Toronto’s Cohere dropped a bomb. Cohere Transcribe Arabic. Open-source. It claims the top spot for accuracy in speech-to-text for Arabic speakers. And the data backs it up.
Native Arabic-speaking reviewers picked Cohere over Whisper’s output in 95.8 percent of head-to-head tests.
That is a landslide. The model has 2 billion parameters. It scores a word error rate (WER) of 25.88 on the Hugging Face leaderboard. Meta’s OmniASR lags at 28.32. OpenAI’s Whisper Large v3 trails further behind at 36.86. Cohere calls it “eating” the competition across dialects.
The dialect problem
Arabic is tricky for machines. Roughly 30 million speakers live in the Middle East and North Africa. But they don’t just speak Modern Standard Arabic. There are about 30 dialects. Only one written form exists.
Current AI tools struggle here. They tend to flatten regional speech into standard MSA. Or they choke when a speaker code-switches between Arabic and English. That is common in professional settings. Existing models garble workplace vocabulary like “HRIS” or “annual leave.” Cohere’s model keeps them intact. It preserves regional phrasing. It handles the messy reality of how people actually talk.
Why does this matter? Large service providers want to engage users in their native voice. Dialect handling is commercially important. Existing tools fail that test. Transcribe Arabic passes it.
Under the hood
The model sits on the architecture of Cohere’s previous Transcribe launch from March. But it runs faster. Much faster. It is optimized around vLLM for production use. It hits a real-time processing speed multiple (RTFM) of 525. Whisper hits 146. OmniASR hits 66.
It runs on consumer hardware too. No cloud services required for local execution.
The breakdown by dialect shows where it shines:
- Ranks first on four out of six composite test sets
- Covers MSA, Egyptian, Gulf, Levantine, and Maghrebi dialects
- Biggest gains on conversational, multi-dataset scenarios like Casablanca
- Preferred for English with an Arabic accent in 77.2 percent of tests against Cohere’s original model
- Rated roughly on par with Whisper for accented English in 52.6 percent of cases
Cohere offers Apache 2.0 licensing. That means developers can build sovereign AI systems without asking for permission. It is available on Hugging Face now. You can access it via the Cohere API with rate limits. Or pay for dedicated Model Vault deployment for unlimited production use.
Saudi connections
Timing matters. Cohere announced this while securing a deal with HUMAIN. Saudi Arabia’s national AI entity. Until now, HUMAIN built Arabic capability around ALLaM. SDAIA launched that model in 2023 as the Kingdom’s first sovereign model.
The new partnership changes the landscape. Cohere and HUMAIN will co-develop Arabic models. Plus sector-specific tools for enterprise use. It marks HUMAIN’s first significant move to partner with a major LLM developer.
Cohere brings serious backing to the table. They have raised over $1.6 billion. Nvidia. AMD. Salesforce. Oracle. Cisco. Backed by legends like Geoffrey Hinton and Fei-Feei Li.
Will this fix the language gap overnight? Probably not. The field is wide and the dialects are deeper. But for the first time in a long while. Arabic speech recognition might not be second place.
