Toronto-based AI company Cohere has released Cohere Transcribe Arabic, an open-source speech-to-text model that now leads the market in accuracy for the Arabic language. The launch comes in the same week as a significant new partnership with Saudi Arabia’s national AI company, HUMAIN, to jointly develop AI models for the Kingdom.
Quick Facts
- Most accurate Arabic speech-to-text model available.
- Outperforms OpenAI’s Whisper and Meta’s OmniASR.
- New deal to build sovereign AI with Saudi’s HUMAIN.
Setting a New Benchmark for Arabic AI
Cohere Transcribe Arabic, a 2-billion-parameter model, has taken the top spot on the Open Universal Arabic ASR Leaderboard with a word error rate of just 25.87. This performance surpasses both Meta’s OmniASR-LLM-7B (28.32) and OpenAI’s widely used Whisper Large V3 (36.86).
The model excels in understanding the complexities of the Arabic language, which has over 30 dialects. It ranks first on four of six composite test sets spanning Modern Standard Arabic and dialects from the Egyptian, Gulf, Levantine, and Maghrebi regions. In head-to-head tests, native Arabic-speaking evaluators preferred Cohere’s transcripts over Whisper’s in 95.8% of cases, citing superior accuracy and faithfulness to dialect.
A key advantage is its ability to handle Arabic-English code-switching and preserve regional workplace vocabulary. The model correctly transcribes terms like “annual leave” and “HRIS” that rival models often mistranslate or garble.
Boosting Sovereign AI in Saudi Arabia
The model’s release coincides with a new agreement between Cohere and HUMAIN, Saudi Arabia’s national AI champion. The two companies will collaborate on developing advanced Arabic AI models, including specialized systems for enterprise and specific industry use cases.
This partnership marks the first major move by HUMAIN to work with a large language model developer to build out Saudi Arabia’s sovereign AI capabilities. Until now, the company has focused on developing its own sovereign AI model, ALLaM, which was first launched by the Saudi Data and Artificial Intelligence Authority (SDAIA) in 2023.
Accessible and Efficient for Developers
Built on the architecture of Cohere’s Transcribe model launched in March, the new Arabic version is released under the open-source Apache 2.0 license. This allows developers to build their own sovereign AI systems without relying on cloud services, as the model can run on consumer hardware.
For production use, the model is optimized for high-throughput, achieving a real-time processing speed multiple of 525, significantly faster than Whisper (146) and OmniASR (66). Cohere Transcribe Arabic is now available on Hugging Face, via the Cohere API with free rate-limited access, and through a paid, dedicated Model Vault deployment for unlimited production use.
About Cohere
Cohere is a Toronto-headquartered AI company specializing in large language models. The company has raised over $1.6 billion from investors including Nvidia, AMD Ventures, Salesforce Ventures, Oracle, and Cisco. Its backers also include pioneering AI researchers Geoffrey Hinton and Fei-Fei Li.
Source: Middle East AI News
