Keeping presentation coaching on device
A practical look at privacy, latency, and synchronizing local speech analysis with a live Keynote deck.
Privacy changes the architecture
Presentation coaching handles material people may not want to upload: unreleased work, research, client information, or a nervous first rehearsal. Waverd treats local processing as a product requirement rather than a technical preference.
WhisperKit and Core ML keep transcription on the Mac. That removes a cloud round trip and makes the privacy promise easier to explain: the recording and analysis stay close to the person presenting.
A transcript is only half the context
Feedback becomes more useful when it can answer which slide was visible when a phrase, pause, or filler occurred. Waverd polls Keynote through AppleScript and aligns slide changes with the speech timeline.
The resulting model is temporal rather than document-based. Speech events, slide transitions, and recording progress share a common clock, which makes per-slide feedback possible after the rehearsal.
Useful intelligence stays quiet
The goal is not to show every signal the models can produce. It is to surface a small set of observations that someone can act on before the next rehearsal. Local intelligence earns trust when the interface remains calm, specific, and easy to dismiss.