v0.5.5-beta1 — Speaker tagging (first Premium feature)
The first feature gated behind the Premium tier. Voice tagging detects each distinct speaker in your meeting and lets you label them by name.
What’s new (Premium)
- Speaker panel: right-side panel in every meeting detail shows auto-detected speakers as “Speaker 1, Speaker 2, …”.
- Tag once, propagate everywhere: rename a row to “Marisa” and every transcript segment in that cluster instantly relabels.
- Voice sample playback: click ▶ on any row to hear a 3 second sample of that voice.
- Split & merge: fix model mistakes by splitting a segment off into its own speaker, or merging two rows that are actually one person.
- Mode setting: in Settings → License, choose “After recording stops” (default, lighter) or “Live during recording” (faster results, slightly hotter). Live mode automatically defers to After-stop on Low Power Mode.
- 100% on-device: voice fingerprints are computed locally via Apple’s Neural Engine. They never leave your Mac.
Existing meetings
Voice tagging is available for meetings recorded with v0.5.5 onwards. Recordings made before this build don’t have voice fingerprints, so the speaker panel will show a note. No retroactive backfill.
Free tier
Free users see the speaker panel as a billboard explaining Premium. No voice processing runs on Free; nothing changes for Free meetings.
Under the hood
Voice embeddings are produced by ECAPA-TDNN (Apache 2.0, SpeechBrain), converted to Core ML and bundled with the app. Each meeting’s system-track segments are run through the model on the Neural Engine (CPU/GPU fallback on Intel), then clustered via agglomerative cosine similarity with threshold 0.75. App size grows by ~40 MB to ship the model offline.