Enterprises seeking reliable voice transcription have long faced a difficult choice: use closed APIs that pose data privacy concerns, or opt for open models that sacrifice accuracy for flexibility. Cohere has entered this space with Transcribe, an open-weight automatic speech recognition model designed to eliminate this trade-off. With a word error rate of just 5.42%, Transcribe delivers accuracy that rivals or surpasses established leaders whilst offering the control and deployability that enterprises require for production environments.
Transcribe is a 2-billion parameter model released under the Apache-2.0 licence, making it commercially viable from launch. Unlike OpenAI's Whisper, which was initially positioned as a research tool, Transcribe is explicitly built for enterprise production pipelines. The model supports 14 languages, including English, French, German, Italian, Spanish, Greek, Dutch, Polish, Portuguese, Chinese, Japanese, Korean, Vietnamese, and Arabic. Cohere has optimised the model to run efficiently on local GPU infrastructure, allowing organisations to host transcription workloads in-house rather than routing sensitive audio data through external APIs.
Benchmark results place Transcribe at the top of the Hugging Face ASR leaderboard, outperforming Whisper Large v3 (7.44% WER), ElevenLabs Scribe v2 (5.83% WER), and Qwen3-ASR-1.7B (5.76% WER). On the AMI dataset, which evaluates meeting comprehension and dialogue analysis, Transcribe scored 8.15%. For the Voxpopuli dataset testing accent recognition, it achieved 5.87%, narrowly trailing only Zoom Scribe. Early adopters have highlighted the model's combination of accuracy and self-hosted deployment as particularly valuable for teams building retrieval-augmented generation pipelines or agent workflows that incorporate audio inputs.
The model is accessible via Cohere's API or through Cohere's Model Vault as cohere-transcribe-03-2026. By offering production-grade accuracy alongside local deployment capabilities, Transcribe addresses longstanding pain points for engineering teams that require both performance and data sovereignty. This positions the model as a viable alternative to closed transcription APIs for organisations building voice-enabled workflows at scale.
Fuente Original: https://venturebeat.com/orchestration/coheres-open-weight-asr-model-hits-5-4-word-error-rate-low-enough-to-replace
Artículos relacionados de LaRebelión:
- Dash0 Unicorn Status AI Fixes Production Problems
- AI Agents Need Processes Are Your Operations Ready
- Alibabas Tiny AI Model Outshines OpenAIs
- Panasonic TV Production From Plasma King to Outsourcing
- Netflix Compra Ready Player Me Avatares 3D para el Futuro
Artículo generado mediante LaRebelionBOT
No hay comentarios:
Publicar un comentario