On-device Android text-to-speech built on Inflect by Owen Song (@owenawsong): complete local text-to-waveform stacks under ~10M parameters. This app uses the official ONNX exports Inflect-Micro-v2-ONNX (9.36M, default) and Inflect-Nano-v2-ONNX (3.96M, faster)—24 kHz mono, Apache-2.0. Graphs download once from GitHub Releases (not bundled in the APK). For what the models are, how they were measured, and runtime benchmarks, see Owen’s Inflect v2 evaluation writeup and the Micro v2 / Nano v2 model cards on Hugging Face.
TensorSpeak is a separate Android port and system TTS wrapper; the voices and graphs are Owen’s Inflect releases. Choose Micro or Nano in the app or in system TTS settings, and install the matching model pack in the app’s On-device models section (or pick the voice and let the app download it).
Install and update from GitHub Releases with Obtainium, or download a per-ABI APK manually. Releases are signed; ONNX model packs are separate assets on the same release.
| Device | Asset | Obtainium APK filter |
|---|---|---|
| Phones (almost all) | TensorSpeak-*-arm64-v8a.apk |
arm64 |
| x86_64 emulator / Chromebook | TensorSpeak-*-x86_64.apk |
x86_64 |
| On-device Micro graphs | TensorSpeak-model-micro.zip |
(not an APK — install in-app) |
| On-device Nano graphs | TensorSpeak-model-nano.zip |
(not an APK — install in-app) |
In Obtainium, set APK filter to arm64 on phones so updates do not pick the x86_64 build or the model ZIPs.
After install, enable TensorSpeak under Settings → Accessibility → Text-to-speech output (or your device’s equivalent).
Voice settings: Micro uses variation 0.62 and Nano 0.58 in the default “Balanced” profile; “Stable pronunciation” lowers variation for steadier phrasing.
text → normalize → eSpeak-ng (IPA) → token ids → duration.onnx → decode.onnx → 24 kHz audio
Duration expansion is inside duration.onnx; callers only add Gaussian zp_noise between the two graphs. Tensor names and shapes are documented in docs/TENSOR_CONTRACT.md. Latency levers and on-device benchmarks are in docs/LATENCY.md.
The same pipeline is implemented twice: a Python reference (src/inflect_sandbox/, bit-exact vs upstream) and an Android app (android/, Kotlin + ONNX Runtime + vendored eSpeak-ng 1.52.0). Phoneme and token parity are tested against the reference; waveforms differ slightly on Android because the random number generator is not NumPy’s.
See CONTRIBUTING.md for setup, parity rules, and PR expectations.
Requires Nix (flakes) and direnv.
git submodule update --init --recursive # eSpeak-ng 1.52.0
direnv allow # Python 3.11, uv, Android SDK/NDK
python scripts/fetch_model.py # models/micro/ and models/nano/ (not in git)To work on the Python sandbox only (skip the multi-GB Android SDK):
INFLECT_ANDROID=0 direnv reloadpython scripts/synthesize.py --text "Hello world." --output out/sample.wav
python scripts/parity_check.py # waveform parity vs upstream (micro)
python scripts/parity_check.py --model nano
python scripts/inspect_graphs.py # regenerate docs/TENSOR_CONTRACT.mdpython scripts/export_android_assets.py --skip-models --espeak-data
python scripts/pack_model_assets.py
python scripts/export_frontend_golden.py
cd android && ./gradlew :app:assembleDebug :app:testDebugUnitTestRelease/debug APKs omit the ~54 MB of ONNX graphs (downloaded on demand). Remaining size is mostly ONNX Runtime natives, trimmed eSpeak voice data (~0.9 MB), and libtensorspeak_espeak.so per ABI. For offline debug work you can still run export_android_assets.py --espeak-data without --skip-models so graphs stay in assets as a fallback.
Frontend parity (normalization, phonemizer-compatible punctuation, IPA) is covered on the JVM and on the host where possible:
| Check | Where |
|---|---|
TextNormalizerTest, TextChunkerTest, PhonemeTokenizerTest |
JVM (testDebugUnitTest) |
check_frontend_compat.py |
Host (149-row golden corpus) |
EspeakParityTest |
Device (connectedDebugAndroidTest) |
Agent-oriented architecture notes, release signing, and Obtainium checklist live in AGENTS.md.
Models are gitignored; fetch, pack ZIPs, and export slim assets before a release build (or use the release script, which does all of this):
python scripts/fetch_model.py
python scripts/export_android_assets.py --skip-models --espeak-data
python scripts/pack_model_assets.pyConfigure signing via .signing.env.example → .signing.env (or android/keystore.properties). Keep the keystore outside the repository.
set -a && source .signing.env && set +a
./scripts/build_signed_release_apk.sh # → APKs + out/model-packs/*.zip
./scripts/build_signed_release_apk.sh --upload # refresh APKs + model ZIPs on GitHubPublish releases (not draft-only) with tag matching versionName in android/app/build.gradle.kts, both per-ABI APK assets, both model ZIPs, and links to THIRD_PARTY_NOTICES.md and docs/MODEL_ATTRIBUTION.md.
| Inflect (models, training, Hugging Face releases) | Owen Song (@owenawsong) — Inflect, evaluation & benchmarks |
| ONNX packaging | Robert Bak (webtts-inflect) — graphs used by this app |
| TensorSpeak (Android app, this repo) | AljGe |
TensorSpeak is GPL-3.0-or-later (Copyright (C) 2026 AljGe). Inflect ONNX weights are Apache-2.0; eSpeak-ng is GPL-3.0. See THIRD_PARTY_NOTICES.md and docs/MODEL_ATTRIBUTION.md; the in-app open-source screen mirrors those notices offline.