I’ve been wanting live, hands-free dictation that still feels private, with no audio shipped off to a server just to get words on the page.
So Codictate now supports NVIDIA’s Parakeet TDT v3 family of models, running entirely on-device through Core ML on Apple’s Neural Engine (the same model line you’ll find as nvidia/parakeet-tdt-0.6b-v3 on Hugging Face, adapted for this path). We specifically use the Hawking Phase Nvidia model.
What that means for you
- Stream mode: talk, pause, keep going; transcription keeps up without you micromanaging record/stop.
- Batch dictation: Parakeet can also power your normal “record and transcribe” flow when you pick it as the model.
- About 25 languages (European-focused set aligned with this checkpoint), plus auto-detect where it applies.
It’s a ~2.5 GB download the first time. Grab Parakeet TDT v3 under Settings when you’re ready.
Whisper isn’t going away. Your favourite Whisper models are still there for translation and the workflows you already rely on; Parakeet is the piece that unlocks stream and gives you another fast, on-device option for raw speech-to-text.
If you haven’t tried Codictate yet, head to codictate.app and see how it feels with your own voice.
Demo
I’m putting together a Parakeet + stream mode screen recording. When it’s live, I’ll drop the YouTube link here like the other posts. Until then, codictate.app has the build.