react-native-tts-kit

Neural text-to-speech for React Native and Expo. On-device. Sub-100ms. 31 languages.

import TTSKit from 'react-native-tts-kit';

await TTSKit.speak('Hello, world.');

No API keys. No network. No robotic system voice.

Status: v0.1 alpha. iOS verified on iPhone (iOS 26+). Android scaffolded but not yet validated on a real device. Feedback via GitHub issues.

Why

| | Quality | Offline | Cost | Languages | |---|---|---|---|---| | expo-speech (system) | robotic | ✅ | free | OS-bound | | ElevenLabs / OpenAI TTS | excellent | ❌ | per-request | 30+ | | react-native-tts-kit | neural | ✅ | free | 31 |

There was no good answer for on-device neural TTS in React Native. We needed it for Flowent, so we built it and open-sourced it.

Install

npx expo install react-native-tts-kit
npx expo prebuild --platform ios
cd ios && pod install && cd ..
npx expo run:ios --device

Bare RN: same flow, just install expo-modules-core as a peer dep.

Custom native code can't run inside Expo Go — use a dev build. Standard Expo workflow in 2026.

Use

import TTSKit from 'react-native-tts-kit';

// Default: F1 voice, English
await TTSKit.speak('Hello, world.');

// Pair any voice with any of 31 languages
await TTSKit.speak('Bonjour le monde',  { voice: 'F1', language: 'fr' });
await TTSKit.speak('こんにちは',          { voice: 'M2', language: 'ja' });
await TTSKit.speak('안녕하세요',          { voice: 'F3', language: 'ko' });

// Stream long text — first audio arrives before synthesis finishes
const stream = TTSKit.stream(longArticle, { language: 'en' });
stream.on('chunk', (pcm) => { /* PCM16LE @ 44.1 kHz */ });
stream.on('end',   () => {});

// Stop in-flight synthesis
await TTSKit.stop();

First-launch UX

The model is ~210 MB (multilingual split — 31 languages, fp16 weights) and downloads on first use. Call prefetchModel() from a settings screen so users aren't surprised mid-conversation:

await TTSKit.prefetchModel((p) => {
  setProgress(p.percent); // 0–100
});

Once downloaded, all calls are instant and offline forever.

Voices

10 voices, all language-agnostic. Pair any voice with any language.

| Male | M1, M2, M3, M4, M5 | |-------|---------------------| | Female | F1, F2, F3, F4, F5 |

const voices = await TTSKit.getVoices();
// [{ id: 'F1', name: 'F1', gender: 'female', engine: 'supertonic' }, ...]

Languages (31, verified on-device)

All 31 languages produce intelligible neural-quality audio on iPhone. The example app (example/App.tsx) ships a tappable sample sentence for every one.

| | | | | | |---|---|---|---|---| | Arabic (ar) | Bulgarian (bg) | Czech (cs) | Danish (da) | German (de) | | Greek (el) | English (en) | Spanish (es) | Estonian (et) | Finnish (fi) | | French (fr) | Hindi (hi) | Croatian (hr) | Hungarian (hu) | Indonesian (id) | | Italian (it) | Japanese (ja) | Korean (ko) | Lithuanian (lt) | Latvian (lv) | | Dutch (nl) | Polish (pl) | Portuguese (pt) | Romanian (ro) | Russian (ru) | | Slovak (sk) | Slovenian (sl) | Swedish (sv) | Turkish (tr) | Ukrainian (uk) | | Vietnamese (vi) | | | | |

import { SUPERTONIC_LANGUAGES } from 'react-native-tts-kit';

await TTSKit.speak('こんにちは', { voice: 'F1', language: 'ja' });
await TTSKit.speak('Привет',     { voice: 'M2', language: 'ru' });
await TTSKit.speak('नमस्ते',      { voice: 'F3', language: 'hi' });

Voices are language-agnostic — any voice can speak any language.

