HayaiTTS

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

Latest stable / beta APKs: github.com/HayaiApp/HayaiTTS/releases

After 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.

Build from source

You'll need:

• JDK 21 — the JBR that ships with Android Studio works.
• Android SDK with platform 36 + build-tools 36.0.0.
• Git LFS — the bundled Piper voice (28 MB .onnx) and the vendored sherpa-onnx-1.13.2.aar (55 MB) are LFS-tracked.

git clone https://github.com/HayaiApp/HayaiTTS.git
cd HayaiTTS
./gradlew assembleDebug
adb install -r app/build/outputs/apk/debug/app-debug.apk

On Windows, point Gradle at the JBR explicitly so it doesn't pick up a random JRE from PATH:

.\gradlew assembleDebug `
  -Dorg.gradle.java.home="C:\Program Files\Android\Android Studio\jbr"

For signed release builds, see ONBOARDING.md.

Release streams

Stream	Tag pattern	Channel
Stable	v1.0.0	Releases
Beta	v1.0.0-b1	Releases (prerelease)
Nightly	r17	Releases (prerelease) — every push to main

In-app auto-updaters can discriminate cleanly: tags starting with v are stable/beta (look for the -b suffix); tags starting with r are nightlies; all betas and nightlies carry the GitHub prerelease=true flag.

Catalog

The voice catalog (catalog/v1/models.json) is regenerated by tools/catalog/build_catalog.py, which scrapes the sherpa-onnx TTS model index, streams every release tarball through SHA-256 to compute integrity hashes, and emits a deterministic JSON file. The catalog-refresh workflow runs the script every Monday on a free GitHub-hosted runner and commits the new catalog directly to main.

Voice bundles aren't on every release tarball's allow-list yet — when the runtime can't synthesize a family (because sherpa-onnx hasn't shipped the JNI for it), the catalog flags those entries available: false and the Browse screen surfaces them as "Coming Soon".

Licensing

The app itself is GPL-3.0 — see LICENSE.

Per-voice licenses vary and are captured in each catalog entry's license field. The bundled Piper Amy voice is MIT upstream and is redistributed under the same terms. Check the upstream source on Hugging Face / GitHub before redistributing any individual voice commercially.

Related

• Hayai — the manga / novel reader that surfaces HayaiTTS as a TTS engine in its novel reader settings.
• sherpa-onnx (Apache-2.0) — the on-device inference runtime.
• Piper voices — the upstream release artifacts for the Piper family.

Contributing

PRs welcome. The build_check workflow validates every PR against assembleRelease + assembleDebug on Blacksmith. For UI changes, please attach a screenshot or short screen recording — Compose previews and unit tests don't capture motion/haptics, which is where most of our UX value lives.

For architecture, phase-by-phase history, the runtime smoke-test plan, and known limits, see ONBOARDING.md.