NOTICE.mdWhat the app is built from
The offline voice (tools/fetch-offline-voice.sh; none of it is committed except Tts.kt)
| Part | What | Licence |
|---|---|---|
| Engine | sherpa-onnx 1.13.8, libsherpa-onnx-jni.so, and its Kotlin API file Tts.kt (committed under app/src/main/kotlin/com/k2fsa/sherpa/onnx/) | Apache-2.0 |
| Runtime | ONNX Runtime, libonnxruntime.so (shipped inside the sherpa-onnx release) | MIT |
| Phonemiser | eSpeak NG, linked into sherpa-onnx, and its espeak-ng-data | GPL-3.0-or-later |
| Voice | Piper en_US-ljspeech-medium, a female US English voice trained by Bryce Beattie on the LJ Speech dataset by Keith Ito, repackaged for sherpa-onnx as vits-piper-en_US-ljspeech-medium-fp16 | The recordings are public domain (the model card says "License: public domain"; LJ Speech is LibriVox readings); the Piper code and voice-model packaging are MIT |
Sources, read 2026-10-04: the model card in the voice archive and https://keithito.com/LJ-Speech-Dataset/.
Why this voice and not the smaller Amy: Piper's en_US-amy card gives its dataset licence only as "See URL" (the
mimic3-voices repository), which settles nothing. LJ Speech's terms are plain, so it is the one used.
eSpeak NG is GPL-3.0. The app is private and goes to nobody but its builder. If it were ever handed to anyone, that licence would have to be dealt with first (the phonemiser would be replaced, or the app's source offered).