What problem does it solve? Recorded tabletop RPG sessions contain hours of multi-speaker audio that must be converted into labeled, timestamped text before session notes can be ingested into the campaign wiki. Manual transcription and speaker attribution is slow and error-prone. ## Core Features & Use Cases - Speaker-Labeled Transcription: Runs Whisper large-v3 over isolated per-mic m4a tracks and writes per-part CSV chat logs (ID, Start, End, Speaker, Text) to .raw/sessions/session-NN/transcripts/raw/. - Three-Layer Speaker Identification: Combines saved voice profiles (resemblyzer d-vectors plus prosody), per-mic channel priors, and stable fallback labels, with optional pyannote diarization when HF_TOKEN is set. - Voice Profile Enrollment: Harvest actor and character-voice (persona) profiles from a session via --save-profile, --from-mic, and --actor flags so future sessions label speakers automatically. - Use Case: After recording session 7, run the transcribe script to produce per-part CSVs, assemble them into one transcript, then hand off to the session-ingest skill for speaker resolution and promotion into campaign canon. ## Quick Start Ask the agent to transcribe session 7 by saying "transcribe the session 7 recording into a speaker-labeled transcript".
