This application automates private, server‑side transcription and AI post‑processing of audio and video recordings for a distributed team using Nextcloud StorageShare.

It is designed for:

• Zoom local recordings saved into Nextcloud
• Local audio recorders that sync into Nextcloud
• Fully private processing (no SaaS transcription services)
• Clear separation between transcription and AI enrichment

The system runs as two independent workers:

1. Transcription worker (dev) — converts audio/video → .txt
2. AI worker (ai) — converts selected transcripts → structured Markdown documents

A third script exists for test seeding only.

• Each team member has a private *-Transcripts folder in Nextcloud
• An automation user is granted access to each space
• Files are pulled via WebDAV, processed locally, then written back
• No polling state is stored remotely; idempotency is enforced by file moves

<Name>-Transcripts/
├── New-Recordings/
│   ├── Audio/                  (used for default save location for voice recorders)
│   ├── <Zoom meeting folders>/
│   └── *.m4a / *.mp3 / *.mp4
├── Transcripts/                (generated .txt files)
├── Completed/
│   ├── Audio/
│   └── Video/
├── AI/
│   └── *.txt                   (user‑selected AI inputs)
├── AI/Output/                  (generated .md files)
└── Hold/                       (quarantined failures)

Purpose

• Detect new audio/video files
• Transcribe with whisper.cpp
• Normalize known terminology
• Upload .txt transcripts
• Move processed media out of inboxes

Behavior

• Supports loose files and Zoom folders
• Multi‑audio Zoom folders produce numbered transcripts
• Zero‑byte or invalid files are moved to Hold
• Re‑runs are safe (existing transcripts are skipped)

Audio formats

• .m4a, .mp3, .mp4
• Audio is converted to mono 16kHz WAV before transcription

Purpose

• Process only transcripts explicitly placed in AI/
• Generate structured Markdown documents via OpenAI
• Move original .txt back to Transcripts/

Key rules

• If AI/ does not exist → skip
• If AI/ exists but is empty → no‑op
• AI/Output/ is created only when needed
• Output filenames are deduplicated (-2, -3, etc.)

Output

• Markdown documents (meeting summaries, notes, etc.)
• Schema‑validated JSON → Markdown
• Company name normalization enforced

The system uses whisper.cpp with a local model file.

Required

MODEL_PATH=/absolute/path/to/ggml-base.en.bin

Example deployment mount:

-v /var/lib/whisper-models:/models:ro
MODEL_PATH=/models/ggml-base.en.bin

The container will fail fast if the model is missing.

NC_BASE=https://nextcloud.example.com
NC_USER=automation-user
NC_PASS=app-password
NC_ROOT=/Transcripts

NC_TEMPLATE_INBOX=New-Recordings
NC_TEMPLATE_INBOX_AUDIO=New-Recordings/Audio
NC_TEMPLATE_TRANSCRIPTS=Transcripts
NC_TEMPLATE_COMPLETED_AUDIO=Completed/Audio
NC_TEMPLATE_COMPLETED_VIDEO=Completed/Video
NC_TEMPLATE_HOLD=Hold

OPENAI_API_KEY=sk-...
OPENAI_MODEL=gpt-5-mini

src/seed-testdata.mjs

• Create deterministic, realistic test data in Nextcloud
• Validate transcription and AI behavior end‑to‑end
• Exercise failure paths safely

• DESTRUCTIVE: archives any existing test root
• Must use:

NC_ROOT=/Transcripts/_TEST_LATEST

• Never run against production paths

• Archives previous test runs to _TEST_RUNS/<timestamp>
• Creates multiple user spaces
• Seeds:
  • Loose audio
  • Zoom folders (single/multi/no audio)
  • Bad audio
  • Zero‑byte files
  • AI input transcripts

• AI/Output is never created by the seed script
• This avoids duplicate folder artifacts in Nextcloud

npm run seed:test
npm run dev:test
npm run ai:test

• Docker image contains all runtime dependencies except the Whisper model
• Systemd timers trigger:
  • Transcription worker (e.g. every 15 minutes)
  • AI worker (independent cadence)

Workers are intentionally stateless and safe to rerun.

• Private by default
• No background daemons
• No implicit AI processing
• Clear file‑based user intent
• Fail fast, quarantine safely
• Replaceable infrastructure

• Real‑time transcription
• Multi‑language support (currently)
• SaaS transcription vendors
• Automatic AI processing without user intent