Cohere Transcribe 03-2026
Description
Cohere Transcribe 03-2026 is an open speech recognition model with about 2 billion parameters, built only for turning audio into text. Per the model card, it supports 14 languages (English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese, Japanese, Korean, Vietnamese and Arabic) and long recordings: the feature extractor splits long audio into chunks and the processor reassembles them. Punctuation is on by default and can be switched off. The model does not return timestamps or separate speakers. The card warns that silence can cause hallucinated text and recommends a noise gate or voice activity detection. The card does not state a VRAM requirement. On the Open ASR Leaderboard (English short-form speech, nine test sets, checked 2026-10-20) it has an average word error rate of 6.26%. Released 2026-03-24. The weights are gated: you need to log in to Hugging Face and agree to share your contact information to download them. The license is Apache-2.0, which allows commercial use. The authors offer an official demo Space on Hugging Face. Date of check: 2026-10-20.
Pros and cons
Pros
- Apache-2.0 license, commercial use allowed
- 14 languages and long recordings handled automatically
- Low average error rate on the Open ASR Leaderboard
- Punctuation can be switched on or off
- Official Space for trying it without installation
Cons
- Weights are gated: login and agreement to share contact information required
- No timestamps and no speaker separation
- Can produce text on silence; voice activity detection recommended
- The card does not state a VRAM requirement
Key facts
- License
- Apache-2.0
- Commercial use
- Allowed
- Developer
- Cohere Labs
- Interface languages
- English
- Checked on
- Oct 20, 2026