Skip to content

Call Note 0.2.12 — More stable speaker labels

Choose a tag to compare

@slyapustin slyapustin released this 14 Oct 19:01

More stable speaker labels

  • Short phrases and fillers no longer create a new speaker identity. Uncertain matches display Speaker uncertain instead of inventing another participant.
  • Clear matches gradually update the voice reference during a call.
  • Overlapping speaker regions are transcribed once; only the conflicting overlap loses its speaker label.
  • Failed speaker turns cannot leave an earlier labeled row beside a duplicate fallback phrase. Genuine repeated phrases at different times are preserved.

These changes apply to new transcription. Existing saved labels are unchanged because Call Note does not retain the original audio.

Validation: 71 Python tests and 143 desktop tests. The bundled solo sample stayed at one identity; a three-person sample split into three-second chunks improved from four identities to three. Full streaming preserved its expected speaker sequence with no warnings. This does not establish an accuracy rate for real calls, and the reported one-person call was not available as audio. See docs/speaker-validation.md for scope.

Install

Download the .dmg, quit Call Note, and drag it to Applications, replacing the old copy. Saved calls and preferences remain in your user Library. Apple Silicon, macOS 14.2 or later.

The app and installer are Developer ID signed and Apple-notarized. Automatic updates check after launch and every four hours, download in the background, and install when you quit. The ZIP and latest-mac.yml are updater assets. Live version-to-version update installation was not exercised.

Includes 0.2.11’s Mac capture change and optional automatic call detection. Live call capture and detection validation remain limited; report persistent problems with Settings → Support → Export diagnostics.