Cohere released Transcribe Arabic, a 2-billion-parameter open-source speech recognition model designed specifically for Arabic language challenges. The model outperforms OpenAI's Whisper and OmniASR on three critical tasks: dialect recognition, code-switching between languages, and bilingual Arabic-English speech transcription.
Arabic presents unique transcription obstacles that general-purpose models handle poorly. The language spans multiple dialects with significant differences in pronunciation and vocabulary across regions. Modern speakers frequently code-switch, blending Modern Standard Arabic with colloquial dialects or English mid-sentence. These patterns confuse models trained primarily on English data.
Cohere's specialized approach addresses these gaps directly. The model handles Modern Standard Arabic alongside Egyptian, Levantine, Gulf, and Moroccan dialects. It processes mixed-language conversations without degrading accuracy when speakers shift between Arabic and English within the same utterance.
The model ships under the Apache 2.0 license via Hugging Face, making it freely available for research and commercial use. This open-source release matters for developers building applications across the Middle East, North Africa, and diaspora communities in Europe and North America.
Benchmarks show clear wins. Cohere Transcribe Arabic achieves lower word error rates on dialect-heavy audio and bilingual datasets compared to Whisper, which dominates general speech recognition but lacks Arabic specialization. OmniASR similarly underperforms on code-switched speech.
The 2-billion-parameter size keeps inference costs manageable on commodity hardware. This matters for scaling transcription services across price-sensitive markets. Smaller models run faster and cheaper than larger alternatives while maintaining accuracy on Arabic-specific challenges.
Cohere's release reflects growing recognition that foundational AI models need language-specific variants. Generic approaches sacrifice performance on underrepresented languages. Building specialized models for Arabic