yaps-meeting-transcription

Make a meeting transcript with speaker labels in Yaps. Correct it, export it, or ask questions. New users: install Yaps and sign in.

Read the runtime guide before the first operation. It defines <yaps>, account readiness, local permissions, and file handling.

Required Yaps version: 2.3.124 or newer. Check installed command help for later capabilities.

Feature readiness: Meeting/Sherpa; optional MOSS on supported Apple Silicon systems. Inspect features and group help before selecting an engine. Recaps also need the local chat model.

Workflow

Write a private temporary JSON request with an absolute input, optional title, engine (default auto), and optional integer speakers from 1 to 20. Run node <skill-root>/runtime/run.mjs meeting-file <request.json>. Remove only the request file. The helper extracts a temporary WAV for recognized video types through Yaps, then creates the durable meeting project; Yaps retains its own project audio.

Use auto by default. Use sherpa for cross-platform execution or a supplied speaker count. Use moss only when installed and supported on Apple Silicon; it detects speakers itself and cannot take a speaker-count hint. Treat the returned engine and reason as authoritative. A project with no transcript segments is not a successful meeting.

Inspect <yaps> meeting show <id> and meeting speakers <id> before editing. Use stable segment IDs with meeting correct <id> --segment <id> --text-file <file> or meeting assign <id> --segment <id> --speaker <number>. Roster numbers are 1-based. Use rename-speaker for a name correction; use merge-speakers <id> --speakers 3,4 --into 2 only after the user identifies the same participant. Never infer identity from voice alone. Export using meeting export <id> --output <new.md>.

For requested recaps, use meeting summarize <id> and then meeting chapters <id>. Reuse a recap unless regeneration is requested or its transcript/template changed. For a grounded question, use meeting ask <question> --meeting <id> --scope meeting; --scope all requires an explicit cross-meeting request. Preserve returned source citations. Never edit project JSON directly.

Boundaries

• No joining calls or starting live recording on the user's behalf.
• Speaker labels are diarization labels until the user supplies identities.
• Recaps are generated interpretations, not verbatim transcript. Preserve project audio and existing exports.