Cohere Releases Open-Weight ASR Model 'Transcribe' at 5.4% Word Error Rate
Cohere has released Transcribe, an open-weight automatic speech recognition (ASR) model under Apache 2.0 license that achieves a 5.42% word error rate — accurate enough to replace closed speech APIs in production pipelines.
Model Details
| Specification | Value |
|---|---|
| Parameters | 2 Billion |
| License | Apache 2.0 |
| WER | 5.42% |
| Languages | 14 |
| API Name | cohere-transcribe-03-2026 |
Supported Languages
English, French, German, Italian, Spanish, Greek, Dutch, Polish, Portuguese, Chinese, Japanese, Korean, Vietnamese, and Arabic.
Why It Matters
Until now, enterprise transcription has been a trade-off:
- Closed APIs (Whisper API, Google): Accurate but lock in data, ongoing costs
- Open models (Whisper OSS): Control but lower accuracy, not commercially licensed
Transcribe breaks this trade-off by offering:
- Production-ready accuracy: 5.4% WER competitive with best closed APIs
- Self-hosted deployment: Run on your own GPU infrastructure
- Commercial license: Apache 2.0 with no restrictions
- Manageable footprint: Optimized for local GPU inference
Technical Achievement
Cohere says Transcribe "extends the Pareto frontier" — delivering state-of-the-art accuracy while maintaining best-in-class throughput within the 1B+ parameter cohort.
Impact
This could fundamentally change enterprise voice workflows by eliminating the dependency on cloud-based speech APIs, reducing costs, and addressing data residency concerns — particularly relevant for healthcare, legal, and financial services.