AI company Cohere has released Transcribe, an open-weight automatic speech recognition model that the company says achieves a 5.42% word error rate, positioning it as a production-ready alternative to existing closed APIs for enterprise voice applications.

The 2-billion parameter model, available under Apache-2.0 license, supports 14 languages including English, French, German, Italian, Spanish, Greek, Dutch, Polish, Portuguese, Chinese, Japanese, Korean, Vietnamese and Arabic. The model can be accessed through Cohere's API or downloaded from their Model Vault as cohere-transcribe-03-2026.

According to the Hugging Face ASR leaderboard, Transcribe outperforms several established models including OpenAI's Whisper Large v3 at 7.44% word error rate and ElevenLabs Scribe v2 at 5.83%. On specialized datasets, the model scored 8.15% on the AMI meeting dataset and 5.87% on the Voxpopuli accent dataset.

Cohere designed the model specifically for enterprise deployment, allowing organizations to run transcription workloads on their own infrastructure rather than routing audio data through external APIs. The company says this addresses data residency concerns while maintaining production-grade accuracy and throughput performance.

The release targets enterprises building voice-enabled workflows, transcription pipelines, and audio search systems. Unlike research-focused models, Transcribe is commercially licensed from release and optimized for local GPU deployment, according to Cohere.

The model's availability reflects growing demand for speech recognition capabilities that can be deployed on-premises while maintaining competitive accuracy rates compared to cloud-based services.