Cohere goes open-source with 2B speech model Transcribe

Cohere Transcribe is a lightweight open-source speech-to-text model designed to handle voice note-taking, transcription, and speech analytics on your own hardware.

Packing just 2 billion parameters, it is built to run on consumer-grade GPUs, so you can self-host voice workloads without shipping private audio off to third-party clouds.

What it does and system specs

Transcribe is trained from scratch for everyday production use across 14 languages, including English, Chinese, Japanese, Korean, French, German, Spanish, and Arabic.

Speed is its biggest flex. The model can churn through 525 minutes of audio in a single minute, chewing through multi-hour recordings in just a few seconds.

Because the footprint sits at a modest 2B parameters, you do not need enterprise cluster hardware. A standard consumer GPU is more than enough to handle local inference for desktop dictation or internal meeting summarization.

Architecture and benchmark results

Under the hood, Transcribe relies on a Conformer-based encoder-decoder design. A large Conformer encoder extracts acoustic features from log-Mel spectrograms, followed by a lightweight Transformer decoder generating tokens.

On the Hugging Face Open ASR leaderboard, Transcribe grabbed the top spot with an average word error rate (WER) of 5.42. That puts it ahead of rivals like ElevenLabs Scribe v2, IBM Granite 4.0 1B, Qwen3-ASR-1.7B, and Whisper Large v3, holding up especially well in noisy multi-speaker meeting rooms.

In blind human evaluations measuring transcription accuracy and coherence, Transcribe scored a 61% win rate over alternatives. The caveat: it still trailed behind competitors when transcribing Portuguese, German, and Spanish.

How to get it

Transcribe is released under a permissive Apache 2.0 license, meaning the weights are free to download from Hugging Face right now for commercial use.

If you prefer not to manage infrastructure, Cohere is also offering free API access to the model, listing it on their Model Vault platform, and integrating it into North, their enterprise agent workspace.

Why it matters

If you are building meeting assistants, transcription wrappers, or local voice search, Transcribe is an immediate must-try. Beating Whisper Large v3 on accuracy while staying small enough to run on a single consumer card is no small feat, and the Apache 2.0 badge gives builders complete deployment freedom.

That said, if your user base talks predominantly in Spanish, Portuguese, or German, do not swap out your pipeline just yet. Benchmark it against your own production audio first to make sure those specific language dips do not hurt your downstream accuracy.