A system text-to-speech engine (TextToSpeechService) for Android that speaks through
a Kokoro server, so any app that can read text aloud can use it.

It is not a plain HTTP client. Kokoro's hard limit is 510 phoneme tokens, and because
that counts the length of the IPA phoneme string rather than characters, a character
budget cannot enforce it: a 42-character Chinese sentence is 545 tokens, while English
prose of the same length is about 42. Text is therefore split on the device by a cost
model fitted to real Kokoro G2P output, and every split is verified lossless - the
same words, in the same order.

What it does:

* Splits text on the device, sized by estimated phoneme tokens rather than characters
* Trims the silence Kokoro pads around every chunk (50 ms before, 360 ms after)
* Inserts pauses at chunk joins, or none at all if the app you listen with already
adds its own
* Starts speaking after a deliberately small first chunk, so turning a page begins in
about a second instead of six
* Fetches the next chunk while the current one is still playing
* Fetches the voice list from the server, so every voice it offers is selectable

Requires a Kokoro server that you run yourself. Any OpenAI-compatible
/v1/audio/speech endpoint works; this was developed and measured against
Kokoro-FastAPI. The endpoint is configurable, so nothing is tied to one host, and
the app sends nothing anywhere except to the server you point it at.

Both http:// and https:// endpoints work, so a server behind a reverse proxy with a
certificate can be reached over TLS. The default is plain http:// because the usual
case is a Kokoro server on your own network.

No ads, no tracking, no accounts, no network access beyond your own server.