tts-server-proxy speaks the OpenAI /v1/audio/speech dialect: it POSTs
{model, input, voice, response_format} with Bearer auth and plays back the
returned WAV. Any server that implements that contract works — here are the
ones we verified, with honest difficulty ratings.
| Server | Difficulty | GUI | Voice cloning | Italian | Native Windows | License |
|---|---|---|---|---|---|---|
| Chatterbox-TTS-Server | ★☆☆ Easy | Web UI | ✅ reference audio | ✅ (Multilingual engine) | ✅ Portable, zero install | MIT |
| Kokoro-FastAPI | ★★☆ Medium | Optional Web UI | ❌ (preset voices only) | ✅ preset voices | ✅ PowerShell/uv (or Docker) | Apache-2.0 |
| AllTalk TTS V2 | ★★☆ Medium | Web panel | ✅ XTTS/F5 | ✅ (XTTS or Piper it) | ✅ setup script | AGPL-3.0 |
| Qwen3-TTS on Linux | ★★★ Advanced | ❌ CLI + config | ✅ zero-shot | ✅ | ❌ (Linux/WSL2) | varies |
Which one for me?
start.bat, Python embedded — runs from a USB stick).In the app (DiagActivity → ⬇ Import config or directly in the UI):
http://<server-ip>:<port> — no trailing /v1, the app appends /v1/audio/speech itself.label is the name Android shows,
model/voice are the server-side identifiers, lang is it/en/…Example:
{
"endpoint": "http://192.168.1.50:8004",
"api_key": "",
"models": [
{ "label": "Italian voice", "model": "chatterbox", "voice": "my-clone", "lang": "it" }
]
}
Security note: these servers are meant for your LAN. Don’t port-forward them to the internet; if you need remote access use a VPN (Tailscale works great — the author’s setup runs over it).