An on-device text-to-speech library for Android and iOS via Kotlin Multiplatform. Powered by the Supertone supertonic-3 model running locally through ONNX Runtime — no network calls at inference time.

10 built-in voices (5 male, 5 female) with support for 31 languages.

Add the dependency to your build.gradle.kts:

implementation("coach.kima.tts:tts-supertonic-3:1.3.0")

The supertonic-3 ONNX model files (~400 MB) must be bundled with your app. Download them from HuggingFace and place them in your app's assets (Android) and bundle resources (iOS).

The model is licensed under OpenRAIL-M.

Place the model files in your Android assets directory:

app/src/main/assets/tts_models/Supertone_supertonic-3/
  ├── duration_predictor.onnx
  ├── text_encoder.onnx
  ├── vector_estimator.onnx
  ├── vocoder.onnx
  ├── tts.json
  ├── unicode_indexer.json
  └── F1.json, F2.json, ... M5.json   (voice styles)

Initialise the context holder in your Application.onCreate():

AndroidContextHolder.set(this)

Then load the bundled model:

val modelPath = getBundledModelPath("Supertone/supertonic-3")
val sdk = TtsSdk(TtsConfig(modelDir = modelPath!!))
sdk.initialize()

The SDK extracts model files from assets to internal storage on first launch.

Add the Supertone_supertonic-3/ directory to your Xcode project as a folder reference so it's included in the app bundle. Then initialize with the bundle path:

let modelDir = Bundle.main.path(forResource: "Supertone_supertonic-3", ofType: nil)!

val sdk = TtsSdk(TtsConfig(
    modelDir = modelPath,
    defaultVoiceStyle = "M2",
))
sdk.initialize()

sdk.speak("Hello, world!")

sdk.setCallback(ttsCallback {
    onStart { id -> println("Started: $id") }
    onDone  { id -> println("Done: $id") }
    onError { id, err -> println("Error: $err") }
})

// List available voices
val voices = sdk.availableVoices.value  // F1–F5, M1–M5

// Set voice by ID
sdk.setVoiceStyle("F1")

// Or by TtsVoice object
sdk.setVoice(voices.first { it.gender == "FEMALE" })

sdk.speak(SpeakRequest(
    text = "Bonjour, comment allez-vous?",
    voiceStyle = "F3",
    language = "fr",
    speechRate = 1.1f,
    volume = 0.8f,
))

Rendering is the expensive part. Two knobs trade it against audio quality:

val sdk = TtsSdk(TtsConfig(
    modelDir = modelPath,
    flowSteps = 5,          // steps per render when nothing asks for less (model default)
    inferenceThreads = null, // ONNX threads per session; null = platform default
))
sdk.initialize()

// While the app needs the CPU for something else (e.g. live detection), ask for a
// lighter render; raise it back when it doesn't. Applies from the next piece rendered.
sdk.flowSteps = 4

flowSteps is the number of flow-matching steps per render: each step is roughly a fifth of the render time, and below a device-dependent floor (around 3-4) the voice audibly degrades. prerender() always renders at the configured value and that audio is always used, so lowering the live value only affects lines that are not already cached. inferenceThreads is read once at initialize().

Synthesise speech directly to a WAV file without playing it:

val request = SpeakRequest(
    text = "This will be saved to a file.",
    voiceStyle = "F1",
    language = "en",
)

// Write to a WAV file
val sampleRate = sdk.synthesiseToFile(request, "/path/to/output.wav")

// Or get the WAV bytes in memory (e.g. for uploading)
val wavBytes: ByteArray = sdk.synthesiseToBytes(request)

// Or get raw PCM float samples for custom processing
val pcmSamples: FloatArray = sdk.synthesiseToFloatArray(request)

sdk.stop()      // Cancel current speech
sdk.shutdown()  // Release all resources

// In Compose
val isSpeaking by sdk.isSpeaking.collectAsState()
val voices by sdk.availableVoices.collectAsState()

supertonic-3 supports 31 languages:

Codes are ISO 639-1. An unrecognised code falls back to en. The full list is available at runtime from ModelConfig.supportedLanguages, or as SupportedLanguages.all (code + English display name).

Input text is NFKD-normalised before tokenisation, matching the official reference implementation. This is required, not cosmetic: supertonic's character vocabulary is built on decomposed codepoints, so a precomposed "ä" or a Hangul syllable such as "안" is not in the vocabulary at all. Without decomposition those characters are dropped — Korean loses roughly half its input and synthesises near-silence. Any character still outside the vocabulary after normalisation is dropped rather than passed through, since the unsupported marker (-1) would otherwise be read by the embedding lookup as an index from the end of the table and silently substitute an unrelated character.

See CONTRIBUTING.md for project structure, build instructions, and development setup.

MIT License

Copyright (c) 2026 Nate Holland

Permission is hereby granted, free of charge, to any person obtaining a copy
of this software and associated documentation files (the "Software"), to deal
in the Software without restriction, including without limitation the rights
to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
copies of the Software, and to permit persons to whom the Software is
furnished to do so, subject to the following conditions:

The above copyright notice and this permission notice shall be included in
all copies or substantial portions of the Software.

THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
SOFTWARE.

Model weights are licensed separately under OpenRAIL-M by Supertone.