Androidβs built-in TTS voices are robotic, locked to Googleβs cloud, or missing entirely for your language. If you run a Qwen3-TTS server (or any OpenAI-compatible speech endpoint) you already have great, clonable voices β but Android apps canβt use them.
tts-server-proxy bridges that gap: it installs as a first-class TextToSpeechService,
so it appears in Settings β System β Languages & input β Text-to-speech output like any
commercial engine. Whatever the system asks it to speak, it synthesizes on your server and
streams back to the app that requested it.
Use cases:
ββββββββββββββββ TextToSpeech API ββββββββββββββββββββββ OpenAI-compatible βββββββββββββββββββββββ
β Any Android β ββββ bind βββββββββΊ β tts-server-proxy β ββββ POST ββββββββββΊ β Your GPU server β
β app β βββ PCM 16-bit ββββ β (system engine) β βββ WAV/MP3 ββββββββ β Qwen3-TTS β
ββββββββββββββββ ββββββββββββββββββββββ /v1/audio/speech βββββββββββββββββββββββ
The app speaks OpenAIβs audio-speech dialect:
POST {endpoint}/v1/audio/speech
{
"model": "qwen3-tts-1.7b_G0",
"voice": "clone:VoceItalianaAdulta",
"input": "Text to speak",
"response_format": "wav"
}
MediaCodec plays) is
decoded automatically as a fallback.tts-server-proxy-v0.6.1.apk from the latest release
(or from apk/ in this repository).The engine also works before you ever open the app: if no configuration is found, built-in defaults are used, so a fresh install (or a reboot) keeps talking.
git clone https://github.com/manu69x/tts-server-proxy.git
cd tts-server-proxy
export ANDROID_HOME=/path/to/android-sdk
gradle :app:assembleRelease # or :app:assembleDebug for a debug build
Requirements: JDK 17, Android SDK 35, Gradle 8.10+ (AGP 8.6). Release builds expect
app/keystore.properties + app/release.keystore.jks (not committed β see below);
debug builds work out of the box.
Any endpoint that accepts OpenAI-style /v1/audio/speech and returns WAV/MP3 audio will
work. The stack this app was built and tested against:
Android phone
β Phoenix (wake-on-LAN proxy, OpenAI-compatible front door)
β llama-swap (model router / pool, GPU instances)
β Qwen3-TTS (FastAPI serving Qwen3-TTS 0.6B/1.7B models on one or more GPUs)
You donβt need all of that β a single FastAPI service exposing /v1/audio/speech is enough.
Quick sanity check from any machine on the LAN:
curl -X POST http://YOUR_SERVER:PORT/v1/audio/speech \
-H "Content-Type: application/json" -H "Authorization: Bearer sk-..." \
-d '{"model":"qwen3-tts-1.7b_G0","voice":"clone:MyVoice","input":"hello","response_format":"wav"}'
Any OpenAI-compatible /v1/audio/speech server works β including native
Windows options that need nothing but a double-click:
| Server | Difficulty | GUI | Cloning | Italian | Docs |
|---|---|---|---|---|---|
| Chatterbox-TTS-Server | β ββ Easy (Windows portable, zero install) | β | β | β | guide |
| Kokoro-FastAPI | β β β Medium (Windows/uv or Docker) | β | β preset voices | β | guide |
| AllTalk TTS V2 | β β β Medium (Windows setup script) | β | β XTTS | β | guide |
| Qwen3-TTS on Linux | β β β Advanced | β | β zero-shot | β | guide |
Full comparison + per-server install instructions and the exact app config: docs/servers/README.md.
| Configuration | Synthesis test | Diagnostics |
|---|---|---|
![]() |
![]() |
![]() |
Material 3 UI, follows system dark mode. One model in the list = one voice in Androidβs voice picker.
TextToSpeechService, BIND_TEXT_TO_SPEECH_ENGINE)adb push (one-tap import in the Diagnostics screen)logcat is disabled)Authorization: Bearer| Symptom | Cause / fix |
|---|---|
| Engine doesnβt appear in the TTS list | Reboot once after install; the engine registers itself β no account needed |
| First synthesis takes 10 sβ3 min | Cold start: the server wakes and loads the model. Next requests are fast (~3 s) |
| βSpeech engine problemβ dialog | Update to β₯ v0.3.2 β cancelled requests no longer surface errors |
| No audio on some servers | The app auto-detects MP3 and decodes it; ensure the endpoint returns valid audio |
| Wrong voice used by default | The engine picks the first model matching the requested language β order your model list accordingly |
| Need to see whatβs happening | Open Diagnostics β π debug.log: the service logs every synthesis step there |
Speech, text and API key never leave your LAN. HTTP is used in cleartext by design β run it
on a network you trust (or wire up TLS on your front proxy). Signing keys are not in this
repository: app/keystore.properties and app/release.keystore.jks are gitignored; generate
your own with keytool to sign your builds.
MIT β Β© 2026 Emanuele Viani