https://github.com/HayaiApp/HayaiTTS/compare/v1.2.0-b1...v2.0.0-b1/

Offline neural text-to-speech for Android, with Material 3 Expressive. HayaiTTS registers itself as a system-wide text-to-speech engine on Android, so any app that talks via android.speech.tts.TextToSpeech — including the Hayai novel reader — speaks with neural voices instead of the stock Google or Samsung engine. Everything runs on-device. No accounts, no cloud, no network during synthesis. Highlights - System TTS engine — appears in Settings → System → Languages & input → Text-to-speech engines. Pick HayaiTTS, set it as the default engine, and every other app on the device now uses neural voices. - 186-voice catalog scraped weekly from the upstream k2-fsa/sherpa-onnx model index, covering 8 model families: - Piper (172 voices, 40+ languages — small, fast, MIT) - Kokoro 82M (3 multilingual voices — SOTA quality) - Kitten (7 nano-tier English voices — fastest) - Matcha-TTS (3 diffusion voices — very high quality) - Supertonic (1 voice — newest 2026 model) - ZipVoice + Pocket (placeholders — will populate on the next catalog refresh after the slug-matcher fix lands) - VITS classics (engine wired; voices appear as upstream re-uploads) - Bundled-in-APK Piper Amy so the app speaks the moment it's installed, before any downloads. - Material 3 Expressive end-to-end: MaterialExpressiveTheme, MotionScheme.expressive() springs, WavyProgressIndicator, LargeFlexibleTopAppBar, MultiChoiceSegmentedButtonRow, MaterialShapes morphing, container-transform-style nav. - Quick voice switcher — a global bottom sheet (top-bar icon on every screen) lists installed voices grouped by language. Tap to set the default for that locale, with haptic feedback. - Featured + favorites + reorder — pinned voices float to the top of the Library, long-press a card to drag-reorder, star toggles favorites. - Per-family visual identity — every voice card carries a gradient tinted by the family's seed color (Piper=teal, Kokoro=violet, Kitten=amber, VITS=cyan, Matcha=green, ZipVoice=indigo, Pocket=rose, Supertonic=gold). The family badge morphs between rounded and polygon shapes during downloads. - Live waveform preview — Voice Detail's Play button feeds a 32-bin amplitude-driven LinearWavyProgressIndicator so you actually see the audio. - Custom voice import — pick any sherpa-onnx-compatible bundle (.tar.bz2, .tar, .zip, or a bare .onnx) via SAF. The analyzer detects the family from the file layout, parses any speakers.txt / speaker_id_map JSON for real multi-speaker support, lets you confirm or override, then installs into the engine's voice library. - Per-locale defaults — set a different voice for English vs. Japanese vs. Chinese; the TTS service routes synthesis requests to the right model based on the locale Android passes in. - Storage location toggle — internal storage or external SD card. When you flip the preference, every installed voice is physically relocated with progress UI (not just persisted). - R8 + per-ABI splits — release APKs land at 96–112 MB per ABI vs. 170 MB for the universal build. - Signed releases via apksigner running in GitHub Actions against a real RSA 4096 keystore stored in encrypted repo secrets. Install 1. Open the app once so it can mirror the bundled Amy voice into private storage. 2. Go to Settings → System → Languages & input → Text-to-speech engines (the path varies by Android skin — look for "Preferred engine"). 3. Pick HayaiTTS. Hit "Listen to an example" to verify Amy speaks. 4. Open HayaiTTS, tap the Browse FAB, install whichever neural voices you want. Hayai's novel reader settings auto-detect HayaiTTS once installed and surface a "Manage neural voices" button that deep-links straight to the Library screen. Licensing Copyright © 2026 Ahmed Mohamed GNU General Public License v3.0