Lyra
A Thai audio transcription and summarization dwarf planet — hears what is spoken.
Lyra is the ecosystem's audio transcription utility on port 5005, named after the musical constellation. Combines the diarize library (Silero VAD + WeSpeaker) for speaker diarization, Systran faster-whisper-large-v3 (Thai STT), and qwen2.5:3b-instruct via Ollama for summarization. CPU-only on UM890 Pro. Dwarf planet — specialist utility called by other planets or used directly. Tailscale-only.
Overview
Lyra is the ecosystem's Thai audio transcription utility on port 5005, named after the musical constellation. A dwarf planet — a specialist service called by other planets or used directly, never user-facing on its own.
Architecture
Three FastAPI services behind one entry point. A conductor handles audio preprocessing and async job dispatch; a speech-to-text service runs faster-whisper Large-v3 quantised to int8; a diarizer separates speakers using Silero VAD and WeSpeaker embeddings. Summarisation calls the shared Ollama instance. Everything runs CPU-only on the UM890 Pro, reachable only over Tailscale.
Problem It Solves
Multi-speaker Thai audio is poorly served by general transcription tools, and sending personal recordings to a hosted API is not an acceptable trade. Lyra keeps the whole pipeline local.
Current State
In active development. The Phase 1 pipeline runs on port 5005 and Phase 2 is underway. Model validation was deliberately deferred, so accuracy has never been measured against real Thai meeting audio.