One pass of the watcher, replayed.
A recorded run over the five sample recordings. The clock stamps are the real pace of the run on a desktop with a graphics card; the replay is about ten times faster. The wording is what the scripts print.
Recorded run. Sample data.The model loads once
Transcription happens inside the watcher process, so the speech model is loaded a single time for the whole batch and the files are processed one after another. There is one graphics card; running several transcriptions at once is slower, not faster. After ten idle minutes the model is released and the memory goes back to the machine.
What stays on the machine
Audio extraction, the speech model and the speaker labelling all run locally. The audio file never leaves the computer. The summary step sends the text transcript to an AI assistant through Dana's existing subscription; that is the one step that uses the network, and it only ever sees text.
Honest about uncertainty
When voice ID and the transcript disagree about whether a recording is a client call or an internal session, the call is still filed on the best guess, but it is flagged, printed at the end of the run and listed at the top of the ledger for a human to confirm.