Preparing secure vault
Unlock the local vault to load patient data. Loading runs on this device and can take a moment on a first visit. If this message does not change, reload the page.
Scribe Pro
Parakeet Laptop onlyOn-device clinical scribe: NVIDIA Parakeet TDT 0.6B runs fully in this browser. No audio or transcript ever leaves the device.
1 · Model setup
Total download: about 689 MB. The speech encoder is about 652 MB of that total. Download once on Wi-Fi, then Scribe Pro works offline.
One download, then it works offline. The models come from Hugging Face (~690 MB), are stored in this browser's IndexedDB, and run fully on this device afterwards. Best done on Wi-Fi — you only do this once.
2 · Record
Recording unlocks once the models finish downloading above.
3 · Transcript
No transcript yet. Press Record above, then Stop — the transcript appears here, transcribed on this device.
Experimental — review and verify the transcript before any clinical use. Timestamps are approximate.
Voice enrollment (optional)
Record a 12-second sample of your voice. Later transcripts label matching speech as “You” and everything else as “Patient”. Only the voiceprint (a small numeric embedding) is saved on this device — never audio.
Not enrolled.
Attribution
- Parakeet TDT 0.6B v2 speech model — © NVIDIA Corporation, licensed CC-BY-4.0 (istupakov/parakeet-tdt-0.6b-v2-onnx).
- Silero voice activity detection — MIT license, attribution to Silero (export via christopherthompson81/sortformer_parakeet_onnx).
- WeSpeaker voxceleb ResNet34-LM speaker embedding — licensed CC-BY-4.0 (onnx-community/wespeaker-voxceleb-resnet34-LM).
- ONNX Runtime Web 1.22.0 — MIT (microsoft/onnxruntime), vendored locally.
- Adapted prior art (logic ported, MIT): the TDT greedy decoder follows onnx-asr’s NemoConformerTdt; tensor-reuse and periodic-yield patterns adapted from xiuxiu’s parakeet worker.
Full third-party notices: see THIRD-PARTY-NOTICES.md in the inference core.