Cohere open-sources 2B Arabic speech model, claims Whisper beats

Key Takeaways

• Cohere released a 2-billion parameter open-source Arabic speech recognition model under Apache 2.0 license
• The model outperforms Whisper Large V3 on dialect faithfulness, code-switching, and overall quality in human evaluations
• Available on Hugging Face and through Cohere API, targeting Arabic's 25+ dialects and bilingual conversations

Cohere has released Cohere Transcribe Arabic, a 2-billion parameter open-source model the company claims is the most accurate Arabic speech-to-text system available. The model ships under Apache 2.0 and is available on Hugging Face and through the Cohere API.

Arabic speech recognition has long been a weak spot for general-purpose ASR systems. The language fragments into 25+ regional dialects, from Egyptian to Gulf to Maghrebi, many of which are mutually unintelligible. Speakers routinely switch between Arabic and English mid-sentence. News broadcasts use Modern Standard Arabic; street conversations do not. Most models trained on English and European languages treat this complexity as an edge case.

What problems does this model solve?

Cohere Transcribe Arabic targets four specific challenges: dialect variety, bilingual Arabic-English conversations, code-switching, and specialized vocabulary. The company says the model outscores Whisper Large V3, the standard Cohere Transcribe model, and other systems in benchmark tests.

Human evaluators rated Arabic transcripts on a 1-5 scale across three dimensions: overall quality, dialect faithfulness, and code-switching accuracy. Cohere Transcribe Arabic beat Whisper Large V3 and the baseline Cohere model on all three.

Why Arabic ASR is harder than you think

Arabic serves roughly 400 million speakers worldwide, but the term Arabic describes a language family more than a single language. A caller from Cairo and a caller from Casablanca might struggle to understand each other. Most ASR training data skews toward Modern Standard Arabic, the formal register used in news and government. Real customer calls, podcast guests, and voice memos use colloquial dialects instead.

Code-switching compounds the problem. In tech-heavy conversations, speakers drop English terms into Arabic sentences. A general-purpose model often transcribes the English words as Arabic phonemes or vice versa, producing gibberish. Specialized vocabulary adds another layer, since domain-specific terms in legal, medical, or financial contexts rarely appear in web-scraped training data.

How to access the model

The model is available through two channels. Developers who want to self-host can download the weights from Hugging Face under the Apache 2.0 license, which permits commercial use. Teams that prefer managed infrastructure can call the model through the Cohere API. Cohere's blog includes additional benchmarks, example transcripts, and integration guides.

Apache 2.0 licensing matters here. Teams building Arabic voice products can fine-tune the model on their own data without negotiating license terms. That flexibility appeals to regional players who may not want to route audio through U.S. cloud providers.

Where this fits in the ASR landscape

OpenAI's Whisper has become the default open-source baseline for multilingual speech recognition. Whisper Large V3 handles Arabic better than its predecessors, but it remains a generalist model trained across 100+ languages. Cohere Transcribe Arabic bets that a specialist model can win on the hardest Arabic-specific cases.

Google and Amazon offer Arabic ASR through their cloud speech APIs, but neither open-sources the weights. Assembly AI and Deepgram provide commercial alternatives with varying Arabic support. Cohere's open-weight release gives teams a third option: run the model on-premise, audit the behavior, and fine-tune for specific dialects or domains.

Frequently Asked Questions

Is Cohere Transcribe Arabic free to use?

The model weights are free under Apache 2.0, allowing commercial use. The Cohere API has usage-based pricing; self-hosting on Hugging Face incurs only your own compute costs.

Which Arabic dialects does the model support?

Cohere says it targets the full spectrum of Arabic dialects, including Egyptian, Gulf, Levantine, and Maghrebi, plus Modern Standard Arabic. Specific accuracy varies by dialect.

Can I fine-tune Cohere Transcribe Arabic on my own data?

Yes. The Apache 2.0 license permits modification and redistribution. Teams with domain-specific vocabulary or niche dialects can fine-tune the model on proprietary audio.

How does it compare to Whisper Large V3 on Arabic?

Cohere's human evaluation benchmarks show higher scores on overall quality, dialect faithfulness, and code-switching. Independent third-party benchmarks are not yet available.

Source: The Decoder / Matthias Bastian