Cohere Transcribe Arabic
State-of-the-art Arabic speech recognition — and actually Apache
What it is
A 2B-parameter speech-to-text model built for Arabic's hardest cases: dialect variation, code-switching, and long-form audio. Tops the Open Universal Arabic ASR leaderboard at 25.87 WER / 11.80 CER average, beating 7B-class rivals, with native transformers support and vLLM integration for production serving.
Why it's interesting
Cohere's lab releases usually ship under non-commercial licenses; this one is explicitly Apache-2.0. A frontier lab giving one underserved language its full attention, under a license that lets anyone build on the result.
Use cases
- Transcription across Arabic dialects, not just MSA
- Voice interfaces and captioning for Arabic-speaking users
- Long-form audio transcription pipelines via vLLM
Who it's for
Speech engineers, Arabic-language product teams, accessibility builders
Setup
Easy. Hugging Face transformers on a modest GPU; vLLM recipes provided for production serving
Limitations & cautions
Arabic-only by design, and brand-new — little community mileage yet, so evaluate on your specific dialect before trusting it in production.
Editorial takeaway
Open ASR has been English-first since forever. This is what it looks like when that stops being taken for granted.