Transcripts
Transcribe, on a clip's transcript row in the inspector, writes out what you said, using whisper.cpp on your own computer. The transcript is yours to edit.
The speech model
The picker beside the button chooses the model:
| Model | Download | |
|---|---|---|
small.en | 488 MB | The default. More accurate; about 0.7× real time on a typical laptop. |
base.en | 148 MB | Faster, less accurate. |
The first time you transcribe with a model, pundit downloads it into
~/.cache/pundit/models/ and keeps it — the button says so, as Download
488 MB and transcribe, before you press it. After that, nothing leaves your
machine. The model you choose is remembered for this computer, not for the
project.
While it runs
Transcribing is always something you ask for; recording never starts it. Clips you ask for queue behind each other, and the inspector shows how long the current one has been going.
Recording always wins. Starting a recording stops a transcription in progress and puts that clip back at the front of the queue, so the take gets the whole machine.
Using your own model file
If $PUNDIT_WHISPER_MODEL names a model file, pundit uses that instead of the
picker, which greys out and shows the file.