manage-voice-profiles

Run voices.sh enroll "Name" --record 30 to capture 30 seconds of natural speech as an actor profile. For a character voice, add --actor "Name" so the persona is nested under that actor. You can also enroll from an existing audio file instead of live recording.

An actor is a real person's natural speaking voice, while a persona is an in-character voice that actor performs. Each persona is enrolled under an existing actor, and the DM typically has multiple personas, one per voiced NPC.

Run voices.sh watch to open a live TUI that listens on the default microphone with energy-gated VAD and prints each identified speaker with a confidence score. No audio is recorded or saved, and you stop it with Ctrl+C.

This happens when no voice profiles are enrolled or the channel_priors fallback is not set in tools/audio/config.yaml. Enroll at least one actor profile per participant, or configure mic-based priors as a fallback.

Yes, use the transcribe-session-audio wrapper with --save-profile and --from-mic to build a profile from a session recording. For people who voice multiple characters on one mic, use live --record enrollment or a corrected transcript instead, since mic harvesting blends all voices on that channel.

Re-enrolling merges the new sample with the existing profile by blending 70 percent new embedding with 30 percent old. There is no confirmation prompt, so the merge happens silently.