Cohere의 2B 아랍어 모델, 코드 스위칭에서 Whisper 능가
cohere가 Apache 2.0 라이선스로 공개한 2B 파라미터 모델이 아랍어 벤치마크에서 Whisper large v3를 능가한다고 주장하며, 방언 다양성과 코드 스위칭 문제를 다룹니다.
cohere released cohere transcribe arabic, a 2-billion-parameter open-source asr model optimized for dialect variety, bilingual code-switching, and specialized vocabulary. licensed under apache 2.0, the model is accessible via hugging face and the cohere api, claiming to outperform whisper large v3 and standard cohere transcribe on accuracy. human ratings indicate significant gains in dialect faithfulness and handling mixed arabic-english inputs. this signals a push towards specialized, high-performance open-source models for low-resource or complex linguistic scenarios.