Cohere Transcribe
Also: cohere-transcribe-03-2026
Published
Definition
Cohere Transcribe is an open-source speech recognition model from Cohere, released in March 2026. It has 2 billion parameters, takes audio in and gives text out, and supports 14 languages.
The full name on its model card is cohere-transcribe-03-2026. Cohere describes it as a dedicated audio-in, text-out model: it transcribes, and it does nothing else. There is no translation, no chat and no summary.
It is released under the Apache 2.0 license, so the weights can be downloaded and run on your own hardware, and built into other software.
The 14 languages
The model card groups them by region.
- European: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch and Polish.
- Asia-Pacific: Chinese (Mandarin), Japanese, Korean and Vietnamese.
- Middle East and North Africa: Arabic.
How it is built
Cohere describes a Conformer-based encoder-decoder: a large Conformer encoder extracts the acoustic representation, and a lightweight Transformer decoder generates the text. It was trained from scratch, not adapted from an existing model.
How accurate it is
The model card reports an average word error rate of 5.42 percent on the English sets of the Open ASR Leaderboard, as of 26 March 2026. Cohere said at release that this placed it first on that leaderboard for accuracy. Leaderboards move, so read that as a claim about that day.
The figure is for English. The card does not give one number for the other 13 languages.
What it does not do
The model card is direct about three limits.
- No automatic language detection. You tell it which language is being spoken, and the card says it performs inconsistently when a speaker switches language mid-sentence.
- No timestamps and no speaker labels. It returns text only.
- It is eager to transcribe. In Cohere’s words, like most models of its kind it will transcribe even non-speech sounds, so it benefits from a noise gate or voice activity detection in front of it.
Running it on your own machine
Cohere offers the model through its own service, and also publishes the weights. Converted and quantized builds of those weights run on Apple silicon, which is how a 2 billion parameter model fits in a laptop’s memory next to everything else.
Questions people ask.
Is Cohere Transcribe open source?
Yes. Cohere released the model under the Apache 2.0 license, and the weights are published.
Which languages does Cohere Transcribe support?
Fourteen: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Mandarin Chinese, Japanese, Korean, Vietnamese and Arabic.
Does Cohere Transcribe detect the language automatically?
No. Its model card says it does not feature explicit, automatic language detection, so the language has to be chosen.
How big is Cohere Transcribe?
Cohere describes it as a 2 billion parameter model.
Does Cohere Transcribe give timestamps or speaker labels?
No. The model card says it features neither timestamps nor speaker diarization.
Can Cohere Transcribe run offline?
Yes, when you run the published weights on your own machine. Used through Cohere’s service, it needs a connection.