Engines

TTSKit.setEngine('supertonic'); // default — neural, on-device
TTSKit.setEngine('system');     // expo-speech fallback (robotic but free)

| Engine | Status | Use when | |---|---|---| | supertonic | ✅ shipping | Production neural TTS | | system | ✅ shipping | Fallback, no model download | | neutts (voice cloning) | ⏳ v1.1 | 3-second voice clone | | cloud:eleven / cloud:openai | ⏳ v1.2 | When you want premium quality |

Performance

Measured on iPhone (iOS 26.4, Debug build):

| Metric | Value | |---|---| | Time to first audio (1 sentence) | ~70 ms | | Real-time factor | < 0.5× | | Cold start (first speak after launch) | ~1–2 s | | Memory (peak during synthesis) | ~250 MB |

Reproducible benchmark harness in benchmarks/. Numbers across more devices (iPhone 14, Pixel 8, mid-tier Android) coming with v1.0.

Privacy

Text passed to speak() and stream() is processed entirely on-device. Once the model is downloaded (one-time, on first prefetchModel() call), no text or audio crosses the network at runtime — synthesis runs in the local ONNX session and audio plays through the platform audio engine.

The only network activity this package performs is the initial model download from HuggingFace (see ATTRIBUTIONS.md for endpoints). Downloads are SHA-256-verified against fingerprints baked into the package when present. If you ship your app in a privacy-sensitive context, an offline mode (e.g. bundling the model file with your app's assets) is also supported — see CONTRIBUTING.md.

Architecture

your app
   ↓
TTSKit                        — public API ([src/index.ts](src/index.ts))
   ↓
SupertonicEngine                 — JS wrapper, listens for native events
   ↓
RNTTSKitModule                — Expo Module bridge ([ios/](ios/), [android/](android/))
   ↓
4 ONNX sessions                  — duration_predictor → text_encoder
                                   → vector_estimator (×8 denoising)
                                   → vocoder
   ↓
Float32 PCM @ 44.1 kHz           — AVAudioEngine / AudioTrack playback

Model weights: Supertonic-3 (99M params, OpenRAIL-M). We host a pinned mirror for resilience.

License

| Component | License | |---|---| | This package | MIT | | Supertonic-3 model weights | BigScience OpenRAIL-M | | ONNX Runtime, Expo Modules | MIT |

If you ship this in your app, OpenRAIL-M requires you to:

1. Bind your end users to the Use Restrictions — most importantly no impersonation/deepfakes without consent and AI-generated audio must be disclosed.
2. Ship a copy of licenses/OpenRAIL-M.txt with your app.
3. Preserve the Supertone copyright notice.

A 2-line ToS clause covers it. See ATTRIBUTIONS.md for boilerplate.

Roadmap

• v1.0 — Android validation, benchmarks across 4+ devices, cleaner first-launch UX
• v1.1 — Voice cloning via NeuTTS Air (3-sec sample → cloned voice)
• v1.2 — Cloud engine adapters (ElevenLabs, OpenAI, Cartesia) behind the same API
• v2.0 — @ttskit/web — same API in the browser, WASM/WebGPU

FAQ

How big is the model? ~210 MB at fp16 (the multilingual split that supports all 31 languages — vector_estimator.onnx alone is 138 MB because it carries cross-lingual weights). Downloaded once on first launch and stored in Application Support / app-private files. fp32 fallback is available — see tools/quantize.md.

Does it work in Expo Go? No — custom native code requires a dev build. Run npx expo prebuild and use npx expo run:ios --device.

Why isn't Chinese in the language list? Supertonic-3's open-source weights cover 31 languages; Mandarin isn't one of them. Use a cloud engine (v1.2) if you need it.

Can I use a custom voice? Voice cloning lands in v1.1. For now, you have 10 preset voices.

Does this work in production? It's v0.1 alpha. The pipeline runs but we've validated on one iPhone. Expect rough edges for ~2 weeks post-launch.

How does this compare to expo-kokoro-onnx? Different model (Supertonic-3 vs Kokoro-82M), more languages (31 vs 8), faster TTFA, no espeak-ng dependency. Both are valid choices; we chose Supertonic for Flowent's needs.

Credits

• Supertone Inc. — for Supertonic-3, the model that does the heavy lifting
• Microsoft ONNX Runtime — inference engine
• Expo Modules — native bridge

Built by @ahk-d for Flowent.