Record to Structured Notes

Problem it solves

Phone voice memos and meeting recordings often sit as raw files without titles, summaries, action items, or speaker separation. Plain transcription is fast but still hard to turn into notes in Obsidian, Notion, or Logseq. record2note targets this gap by turning audio into structured, archivable Markdown notes.

How it works

The skill splits the pipeline into transcription and note generation.

• Audio preprocessing: converts input to 16kHz mono WAV with ffmpeg, with optional Silero-VAD silence skipping and afftdn denoising.
• Whisper transcription: runs local Whisper with a fallback chain from GPU to limited threads to CPU, and calculates timeouts from audio length and model size to prevent pipeline stalls.
• Speaker separation: uses pyannote-audio when diarization is enabled, adding speaker labels such as A/B back into the transcript.
• Structured notes: stores transcripts as pending JSON files, then uses the Agent to generate titles, summaries, key points, TODOs, and full transcripts.
• Obsidian mode: emits Callouts, [[wikilinks]], and maintains an index page for backlinks.

Boundaries

This is a local automation skill with meaningful setup costs. First run downloads Whisper models and optional dependencies; large-v3 or speaker diarization can require several GB of space. Automatic monitoring depends on macOS fswatch, Windows scheduled tasks, and file delivery via iCloud, Syncthing, or manual copying. If Agent CLI auto-trigger is used, launchd permissions or missing bash on Windows may cause authentication or execution issues; you can disable auto-trigger and process pending files manually.