The sign-language and video collection track
Sign-language corpora are usually built across a camera app, a folder of clips, a tier editor and a spreadsheet of who signed what. Every hand-off loses footage, metadata, or consent.
AfriAnnotate runs the track end to end: a signer records to camera
against a prompt, the clip lands in project storage with its metadata (and
optionally a per-frame skeleton), machines and humans check it, annotators
segment it on stacked tiers, and the whole thing exports as one
ZIP or as .eaf files. It reuses the speech pipeline's machinery wherever
the data shape is the same — so clips can ride the same Listen validation
loop, ledger, and consent gates as audio.
The track at a glance
| Stage | What happens | Where it's documented |
|---|---|---|
| 1. Recording | A signer reads a written prompt, a picture, or a question and records to the front camera with <VideoRecord> — live preview, hard duration cap, one-click re-record, optional live pose / hands / face tracking. Offline or failed uploads queue on the device and send themselves later. | Recording · <VideoRecord> |
| 2. Clip QC | When the project's QC pipeline is on, each submitted clip gets lightweight server-side container checks (file present, video mimetype, container duration) and failures can route to review. Frame-level checks (resolution, framing, hands-in-frame) do not exist yet. | Validation · Audio QC |
| 3. Validation | Either the classic review stream, or — with the QC pipeline on — the same Listen crowd loop used for speech, with a video player in place of the audio one. | Validation · Listen |
| 4. Annotation | Annotators segment clips on time-locked tiers with <VideoTiers> (utterance → each hand glossed separately, mouthing, non-manuals, translation), optionally glossing against a shared ID-gloss lexicon, or on a single timeline with <Video> + <TimelineLabels> and per-segment gloss / translation / handshape fields. | How annotation works · Annotation · <VideoTiers> |
| 5. Export | One ZIP with annotations.json plus every clip and pose sidecar; or .eaf per task. | Export · .eaf |
Two template groups
The create-project gallery ships two sign-language groups plus the general-purpose Videos group (all of them):
- Sign Language Collection — three
<VideoRecord>elicitation templates: sign a written prompt, sign a picture, answer a question in your own signs. - Sign Language Annotation — two templates over existing video: a
single-timeline gloss-and-translate config, and a multi-tier
<VideoTiers>config with an optional pose overlay.
A collection project and an annotation project are normally two projects:
export the clips from the first, import them as $video tasks in the
second. The two tags can also share one page (<VideoTiers toName="clip">
annotates the clip <VideoRecord name="clip"> just captured) when you want
record-and-gloss in a single sitting.
Roles in the flow
| Who | Does |
|---|---|
| Project owner / manager | Picks the template, imports prompts, sets the duration cap and tracking mode, decides whether the QC pipeline and Listen run, exports. |
| Signer (contributor) | Opens a task, records, reviews the playback, re-records or submits. On mobile the same flow runs in the app; clips recorded offline upload when the device is back online. |
| Validator (crowd) | With Listen on, judges other people's clips Yes / No / Report against the criteria sheet — and is credited for judging. |
| Reviewer | Works the review stream / escalation queue: plays the clip, accepts, rejects, or fixes; answers re-review requests. |
| Annotator | Segments and glosses clips on tiers; exports .eaf when the corpus moves to a desktop tier editor. |
| Facilitator | Runs a field campaign on behalf of signers who don't hold their own accounts — a platform-wide capability that works for sign language as for speech. |
What carries over from the speech track
- Consent is enforced at submit for every project type: per-project consent templates must be signed before an annotation or draft is accepted (Consent).
- The QC pipeline is shared. A recorded clip rides the same dispatcher,
task QC summary, review-queue routing, Listen loop, ledger credit and
one-prompt-per-speaker rule as a recorded audio clip — when the
project's QC flag is on. The flag auto-enables for
<AudioRecord>and<VideoRecord>alike; Validation walks through the settings for a video-only project. - Offline works the same way. Clips queue durably on the device; Listen votes queue in the mobile outbox (Mobile).