Parakeet TDT v3
Use Parakeet TDT v3 with the NeMo-Speech.cpp backend for microphone recordings, pause-aware checkpoints, and audio-file transcription. The model detects the spoken language automatically and produces punctuated text across 25 European languages.
Set up the model
Section titled “Set up the model”For local use on Windows or macOS, follow managed NeMo setup, but choose Parakeet TDT v3 from the catalog instead of Nemotron. Download and start it, then select the built-in NeMo Connection for Voice, Audio file, or both. Freehand supplies the model profile; Parakeet uses completed transcription only.
For a manually configured service, install NeMo-Speech.cpp v0.1.0
on your inference machine. Its parakeet-tdt model entry selects
nvidia/parakeet-tdt-0.6b-v3.
Follow the runtime’s model and server instructions
to load it and expose the HTTP API.
Connect that manual service in Freehand:
- Add a NeMo-Speech.cpp connection.
- Open the Settings cog in Voice transcription or Audio file and select that connection in Transcription options. Configure each workflow independently if you use both.
- Refresh models and select the loaded Parakeet model.
- Choose Parakeet TDT v3 as the model profile, then save.
What appears in Freehand
Section titled “What appears in Freehand”Microphone transcription completes when you stop recording or at a pause-aware checkpoint. You can use cleanup and focus-safe insertion as usual. Select the model separately for audio-file transcription, then copy the result when ready.
| Option | What to expect |
|---|---|
| Spoken language | Automatic detection; recognizes 25 supported European languages without a dropdown |
| Forced language | Unavailable; see NVIDIA’s model card for the language list |
| Context, vocabulary, temperature | Unavailable; shared vocabulary remains saved for other compatible selections |
| Realtime microphone | Unavailable; use completed recordings or pause-aware checkpoints |
NeMo transcription controls let you keep or strip punctuation and request optional server-side number/date normalization or profanity filtering. Those optional controls depend on the server’s configured assets; changing them does not download anything.