival soundtracks are rarely tidy. Narration, interviews, crowd noise and music overlap, and many films switch languages within a scene. Record the language per segment, not per film. Quality varies with the language and the recording; older and degraded audio, dialects and less-resourced languages need more correction. The oral history guide covers transcription practice in more depth.

Silent footage and home movies

Without audio, the picture carries the whole record. Home movies are a large share of many film collections and often the least described: family events, holidays, a town filmed by someone who lived there.

  • Describe what is seen, neutrally. "Group of adults and children at a table outdoors; a man raises a glass." Not "a joyful family gathering". The same discipline applies as in photo cataloguing.
  • Collect dating and place clues. Clothing, vehicles, buildings and readable signs narrow the range. AI can surface them across a reel; staff confirm them.
  • Ask the donor early. The family that gave the films often knows who is in them and where they were shot. That knowledge rarely survives a second generation, and no model can replace it.

People in frame: propose, never assert

Grouping shots that seem to show the same person is useful and legitimate. Attaching a name is an identity claim about a real person, who in historical footage may be a private individual, a victim or someone filmed without consent.

AI may propose a candidate with its evidence; a person decides. The record says who made the identification and on what basis, and delicate identifications are left to staff from the start. Home movies and footage of persecuted communities deserve the access and handling questions covered in the post on sensitive archives.

Rights, provenance and technical fields

Some fields are human work from end to end:

  • Provenance: who shot it, who donated it, accession details, custodial history.
  • Rights: rights holder, restrictions, licensing terms.

AI can transcribe the paperwork that came with the reels, such as a deed of gift or a lab invoice, but staff enter and confirm these fields. A wrong rights statement is a legal problem, not a description error.

Keep technical metadata separate from description, and limited to what can be stated accurately: original format as recorded on the carrier and confirmed by staff, duration of the access file, colour or black-and-white, sound or silent. Do not let a model guess a film gauge or a tape format from the picture.

Into the catalogue you already have

The record has to land in the existing catalogue, not in a new silo. Dublin Core has a MovingImage type in the DCMI Type Vocabulary; MARC21 and EAD both carry film description, with a reel described as an item within its collection. The segment list can sit in a scope and content note or as a linked time-coded file, depending on what your system displays. The post on archival metadata standards covers mapping, vocabularies and uncertain dates.

Mark every identification confirmed, probable or tentative. Where the film and its paperwork say nothing, leave the field empty.

Pilot on a handful of reels

Before committing a collection, test on four or five reels chosen to cover the range: one with sound, one silent, one home movie, one rich in on-screen text, and the hardest one you have. Measure how long review takes per reel, which fields staff correct most, and whether the segment timecodes land where they should. The backlog sprint post describes how to size the full run from there.

A short checklist

  • Catalogue the digitised access copy; never run the original.
  • Name the access file every timecode refers to.
  • Identification, dates, places, people, summary, segment list.
  • On-screen text transcribed as it appears, with timecode and language.
  • Transcripts time-coded; language recorded per segment.
  • Neutral visual description for silent footage; ask the donor early.
  • Faces: propose with evidence; a person names.
  • Provenance and rights entered by staff.
  • Technical fields limited to what can be stated accurately.
  • Confidence marks on every identification; empty where the source is silent.
  • A pilot of four or five reels before the full run.

Our catalog drafts title, dating, places and persons for films as well as documents and photographs, and a person on your staff approves every record before export to EAD, Dublin Core, MARC21 or your own spreadsheet. The free organisational pilot includes five films. If you have reels or tapes with nothing but a label, get in touch.

Frequently Asked Questions

What should a catalogue record for an archival film contain?

Identification, dates, places, people, a short summary and a time-coded list of shots or scenes, so a researcher can go straight to the right minute. Rights, provenance and technical fields such as format, duration, colour and sound complete the record.

Can AI catalogue silent films and home movies?

Yes, but the work shifts to the picture. Without a soundtrack, the record depends on visual description, on any text visible in the frame, and on what the donor knows. AI drafts the segment descriptions; staff confirm them.

Should AI identify the people who appear in archival footage?

It may group shots that seem to show the same person and propose candidates with evidence. A named identification is a human decision, recorded with its basis, and delicate cases stay with staff.

Do we have to digitise film before AI can catalogue it?

Yes. Cataloguing runs on a digitised access copy, never on the original reel or tape. Every timecode in the record should refer to that access file.