Repository navigation
Releases
v0.6.1 — resilient retries & battery-keeper card
What's new
Resilient 5xx retry with exponential backoff. The synthesis POST now retries server errors (HTTP 5xx) up to 6 attempts with a growing backoff (2 s, 5 s, 10 s, 20 s, 30 s). Real-world case: a self-hosted llama-swap box waking from sleep can crash its Vulkan engine spawn 2–3 times in a row before one survives — the previous 3 fixed 2-second retries exhausted their budget while the spawn loop was still dying, and the engine surfaced an error to the client. Fast connection blips (e.g. Tailscale re-establishing after phone idle) are still retried; genuine cold-start timeouts are still not retried, so a sleeping server is only waited on once.
"Keep the engine alive" card. The main screen now has a card that explains the one failure mode no app code can survive — an OEM battery/cleanup manager force-stopping the engine (a force-stopped app cannot receive broadcasts or restarts) — and a one-tap button that opens the system app-info page where the battery/startup exceptions live. Keep the persistent "engine ready" notification visible: if it disappears, open the app once and restart your reader.
Upgrade
Install the APK from apk/tts-server-proxy-v0.6.1.apk in the repository. The in-app config, voices and system default voice survive the update; readers that were connected to the previous engine process need one engine re-selection or app restart to re-bind — an artifact of the update killing the old process, not of daily use.
Compatibility
No protocol changes: still OpenAI-compatible /v1/audio/speech (WAV), same endpoint and configuration format. Full changelog: v0.6.0...v0.6.1
tts-server-proxy v0.6.0 — "Speech engine problem" fixed for good
The reliability release. It fixes the intermittent "Speech engine problem" reported by TTS clients (launchers, e-book readers) after the engine has been idle for a while — requests vanished without a trace even though the server was healthy.
What was happening
The TTS engine service was only ever bound, never started. On Android 12+, a background bind cannot call startForeground(), so the "engine ready" notification was never posted (33 days of uptime: numPostedByApp=0). Whenever the system killed the idle process, clients re-bound to a fresh cached process that could end up frozen — onSynthesizeText() was never invoked and the client showed "Speech engine problem".
What's new in v0.6.0
- Foreground service started from boot — a
BOOT_COMPLETEDreceiver callsstartForegroundService(); boot-started foreground services are exempt from the background-launch restrictions. Verified on Android 16 (nubia ROM): after a reboot, with no app opened, the service runs withisForeground=true,types=mediaPlayback, and the persistent "engine ready" notification exists. START_STICKY— after any process kill the system restarts the service; system-initiated restarts are also exempt from the restrictions.- Self-healing — opening the app re-promotes the service to foreground (always allowed from a visible activity).
- Retry on HTTP 5xx (3 attempts, 2 s apart) — recovers from engines that crash on their first spawn after a cold wake (e.g. a Vulkan rig waking from S5). True timeouts are deliberately not retried, to avoid duplicating slow synthesis.
- New permission:
RECEIVE_BOOT_COMPLETED. On first launch Android still asks for the battery-optimization exemption — keep it allowed.
Upgrade notes
If a client app was bound to the old engine, re-select the TTS engine once after updating (one-time artifact of app updates, not a bug).
Files
APK: apk/tts-server-proxy-v0.6.0.apk in the repo (direct link).
Docs: Compatible TTS servers — ready-made Windows/Linux servers that speak the same OpenAI-style API (Chatterbox TTS Server, Kokoro-FastAPI, AllTalk TTS V2, Qwen3-TTS on Linux).