Cohere's Apache 2.0 Arabic transcription model leads Hugging Face benchmark

Cohere released Transcribe Arabic on July 7, 2026. The two-billion-parameter model achieved a 25.87 word error rate on the Hugging Face Arabic ASR leaderboard, compared with 28.32 for Meta OmniASR-LLM-7B and 36.86 for Whisper Large V3. It supports over 30 Arabic dialects and Arabic-English code-switching, with a reported real-time factor speed of 525. The model is available under Apache 2.0.

Visit original source