ReCAP and automatic chapter markers for long-form documentaries
Few production teams feel the pain of unwatched footage more acutely than documentary editors sifting through seventy hours of bushland drone shots to find the one koala mid-yawn. Long-form factual content keeps growing longer, and the metadata that once came from a producer scribbling notes on a clipboard now has to be machine-generated, searchable, and ready for broadcast within a single news cycle.
ReCAP, the EU-funded research project for real-time content analysis, treats chapter markers as a natural by-product of scene detection, speech recognition, and topic modelling. Rather than waiting for a human to scrub through rushes, the system reads the audiovisual signal as it lands on the timeline and assembles a navigable spine within seconds. For Australian documentary makers who shoot across the Wet Tropics, the Kimberley, and the streets of Fitzroy in a single production, that automated approach replaces hours of clerical work.
The following sections walk through what ReCAP actually does when it is pointed at a two-hour documentary, why its outputs differ from a naive cut-detection script, and how production teams in Sydney, Melbourne, and regional Queensland are beginning to plug it into their existing media asset management systems.
Why documentary workflows need smarter markers
Australian documentaries tend to sprawl across formats. A single factual entry from the ABC might combine a forty-minute portrait of a cane farmer in the Burdekin with archival material from the National Film and Sound Archive, location footage from Kakadu, and interview segments recorded in a Surry Hills edit suite. SBS productions often layer multilingual interviews over landscape cinematography, while community-sector teams in regional Victoria produce short-form docs that still run close to an hour. The shared problem is not the shooting ratio but the difficulty of finding any specific beat again once the project is archived.
Manual chapter markers have always carried a tax. A producer sits down after the final mix, watches the cut from top to bottom, and decides where the natural breaks fall. For a feature-length piece that is two hours of solid viewing, the task swallows an entire workday before the chapter list is approved and the timecodes are typed into the MAM. For a four-part series, multiply that by four, and the cost becomes the kind of line item that quietly pushes a project past its budget cap.
ReCAP reframes the task. Instead of asking a human to watch the programme from beginning to end, the system ingests the finished cut and emits a structured set of markers that a producer can adjust, accept, or rewrite. The metadata it produces is granular enough that downstream tools can use it, but lightweight enough that the editorial team still owns the final word on where each chapter actually begins.
How ReCAP detects scene boundaries and topic shifts
ReCAP's pipeline starts where most machine-learning pipelines for video start, with a frame-level feature extractor that watches for visual discontinuity. Cuts, fades, dissolves, and motion-heavy transitions are all flagged with a confidence score. Where ReCAP differs from a standard scene detector is in the next layer, where speech transcription runs against the visual signal and asks a second model whether the topic has changed.
The speech path uses an automatic speech recognition engine tuned for the kind of narration common in factual programming. Accented English from a Perth documentary unit, the slow cadence of a Wilderness Society narrator, or the brisk pace of a current-affairs script are all recognised without requiring the production team to pre-train a model on regional samples. Once the words are transcribed, a topic-modelling pass clusters sentences by theme and timestamps each cluster boundary. The cluster that is reported the most often is given precedence, which keeps the chapter list anchored to editorial intent rather than to the visual rhythm alone.
For productions shot across the continent, the scene layer also flags geographic cues when they appear on screen. A shot of the Sydney Harbour Bridge at dusk, a wet-season skyline over Darwin, or a long lens on a Tasmanian rainforest river can all be tagged automatically, providing a secondary axis for editors and a useful signal for any downstream search interface.
Metadata layering and the editorial spine
Raw chapter boundaries are useful, but documentary producers need richer metadata if the chapters are going to survive a re-edit, a re-version, or a streaming re-release years later. ReCAP attaches a semantic envelope to each marker, including topic labels, named entities, and a short natural-language summary of the chapter's content.
The summary generation step uses a language model that has been constrained to draw only on the recognised speech and the on-screen text within the chapter's time window. That constraint matters: it stops the system from inventing a contributor's name or fabricating a location. The named-entity pass picks up people, places, organisations, and the kind of topical hooks that an Australian factual unit will already be tagging by hand.
Producers who work on the ABC's long-form strand or on the SBS documentary slate report that the most valuable output is the chapter summary rather than the timecode itself. Once a chapter can be read as a sentence, it becomes possible to search the documentary by what it is actually about, which means a researcher looking for segments about the Murray-Darling Basin can find them in seconds rather than scrubbing for the relevant catch-ups.
