## Summary
/e/OS currently ships eSpeak as the system default text-to-speech engine. eSpeak’s formant synthesis produces very low-quality, robotic output that degrades every component relying on the system TTS service (`TextToSpeech` API): TalkBack/screen readers, navigation voice prompts, read-aloud features.
This issue proposes evaluating and integrating a modern, fully open-source neural TTS engine as the default, keeping eSpeak only as fallback.
Background
- eSpeak was introduced as a replacement after LineageOS removed Pico-TTS (Android 9). Community thread: Alternative to eSpeak?
- eSpeak qualities: GPLv3, ~2 MB including dozens of languages, fully offline. This explains the original choice.
- Limitation: formant-based synthesis, quality far below current standards; recurring user complaints for years, no alternative shipped by default. Users must manually discover and install third-party engines (RHVoice, SherpaTTS…), which is not discoverable.
Proposed solution
Evaluate **SherpaTTS** as candidate default engine:
- F-Droid package: `org.woheller69.ttsengine` (upstream: GitHub - woheller69/ttsEngine · GitHub , GPLv3)
- Based on next-gen Kaldi, loading **Piper** voice models ( GitHub - rhasspy/piper: A fast, local neural text to speech system · GitHub )
- Fully offline after a one-time model download from Hugging Face (only network permission: initial model fetch)
- Multilingual, including French (e.g. `fr_FR-upmc-medium`), German, Spanish, Italian…
- Neural (VITS-based) voices: quality comparable to mainstream mobile TTS engines
Alternative candidates considered:
- **RHVoice** (GPL): good quality, smaller footprint, but **no French support** — insufficient for global coverage alone.
- Pico-TTS: removed upstream by LineageOS; re-adding it would mean maintaining it ourselves.
Suggested integration
- Ship the neural engine as default TTS engine in `/e/OS` builds.
- Fetch/download voice models on demand (per-language), similar to the existing voice data download flow, to keep system image size small.
- Keep eSpeak installed as fallback for low-RAM devices and languages not covered by Piper models.
- Ensure `Settings → Accessibility → Text-to-speech output` correctly lists both engines.
- Update /e/OS documentation accordingly.
Acceptance criteria
- [ ] System TTS output uses the neural engine by default on supported devices
- [ ] French, English (+ other major languages) voices available offline
- [ ] eSpeak remains selectable as fallback engine
- [ ] No Google dependencies introduced (checked against /e/OS de-Googling policy)
- [ ] Verified with TalkBack and a navigation app (e.g. Organic Maps voice prompts)