Cohere Transcribe
Cohere's first audio model — a 2B open-weights transcriber that took #1 on the Open ASR Leaderboard.
About
Cohere Transcribe (cohere-transcribe-03-2026) is Cohere's first audio product, released March 26, 2026 under Apache 2.0. It is a 2B-parameter Conformer encoder with a lightweight Transformer decoder — over 90% of the parameters sit in the encoder — trained from scratch on 500K hours of audio across 14 languages. It immediately took the #1 spot on the Hugging Face Open ASR Leaderboard with 5.42% average word error rate in English, beating Whisper Large v3 (7.44%), ElevenLabs Scribe v2 (5.83%), and NVIDIA's Canary Qwen 2.5B (5.63%).
Trade-offs are honest: the language must be specified per request (no auto-detection), there are no timestamps or speaker diarization, and it does not stream partial results. The weights are free for commercial use (Hugging Face asks for a contact-sharing form before download), and Cohere's API offers a free trial tier at 5 requests/minute and 1,000 calls/month.
Specs
Context window
Pricing
Free quota
API access
Weights
Released
Benchmarks
Independent: Hugging Face Open ASR Leaderboard #1 — 5.42% avg WER (English), vs Whisper Large v3 7.44%.
Provider: multilingual results comparable to or better than all open models across 13 more languages.
Provider-reported unless a source says otherwise. Benchmarks are marketing-adjacent by nature — treat them as directional.
Use cases
Trust score — 8/10
Repository and release blog verified live on Hugging Face (HTTP 200); Apache 2.0 license; the #1 leaderboard position is published by Hugging Face, not the vendor — the strongest independent evidence in this batch. Docked only for: (1) no timestamps or diarization limits subtitle/dictation use; (2) HF download requires a contact-sharing form.