What it is
Qwen TTS Runner is a small LAN-only service on the Zenbook/Jarvis Windows host. It wraps a local Qwen text-to-speech model behind a simple HTTP API so project tools can request voice test clips without opening a GUI on the laptop.
It was built for Sakuga voice experiments first, but the more durable win is the runtime pattern: agent-accessible Windows host, scheduled task, health endpoint, artifacts directory, and a clean source-of-record repo.
Why I run it
Voice experiments need repeatable proof. A one-off notebook or manually launched script is fine for curiosity, but it is poor infrastructure. The runner makes local TTS a service I can smoke test, monitor, and call from other projects.
It also gives future agents a headless path into the Zenbook/Jarvis AI machine. That matters because the Zenbook is where the local model stack lives, but it is not a Proxmox guest.
How I use it
Sakuga treats it as a candidate voice lane: ask the runner for a short clip, fetch the generated WAV, and compare the result against other providers. The CPU smoke path produced a short WAV successfully, but fresh generation reliability and acceleration remain unproven.
For now, CPU smoke is the verified backend. AMD acceleration is interesting, but not proven for this path yet.