What is Cohere Transcribe (03-2026)?
Cohere Transcribe (cohere-transcribe-03-2026) is a 2-billion-parameter speech recognition model released by Cohere on 26 March 2026, its first audio model. It pairs a large Conformer encoder with a lightweight Transformer decoder, was trained from scratch, and takes audio in and returns text, with nothing else in between.
It covers 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese, Japanese, Korean, Vietnamese and Arabic, and Cohere has since posted a separate Arabic checkpoint. The weights are on Hugging Face under Apache 2.0 behind an access form, and run in Transformers 5.4.0 or later and in vLLM 0.19.0, which serves an OpenAI-style transcription endpoint. Cohere's own API offers it free for testing, and Model Vault hosts it on dedicated instances.
Is Cohere Transcribe (03-2026) worth it?
Cohere Transcribe earns its place on English accuracy at a size you can run yourself. On Hugging Face's Open ASR Leaderboard at launch it averaged a 5.42% word error rate across eight English test sets, first on the board, ahead of Zoom Scribe v1 at 5.47%, IBM's Granite 4.0 1B at 5.52%, Qwen3-ASR-1.7B at 5.76% and ElevenLabs Scribe v2 at 5.83%; Whisper large-v3 scored 7.44%. It did that at an RTFx of about 525, meaning 525 minutes of audio a minute, in Cohere's measurement.
Independent testing is less flattering. Artificial Analysis puts Cohere's hosted version at a 4.6% AA-WER, joint 43rd of 62 entries and level with Whisper large-v3-turbo on Groq, while Mistral's Voxtral Mini Transcribe 3 sits at 2.2%. Leaderboard sets reward clean read and broadcast speech; Artificial Analysis mixes in harder audio, and the gap suggests the model is strong on the former rather than a general leader. In Cohere's own human preference tests it won 61% of comparisons on average but lost in Portuguese, German and Spanish.
The model does one thing. The docs say plainly that it has no timestamps, no speaker labels and no automatic language detection, so you must name the language for each file, and it handles code-switched audio inconsistently. It is also eager to transcribe non-speech sounds, so Cohere recommends a voice-activity detector or noise gate in front of it to stop it inventing words in silence.
The catch: the hosted side is unfinished as a product. The API is free but trial keys allow 5 transcription requests a minute, production limits need a sales call, files are capped at 25MB and the docs say nothing about how long audio is kept. Model Vault, the dedicated option, is priced per instance-hour with no public figure. For production you are realistically self-hosting, which the Apache 2.0 licence makes easy.
Our take: Cohere Transcribe is a sound default engine for English-heavy batch work on your own GPUs, as long as you add a speech detector, pick the language yourself and bolt on alignment and diarization from elsewhere.
Where it earns its place
An ML engineer at an investor-relations platform with thousands of earnings calls a quarter
Turn English earnings calls into searchable text without per-minute API fees
Serve the model with vLLM on owned GPUs and send files to its OpenAI-style transcription endpoint; Cohere's own example transcribes a 55-minute earnings call. Its Earnings-22 error rate is 10.84%, so ticker symbols and figures still need a pass with a custom vocabulary check.
A product team at a Japanese note-taking app
Add on-device or private-cloud dictation in Japanese and English
Run the GGUF or mlx-audio ports and set the language per session, since there is no auto-detection. Japanese is one of its 14 languages, but the card warns of inconsistent results on code-switched audio, so sentences that mix English into Japanese are where it wobbles.
A compliance lead at a European bank
Transcribe recorded calls without sending audio to a US speech API
Self-host under Apache 2.0 so audio never leaves the bank's infrastructure, and pair it with a separate diarization model such as pyannote, because Cohere Transcribe does not say who spoke.
Before you commit
- Setup
- Accept the access form on Hugging Face, install transformers 5.4.0 or later and load CohereAsrForConditionalGeneration with a language code; for serving, vllm serve with vllm 0.19.0 and the audio extra.
- Integrations
- Transformers, vLLM (OpenAI-compatible /v1/audio/transcriptions), mlx-audio, transformers.js with WebGPU, community GGUF builds; Cohere API, Model Vault and fal for hosted use.
- Data and privacy
- Self-hosted, audio stays on your hardware under Apache 2.0. Cohere's Transcribe docs do not state retention for API audio; Model Vault is described as private cloud inference.
- Cost at scale
- The weights are free and Cohere says it runs on consumer GPUs. The free API is limited to 5 requests a minute on trial keys; production API limits and Model Vault instance-hour prices are not published.
- Limits
- 14 languages, one per request; no timestamps, diarization or language detection; 25MB per file on the API; long audio is chunked and stitched automatically; hallucinates on silence without a voice-activity detector.
What’s good
- 5.42% average WER, first on the Open ASR Leaderboard at launch
- Apache 2.0, about 2B parameters, runs on a single consumer GPU
- About 525 times real time in Cohere's tests; vLLM serves an OpenAI-style endpoint
- 14 languages including Arabic, Japanese and Vietnamese
What’s not
- No timestamps, diarization or language detection
- 4.6% AA-WER on Artificial Analysis, 43rd of 62
- Transcribes non-speech noise unless gated by a voice detector
- Trial API capped at 5 requests a minute; production pricing by quote
- Lost human preference tests in Portuguese, German and Spanish
Best for
Teams transcribing large volumes of single-language, mostly English audio on their own GPUs who want a clean licence.
Cohere Transcribe (03-2026) alternatives
Cohere Transcribe (03-2026): common questions
- Is Cohere Transcribe (03-2026) worth it?
- Cohere Transcribe is an accurate, permissively licensed English transcriber that leaves timestamps, speaker labels and language detection for you to build. We rate it 4.0 out of 5.
- How much does Cohere Transcribe (03-2026) cost?
- Free under Apache 2.0 (gated download). Cohere's API is free for testing at 5 requests a minute on trial keys; production limits and Model Vault hourly prices are by quote.
- What are the downsides of Cohere Transcribe (03-2026)?
- No timestamps, diarization or language detection; 4.6% AA-WER on Artificial Analysis, 43rd of 62; transcribes non-speech noise unless gated by a voice detector; trial API capped at 5 requests a minute; production pricing by quote; lost human preference tests in Portuguese, German and Spanish.
- Who is Cohere Transcribe (03-2026) best for?
- Teams transcribing large volumes of single-language, mostly English audio on their own GPUs who want a clean licence.
- How much does Cohere Transcribe (03-2026) cost for a team or in production?
- The weights are free and Cohere says it runs on consumer GPUs. The free API is limited to 5 requests a minute on trial keys; production API limits and Model Vault instance-hour prices are not published.
- Is Cohere Transcribe (03-2026) safe for company data?
- Self-hosted, audio stays on your hardware under Apache 2.0. Cohere's Transcribe docs do not state retention for API audio; Model Vault is described as private cloud inference.
- What are the best Cohere Transcribe (03-2026) alternatives?
- The closest alternatives we have reviewed are Qwen3-ASR (1.7B and 0.6B), Granite Speech 5.0 2B, Parakeet TDT 0.6B v3, Whisper large-v3-turbo and Voxtral 3 (Transcribe and TTS).
Answers checked Nov 7, 2026.
Make Cohere Transcribe (03-2026)? Show this rating on your site
