LinkedStorm

Whole-home audio assistant. Deploy browser tabs as microphones throughout the house — LinkedStorm captures voice activity, transcribes speech with Whisper, identifies who is speaking, and maintains a searchable daily log of all household conversations.

All processing is local. No cloud services required.

What It Does

Always-on VAD capture — Open capture.html?room=Kitchen on any phone, tablet, or laptop. Silero VAD detects speech and streams only the active segments (no silence stored).
Whisper transcription — GPU-accelerated Whisper large-v3-turbo via the existing Speaches service.
Speaker identification — resemblyzer voice embeddings + cosine similarity, same pattern as frigate-notify. Review unknowns and tag them once; the system recognizes that voice automatically going forward.
Daily log — Timeline view of all utterances, color-coded by speaker and room.
Full-text search — Search across all transcripts (SQLite FTS5).
Ask AI — Natural language queries against the transcript archive via Ollama (e.g. "What did we decide about the vacation?").
Live monitoring — Dashboard joins LiveKit rooms to monitor any room or replay clips.
Daily summaries — Ollama generates a plain-English summary of each day's activity at midnight.

Container Architecture

Container	Port	Purpose
`linkedstorm-pipeline`	8300	WebSocket ingestion, Silero VAD, Whisper STT, speaker embedding, cosine matching, SQLite writes, LiveKit publish
`linkedstorm-api`	8310	FastAPI dashboard + REST API + static UI
`linkedstorm-cron`	—	Nightly Ollama summary, speaker clustering, WAV clip purge

External Dependencies (all pre-deployed in homelab)

Service	Host	Purpose
Speaches/Whisper	192.168.86.150:8000	GPU speech-to-text
LiveKit	192.168.86.150:7880	WebRTC for live monitoring + replay
Ollama (qwen2.5:14b)	192.168.86.150:11434	RAG queries + daily summaries

Adding a Room

Open https://linkedstorm.hobbs.farm/capture?room=YourRoomName on any device with a microphone
Allow microphone access
The device appears in the Rooms tab of the dashboard within seconds
Use the mute toggle on the capture page to pause recording without closing the tab

Speaker Labeling

Unknown utterances appear in the Speakers tab review queue with a Play Clip and Tag Speaker button. Tag a speaker once — their voice embedding is stored and all future utterances from that voice resolve automatically. Mirrors the frigate-notify person-labeling workflow exactly.

Storage

Database: SQLite at /mnt/docker-data/configs/linkedstorm/linkedstorm.db
Audio clips: /mnt/docker-data/configs/linkedstorm/clips/ — 30-day retention (configurable via CLIP_RETENTION_DAYS). Transcripts and embeddings are kept permanently.
NAS: All data on Synology NAS via NFS at 192.168.86.30

Dashboard

https://linkedstorm.hobbs.farm — 5-tab interface:

Daily Log — timeline of utterances, filter by room/speaker
Search — full-text search with keyword highlighting
Speakers — known speakers + unknown review queue
Ask AI — RAG chat against transcript archive
Rooms — live capture device status + mute controls

Deployment

# On hf-pdocker-01
cd /mnt/docker-data/compose/hf-pdocker-01/linkedstorm
docker compose up -d

Compose file: compose/hf-pdocker-01/linkedstorm/docker-compose.yml in the docker repo.

Full design spec: docs/superpowers/specs/2026-05-27-linkedstorm-design.md

3.7 KiB Raw Permalink Blame History