Cohere Transcribe
CohereCohere's 2B-parameter Conformer encoder-decoder speech recognition model, trained from scratch for transcription in 14 languages. Released on the Hugging Face Hub, where downloading requires accepting a contact-sharing prompt; an Arabic variant followed.
No code repository: the model runs through transformers with remote code.
Openness
3 medium confidence- weights
- open(safetensors behind a click-through contact-sharing gate)
- data
- closed(the card names no training corpus)
- code
- partial(transformers inference code only)
- license
- Apache-2.0(OSI)
Cohere Transcribe's weights are Apache-2.0, an OSI-approved license, behind a gate that asks only for contact details. Cohere does not say what the model was trained on and publishes inference code alone.
- https://huggingface.co/api/models/CohereLabs/cohere-transcribe-03-2026 recorded 2026-09-27
"license:apache-2.0" tag, "gated":"auto", with model.safetensors in the file list.
- https://huggingface.co/CohereLabs/cohere-transcribe-03-2026 recorded 2026-09-27
Model page: "You need to agree to share your contact information to access this model"; the card describes the architecture ("trained from scratch") and inference with transformers, and names no training dataset.
Adoption
3 high confidenceHugging Face downloads of the two released checkpoints, which sit behind a click-through gate.
- https://huggingface.co/api/models?author=CohereLabs&search=transcribe recorded 2026-09-27
cohere-transcribe-03-2026 213,697 and cohere-transcribe-arabic-07-2026 48,692 downloads in the trailing 30 days.
Capability
4 medium confidenceA strong English recognizer a third of a point behind Qwen3-ASR, with narrower language coverage.
- https://huggingface.co/datasets/hf-audio/open-asr-leaderboard-results/resolve/d2c5b384deccdb82834f41aeaffcc618c00efa2f/english_short_latest.csv recorded 2026-09-27
Results CSV row "CohereLabs/cohere-transcribe-03-2026,4.67,906.56,apache-2.0" (14 languages).
Verified 2026-09-27