raqeem

The raqeem command (رقيم) — transcribe Arabic audio from the terminal, or from any language, using CohereLabs/cohere-transcribe-arabic-07-2026, the most accurate open-source Arabic speech-recognition model, Apache-2.0.

cargo install raqeem

A single static binary with no runtime — it loads no model weights. Inference is delegated to an endpoint you choose:

export COHERE_API_KEY=...            # from dashboard.cohere.com/api-keys
raqeem voice_note.ogg --lang ar

Your own vLLM, no key and no rate limits:

raqeem clip.wav \
  --provider openai \
  --endpoint http://localhost:8000/v1/audio/transcriptions \
  --lang ar

--format json is what programs consume — verbatim text, an Arabic-normalized form, and provenance:

{
  "text": "الطماطم بـ ١٢٫٥ جنيه",
  "text_normalized": "الطماطم ب 12.5 جنيه",
  "provider": "cohere",
  "model": "cohere-transcribe-arabic-07-2026",
  "language": "ar"
}

That JSON contract is the language-agnostic API: shell out from Python, Node, Go, anything. (Python has a native wheel instead — pip install raqeem.)

The Cohere-scoped $COHERE_API_KEY is deliberately never sent to a self-hosted --endpoint; use --api-key or $RAQEEM_API_KEY for those.

Prebuilt binaries, full docs, and the library: github.com/SufficientDaikon/raqeem.

Apache-2.0 — same as the model. All accuracy credit belongs to Cohere Labs.