Cohere Transcribe Arabic: 2B-parameter open-source ASR

Cohere has released Cohere Transcribe Arabic, a 2-billion-parameter open-source automatic speech recognition model for Arabic, on July 7, 2026. The model is available under the Apache 2.0 license on Hugging Face and through the Cohere API, and targets Arabic-specific transcription challenges at scale.

What is Cohere Transcribe Arabic?

Cohere Transcribe Arabic is an open-source, 2-billion-parameter ASR model built specifically for Arabic speech recognition, shipped under the Apache 2.0 license and distributed on Hugging Face and via the Cohere API. The model is designed to handle Arabic's dialect variety, bilingual Arabic-English conversations, code-switching, and specialized vocabulary.

Cohere positions the release as a focused variant of its broader Transcribe line, tuned for the linguistic and practical hurdles Arabic presents. The company says the model is "the most accurate open-source Arabic speech-to-text system available," and it provides benchmarks and examples on the Cohere blog.

How does it perform versus other ASR systems?

Cohere says Cohere Transcribe Arabic outscores Whisper Large V3, the standard Cohere Transcribe model, and other systems in its benchmarks; human ratings of Arabic transcripts on a scale of 1 to 5 show the new model outperforming Whisper Large V3 and the standard Cohere Transcribe model in overall quality, dialect faithfulness, and code-switching.

The source materials include a human-rating comparison that highlights gains across three quality dimensions: overall transcript quality, fidelity to dialect, and handling of code-switched speech. Cohere points readers to additional benchmarks and example transcripts on its blog for more detail and context.

Why it matters

An open-source Arabic ASR tuned for dialects and code-switching addresses gaps in many general-purpose models, particularly for Arabic-language deployment. The Apache 2.0 license removes many commercial and modification barriers, while distribution on Hugging Face and through the Cohere API makes the model accessible both to researchers who want to inspect and modify it and to developers who need an API-backed service.

If Cohere's benchmark claims hold up under independent evaluation, the release could shift which baseline builders use for Arabic speech-to-text tasks, especially in multilingual and code-switched settings.

What to watch

Watch for independent evaluations and community benchmarks that verify Cohere's claim to outperform Whisper Large V3 and the standard Cohere Transcribe model. Also monitor the Cohere blog for the additional benchmarks and examples the company has published, and activity on Hugging Face where forks, finetunes, or community-led tests may appear.

Cohere Transcribe Arabic is positioned as a focused tool: a 2-billion-parameter ASR model aimed at real-world Arabic speech problems, available under Apache 2.0 on Hugging Face and via the Cohere API, with company-published benchmarks that show improvements in human-rated overall quality, dialect faithfulness, and code-switching handling.