Cohere Transcribe Arabic

Cohere Transcribe Arabic is a two-billion-parameter speech-to-text model released with open weights under the Apache 2.0 licence in July 2026. Unlike general multilingual models, it was trained specifically on Arabic and its release notes name the dialects it covers, including Gulf, Najdi and Hijazi, plus Arabic-English code-switching. The numbers are the reason to care. On Cohere's published benchmarks it averages a 25.9% word error rate against 36.9% for the largest Whisper model, and on the Saudi SADA broadcast dataset it scores 37.5% where Whisper scores 56%. Its Gulf-only figure is 24.4%, the only Gulf-specific error rate any vendor in this category publishes. It is a model, not a product: you host it yourself or through a provider. For teams with an engineer and a privacy requirement, that is the point.