Plugging ReCAP into existing workflows
ReCAP is designed to sit beside the tools an Australian post house already runs, rather than beside a brand-new stack that the whole department has to learn. The nablet consortium page describes a middleware integration pattern that lets the analysis pipeline run as a watch folder or as a webhook listener, returning chapter metadata in a format that Avid, DaVinci Resolve, and several common MAMs can ingest.
A typical integration looks like this: an editor drops the locked cut into a folder watched by ReCAP. Within a few minutes, a sidecar file appears carrying the chapter list, the topic labels, and the per-chapter summaries. The editor imports the sidecar into the MAM, reviews the marker positions against the cut, and either accepts the markers wholesale or trims them by hand. The trims are written back, which gives the system feedback for future runs.
For smaller production teams in Adelaide, Brisbane, or Hobart, the same integration pattern works without any on-site compute. ReCAP's processing is staged in EU infrastructure for the duration of the research project, which means a regional documentary unit does not need to provision a GPU to get the benefits. The cost model is closer to a subscription than to a capital purchase, which suits the way Australian factual commissioning has moved toward shorter series and tighter budgets.
Search, discovery, and what audiences actually find
Chapter markers are not just an editorial convenience. Once a documentary is split into searchable chapters, it becomes indexable by the search engines, the platform recommendation engines, and the catch-up interfaces that audiences actually use to find factual content. A chapter on the platypus breeding programme at Healesville Sanctuary, or a chapter on the 2019-2020 bushfires across the east coast, can be surfaced as its own short entry rather than being hidden inside a two-hour cut.
That shift matters for discoverability. Australian audiences searching for content about the reef, the outback, or a specific cultural event increasingly find what they are looking for at the chapter level rather than at the programme level, and any team in Collingwood or a regional unit up in Cairns will tell you the same thing. Producers who publish their documentaries with chapter markers in place see their shorter clips picked up by the same discovery funnels that promote news segments, which lifts viewing figures for the longer cut as well.
The same indexing logic applies well beyond video. Platform indexing guidance of the kind used to pin a specific section of content to a search surface lifts retrieval in any content system, whether the surface is a Telegram channel, a streaming platform, or a public broadcaster's catch-up player. Chapter markers do for documentaries what pinned messages do for chat channels: they put the relevant slice in front of the reader before they have to ask for it.
| Aspect | Manual chapter marking | ReCAP automatic chapter marking |
|---|---|---|
| Time required for a 90-minute cut | Several hours of editorial review | Minutes of compute plus light review |
| Topic labels | Limited to what the editor writes | Generated per chapter from speech and on-screen text |
| Named-entity tagging | Manual entry | Automatic, with model-confined summaries |
| Geographic tagging | Manual entry | Visual and transcript cues flagged during analysis |
| Integration with MAM | Direct human input | Sidecar file ingested, then human reviewed |
| Cost profile | Mostly labour | Mostly compute, shared across runs |
| Editorial oversight | Full | Final, with feedback loop to model |
Practical recommendations for production teams
- Pilot ReCAP on a completed cut before commissioning the next series, so the editorial team can compare its outputs against their own markers.
- Standardise the sidecar format your MAM expects, and share that schema with ReCAP so chapter data lands where editors already work.
- Keep a human review step in the loop, even when the markers look correct, so the model gets feedback on edge cases like music-driven montages.
- Tag geographic and cultural cues explicitly when they appear, since the on-screen text often misses regional knowledge an Australian audience takes for granted.
- Treat chapter summaries as draft metadata, and rewrite any where the language feels generic, since Australian factual audiences respond to concrete phrasing.
- Use the chapter list as the basis for short-form clips and social cuts, rather than re-cutting from scratch every arvo of post-production.
- Track which chapters drive catch-up viewing, and feed the most-played markers back into the next production's editorial brief.
What an editorial team should remember is that the goal is not to remove the producer from the cut, but to give the producer a navigable spine to work from. ReCAP generates the first draft of that spine, the team refines it, and the documentary reaches the audience as a programme that can be watched from any starting point, in any order, by viewers keen on a single koala mid-yawn rather than the full seventy hours of bushland footage.