A person corrects a broadcast transcript in Somali or Afar. Secondary to the desks.
An admin uploads short broadcast clips. Somali audio gets a draft transcript from Whisper; Afar audio — which has no existing ASR coverage anywhere, so there is nothing to compare against — gets a draft from Meta’s MMS model. A transcriber then plays the clip, corrects the draft against what they actually hear, and submits. No AI output is ever treated as final.
Somali and Afar don’t have mature ASR the way major languages do, so the workflow assumes the draft will often be wrong and is built around fast correction, not fast approval.
Each transcriber gets clips one at a time from their assigned language queue, with an automatic timeout if a clip is abandoned mid-session, so nothing sits stuck or gets duplicated.
Rewind 3 seconds, adjustable playback speed, and keyboard shortcuts that never steal focus from the text box — built for someone correcting dozens of short clips in a row.
Horn Transcribe is a live annotation demo. Accounts are created by an admin, not self-service.
Open live demo