CohereLabs /

cohere-transcribe-03-2026

A 2B-parameter open model that turns speech in 14 languages into written text.

1.2k likes
101.8k downloads a month
2.1B parameters
Some experience
Speech

Open on Hugging Face (opens in a new tab)

Cohere Transcribe is an automatic speech recognition (ASR) model with 2 billion parameters. It uses a Conformer encoder to read the audio and a lightweight Transformer decoder to write out the text. It was trained from scratch on 14 languages: English, French, German, Italian, Spanish, Portuguese, Greek, Dutch, Polish, Chinese (Mandarin), Japanese, Korean, Vietnamese and Arabic. Audio is resampled to 16kHz automatically, and stereo is averaged down to one channel.

It does one job rather than being a general chat model. The card reports an average word error rate of 5.42 on its English leaderboard table, the lowest of the models listed there, and says it runs up to three times faster (real-time factor) than other dedicated ASR models of similar size. The limits are real, though. You need to tell it which language to expect, because it has no automatic language detection and does inconsistently on audio that mixes languages. It gives no timestamps and no speaker labels. It also tends to transcribe non-speech sounds, so quiet background noise can turn into made-up text unless you filter silence first.

What you need to run it

2.1B parameters. The weights take about 4 GB at full 16-bit precision (fits a 8 GB Mac or a GPU with 8 GB), or about 2 GB as a 4-bit download, which runs on a 8 GB Mac or a GPU with 8 GB. Long inputs and images need more memory on top.

What it does well

Low error rate on English benchmarks
The card's English leaderboard table (as of 03.26.2026) lists an average WER of 5.42 for this model, compared with 7.44 for OpenAI Whisper Large v3 in the same table. Lower is better.
14 languages
You pick one of 14 supported languages with a language code, such as "en" or "ja", when you run it.
Handles long recordings
Audio longer than the feature extractor's limit is split into chunks automatically and the text is stitched back together. The card's example transcribes a 55 minute earnings call.
Punctuation switch
Punctuation is on by default. Setting punctuation=False gives lower-cased text with no punctuation marks.
Runs in many places
The card lists support in transformers, vLLM, mlx-audio for Apple Silicon, a Rust version, in-browser use through transformers.js and WebGPU, and several desktop and mobile apps.

What you could build

Transcribe recorded lectures, interviews or meetings into text.
Build a small tool that transcribes a folder of audio files in one of the 14 languages.
Serve a transcription endpoint with vLLM and send audio files to it over HTTP.
Compare transcripts with and without punctuation for a note-taking or captioning project.
Try it in the browser through the card's demo, with no setup.

Get started

Install the Python packages, then load the processor and model from the Hugging Face hub with transformers and run the Quick Start example from the card, passing the language code of your audio. For serving, the card recommends vLLM.

pip install transformers>=5.4.0 torch huggingface_hub soundfile librosa sentencepiece protobuf accelerate

vllm serve CohereLabs/cohere-transcribe-03-2026 --trust-remote-code

Commands are copied from the model card. Read it on Hugging Face for the full setup (the download asks you to accept the licence first).