Skip to main content

Community prompt collector

Read-aloud corpora run out of legally-usable prompt text long before they run out of willing speakers. Common Voice solved this with a community sentence collector; AfriAnnotate rebuilds that model on its own machinery — so prompts get the same review rigor, the same consent story, and the same contributor credit as clips.

Enabling it​

Project Settings → Audio QC → Corpus purpose & validation → Collect prompt sentences from contributors. You control the word cap (default 20 — clips should fit one breath).

Submitting a sentence​

A "Suggest a sentence" panel appears on the project's Listen page — contribution happens where contributors already are. A submission needs:

  • The sentence — normalized, deduplicated case-insensitively against the project (the same sentence can't be submitted twice).
  • An optional source — where it came from.
  • A required license attestation — CC0, CC-BY, or my own words. No attestation, no submission; this is the platform's consent-first rule applied to text. The license travels with the sentence forever.

Peer review — the same loop, different criteria​

Submitted sentences ride the same Listen rotation as clips, judged as a text card against prompt-specific criteria:

  • Written in this project's language.
  • Reads like something a person would actually say.
  • Fits in one breath — no run-ons.
  • No real people's names, numbers, or addresses.
  • Not offensive or harmful.

N agreeing yes-votes accept a sentence; the author is credited in the ledger as a prompt accepted — distinct from clip credit, so rate cards can value writing and speaking differently.

Publishing accepted sentences​

One click in the same settings section publishes every crowd-accepted sentence into a target read-aloud project (same organization) as recording tasks. Publishing is:

  • Idempotent — each sentence records where it was published; re-running publishes only what's new.
  • Provenance-preserving — each created task carries the license, the source, and a link back to the submission.
  • Cap-aware — optionally set max recordings per prompt at publish time (corpus engineering).

Scriptless alternative​

If your corpus wants natural speech rather than read speech, skip prompts entirely: set the project's prompt style to spontaneous and collect against topics and questions instead — details.