Superwhisper is partnering with Cohere. Starting today, Cohere Transcribe is available in the Superwhisper models library, running local dictation on your device.
Speed and accuracy are everything in dictation. Stopping to fix a wrong word eats most of the speed you gained by talking instead of typing. Transcribe gets the words right, so you keep moving.
One of the best transcription models available
Cohere Transcribe is one of the top-ranked models on Hugging Face’s Open ASR Leaderboard, with a 5.35% average word error rate that drops to 1.25% on clean audio. It holds up on messy, multi-speaker recordings too, and it beats Whisper Large v3 on both accuracy and throughput while going head to head with the best speech-to-text models out there like ElevenLabs Scribe v2. It also topped the new Hugging Face leaderboard for far-field audio launched this month.
Transcribe is a 2B parameter model, maintaining high word accuracy with strong throughput for its size class. That combination is exactly what dictation needs: fast enough to keep up with natural speech, accurate enough that you don’t reread the text. It’s open source under Apache 2.0 and supports 14 languages, including English, French, Spanish, Mandarin, Japanese, and Arabic.
If you speak one of those 14 languages and you want an on-device model, this is the one to reach for. Parakeet stays the default download only because it covers more languages.
Introducing: Cohere Transcribe – a new state-of-the-art in open source speech recognition.
Unmatched accuracy with Superwhisper vocabulary
Superwhisper users expect vocabulary, snippets, and text expansion to work everywhere in the product. So we paired Cohere Transcribe with our vocabulary wrapper, which teaches it the names, terms, and jargon you actually use.
Whether that’s a teammate’s name, a product, or a piece of technical shorthand, Transcribe now gets it right the first time.
Two Canadian companies, one homegrown stack
We’re big believers in Cohere and the work they’re doing, and we’re proud to bring their model to our users worldwide. The most popular dictation app and one of the best open transcription models, together in one place.
Cohere and Superwhisper are both built in Toronto, so you can now run a fully Canadian, sovereign stack end to end. Your voice, your words, processed on a Canadian model inside a Canadian app. Sovereign tools, built and run at home, are a key way to push Canadian business adoption of AI from 12% today to 60%, the goal set for 2034 by the national AI strategy. And for everyone else, anywhere in the world, it’s simply one of the best models you can dictate with.
Using Cohere Transcribe in Superwhisper
- 1Open the Models library in Superwhisper.
- 2Find Cohere Transcribe and download it (1.3 GB).
- 3Select it as your voice model and start dictating.
Because it runs locally on your device, Cohere Transcribe works fully offline. Your voice never has to leave your machine, and you can dictate anywhere, even with no connection.
Voice-first computing is the future
More people code, write messages, and get work done by talking now. The less time you spend correcting transcripts or waiting for text to land, the more natural it feels.
Cohere Transcribe keeps up with the pace of conversation, handles 14 languages, and never sends your audio anywhere. Give it a try today in Superwhisper.