Diarization engine
Our own diarization stack — diart (live) + TD, our trained GPU post-pass model.
Diarization answers who spoke when, anonymously. Our own stack has two halves:
- diart — the live, streaming diarizer that drives the on-screen preview during a recording (CPU).
- TD — our own trained model (WavLM-based) that runs the authoritative post-pass on a GPU after the meeting, re-diarizing for accuracy.
We built TD so the authoritative pass runs on a model we own end to end, and tuned the live diart path to run in real time (RTF < 1) on modest CPU. While we scale these across TEE GPUs, a managed engine (Deepgram) currently serves as a temporary fallback so every meeting still gets diarized; engine selection is flag-gated (see Deployment).
Diarization quality is judged by the merged [speaker] text transcript, not by any single metric in
isolation. Speaker labels are machine-generated and disclaimed as such; a name is attached to a voice
only through VFTEE, and only with consent.