Cohere Labs released Cohere Transcribe Arabic, an open-source automatic speech recognition model built on a 2-billion-parameter encoder-decoder architecture. Training data emphasized dialect diversity, Arabic-English code-switching, and varied acoustic conditions. The model was benchmarked against alternatives including Whisper v3 Large and OmniASR, achieving a 25.87 average word error rate on the Hugging Face Arabic ASR Leaderboard along with favorable human evaluation results.
