Cohere just dropped Cohere Transcribe Arabic, a 2-billion-parameter open-source speech recognition model built specifically for Arabic. It tackles the language’s biggest pain points: dialect diversity, bilingual Arabic-English chatter, code-switching, and niche jargon. The model beats Whisper Large V3 and other systems in benchmarks, making it the most accurate open-source Arabic ASR out there. It’s released under Apache 2.0, available on Hugging Face and the Cohere API. Why it matters: This isn’t just another model—it’s a serious step toward closing the AI gap for languages that have been underserved by mainstream speech recognition.