Multi-process by necessity
Transcription is Python's world and Discord voice is Node's, so the system is a Node orchestrator coordinating a Python sidecar over IPC rather than compromising on one runtime for both.
← Back to home
Project dossier
Parley
Self-Hosted Systems / Audio
- Node.js + discord.jsOrchestrator, voice capture
- Python + faster-whisperLocal transcription sidecar
- FastAPISidecar service boundary
- SQLite + FTS5Searchable meeting history
- Gemini / OpenAI / OllamaPluggable summarizers
01 // What it is
A self-hosted Discord bot that records voice channels, transcribes the audio locally, and turns it into structured meeting notes. Privacy is the point: transcription runs on the operator's own machine, and no audio is uploaded to a cloud service.
02 // Why it exists
Every meeting-notes product wants your audio on their servers. For a private Discord community or a small team, that trade is unacceptable — so the transcription step runs locally and the LLM backend is swappable, including a fully offline Ollama option.
03 // Surfaces
Surface
Role
Recorder
Node orchestrator
Per-speaker capture, auto join/leave, concurrent channels
Transcriber
Python sidecar
faster-whisper over IPC behind FastAPI
Archive
SQLite + FTS5
Full-text searchable meeting history
04 // Engineering notes
Per-speaker, concurrent capture
Audio is captured per speaker and multiple channels record concurrently, which keeps attribution intact in the transcript instead of producing one undifferentiated blob.
Swappable summarizers
Gemini, OpenAI, and Ollama sit behind one interface, so the same deployment can run fully offline or use a hosted model without touching the recording pipeline.
05 // Stack
- Node.js + discord.jsOrchestrator, voice capture
- Python + faster-whisperLocal transcription sidecar
- FastAPISidecar service boundary
- SQLite + FTS5Searchable meeting history
- Gemini / OpenAI / OllamaPluggable summarizers
06 // Honest status
Honest status
Open source and self-hostable, with a landing page. Privacy-first by construction: audio never leaves the host machine during transcription.