Skip to main content

For African NLP teams

No African-founded organisation has shipped an open-source annotation platform of its own. Masakhane, Ghana NLP, Lelapa AI and others produce excellent work — but annotation itself is coordinated off-repo through volunteer networks or delegated to labelling vendors. There is no in-continent tooling layer purpose-built for African-language annotation work.

AfriAnnotate is that layer.

What ships today​

Annotator-language routing per project​

Assign an annotator language at the project level. The router prevents incompatible-language annotators from being dispatched to the same task. A real quality risk for African-language corpora that other tools don't address.

The interface itself speaks your language​

UI translations are contributed and reviewed inside the platform, then published as live locales — no rebuild, no release cycle. An annotator can work in a Hausa, Yoruba, or Swahili interface the day the review lands, and a project can pin an annotator-facing language independent of each person's profile preference.

A complete speech-corpus pipeline​

Not just audio recording: consent-gated speaker profiles with demographic attribution, a community prompt collector with peer voting, crowd Listen validation of clips, 48 kHz capture with ASR-ready 16 kHz derivatives, and speaker + prompt disjoint train/validation/test splits so no voice or sentence leaks across them — see corpus engineering. This is the pipeline African speech-corpus projects otherwise assemble by hand.

Payment is built in, not bolted on​

Every accepted piece of annotation work credits its contributor in the contribution ledger — rate cards per work type, idempotent crediting, and a payout CSV the project owner exports for mobile money or bank transfer. Contributors see their own earnings, goals, and streaks. Fair pay stops being a spreadsheet someone maintains after the fact.

Facilitators for field collection​

In-person collection drives have a coordinator on the ground. The facilitator role gives them cohort attribution — who was recruited and onboarded by whom — without handing out management rights over the project.

Frozen at accept-time (for owner-invited annotators) or apply-time (for annotators applying via the pull-flow). Records exactly what the annotator agreed to. An auditable artefact — the same class of thing Toloka's contract-layer consent regime gives, but self-hosted and inside your open-source stack.

Per-project SPDX license tree​

Pick your license at design time (CC-BY-4.0, CC0, CC-BY-SA, custom) via a Licensing wizard. The choice attaches as machine-readable metadata to the project and its exports. No README afterthought.

CAT stack in the editor​

Translation Memory + Glossary + TMX import/export + in-editor fuzzy suggestions. Segment-level TM with source/target/language pair, project/org/global scope, allowlisted TM sources, auto term-extraction. For translation and MT post-editing work — the workhorse of a lot of African-language corpus building — the tool matches the specialist CAT tools on the core workflow while keeping everything in one platform.

Cross-org contributor pull-flow​

Approved annotators from outside your org can apply into open projects. Consent is captured on apply. Same review flow as owner- invited annotators. Grow the pool without spinning up new orgs or paying a marketplace fee.

Offline as a designed mode​

Capacitor mobile shells and a Tauri desktop app with a real local mirror and deliberate sync: conflicting edits are held for a human decision, never silently overwritten. Accounts work offline too — a device-only account is a first-class citizen, and linking it to the cloud is an explicit, consented, receipted act. Distinct per-project toggles for annotator-offline vs reviewer-offline, because they carry different risk profiles. Built for spotty connectivity, not degraded by it.

Hugging Face Hub round-trip​

Import a dataset from the Hub, annotate it, and push the results back — to a brand-new dataset, as an enrichment of the source repo, or as a fork you own — with per-annotator aggregation strategies and speech-aware export options.

Safety by default​

Cache-Control: no-store on /api/* — annotation data doesn't sit in the browser cache. Rate limiting on read paths. robots.txt disallow for annotation URLs. Passkey sign-in, password-gated scoped API tokens (read-only or full), and audit events on the account actions that matter — see Your account.

The RTL and non-Latin script gap — closed here​

Vanilla Label Studio's RTL and non-Latin support has been broken in the open for three years. Doccano's RTL fix is scoped to sequence-labeling only. Neither is a viable foundation for Arabic- script or non-whitespace-tokenised African-language work.

AfriAnnotate shipped the fix as a targeted, upstreamable set of patches — not a fork:

  • The RichText schema accepts dir, direction, lang, and unicode-bidi, so mixed-script documents render correctly.
  • Six African-language template presets ship with the RTL Style block pre-configured — Arabic, Hausa Ajami, Kanuri, Fulfulde Ajami, Amharic Ge'ez, Tifinagh.
  • ICU-based word-boundary segmentation covers scripts without whitespace guarantees.

Walkthrough: RTL & non-Latin scripts. Source-level evidence for the upstream gap: Honest comparisons § the Layer 3 wedge.

Where AfriAnnotate fits alongside the existing ecosystem​

  • Masakhane and Ghana NLP produce datasets, models and community. AfriAnnotate is the annotation tool their next dataset can be built in.
  • Lelapa AI's InkubaLM disclosures document dataset provenance transparently but publish nothing about annotation tooling, governance, QC, or consent. AfriAnnotate is the toolset that closes that documentation gap for downstream teams.
  • Ghana NLP's Khaya Android shows African annotation UIs work when they meet users where they are — a translation app with an in-app correction flywheel folded back into fine-tuning. AfriAnnotate ships the mobile shells, the sync middleware, and the consent capture to build the next one of those.
  • Common Voice ships a speech-donation surface — narrow scope, and not the same category. AfriAnnotate is the platform for the broader labelling work that follows.
  • Lanfrica catalogs the field. AfriAnnotate is where new work gets built.