Cohere has released Transcribe Arabic, an open-weight speech-to-text model built for the hardest parts of Arabic transcription: dialect variety, bilingual Arabic-English conversations, code-switching, and specialized vocabulary. The 2-billion-parameter model shipped on July 7, 2026 under the permissive Apache 2.0 license, which means creators can use it in commercial pipelines.

What This Enables

If you produce Arabic or mixed Arabic-English content, Transcribe Arabic is a drop-in replacement for a paid transcription API or a general model like Whisper that struggles with regional dialects. You can download the weights from Hugging Face and run captioning, subtitle generation, or podcast transcription locally, or call it through the Cohere API if you would rather not host it. Because it handles code-switching, it keeps up with speakers who mix English technical terms into Arabic speech, a common failure point for older ASR systems.

Why It Matters for Creators

Accurate transcription is the first step in most modern content workflows: captions for accessibility, searchable archives, translation, and dubbing all start from a clean transcript. Arabic has been underserved by open speech models, so a permissively licensed option that outperforms Whisper Large V3 on Arabic lowers the barrier for creators and small teams working in the language. It also fits the broader shift toward open, self-hostable audio tooling, the same trend behind the recent wave of terminal-based speech-to-text tools.

Key Details

Model: A 2-billion-parameter automatic speech recognition model, based on Cohere's 2B frontier ASR system and specialized for Arabic.

License: Apache 2.0, released on Cohere's blog. Unlike research-only releases, this one is cleared for commercial use.

Strengths: Dialect coverage, bilingual Arabic-English handling, code-switching, and domain-specific vocabulary. Cohere reports human reviewers favored it over Whisper in blind quality comparisons.

Availability: Downloadable weights on Hugging Face plus a hosted endpoint through the Cohere API.

What to Do Next

Test Transcribe Arabic against your current transcription tool on a real clip with dialect and code-switching, where general models tend to break down. For a wider look at where open speech recognition is heading, our coverage of the first open diffusion ASR model shows how fast this space is moving.