Sky-Kim/supertonic3-unity

supertonic3-tts-unity

Unity 6 demo for Supertonic 3 text-to-speech inference.

References:

• https://github.com/supertone-inc/supertonic
• https://huggingface.co/Supertone/supertonic-3

Environment

• Unity: 6000.4.5f1
• Sentis: com.unity.ai.inference 2.6.1

Quick Start

1. Open Assets/Scenes/SampleScene.unity.
2. Select Supertonic TTS Player.
3. In the Inspector, configure:
   • Language
   • Utterance Text
   • Voice Style Resource Name (M1 by default)
   • Total Step and Speed
4. Enter Play mode.
5. Use the runtime UI:
   • Startup warmup loads the FP16 .sentis models and runs a short silent dry run first, while the lower-right generation label shows the warmup state.
   • Select a language from the dropdown.
   • Select a voice style from the dropdown.
   • Enter the text to speak.
   • Click Play.

Custom Voices

You can create a custom Supertonic voice with Supertonic Voice Builder:

• https://supertonic.supertone.ai/voice-builder

Voice Builder lets you record or upload source audio and create a downloadable voice-style JSON for Supertonic. To use a purchased/exported voice in this Unity demo, place the JSON file under Assets/StreamingAssets/Supertonic/VoiceStyles/. The runtime Voice Style Resource Name dropdown is populated from those JSON files by filename, so a file such as MyVoice.json appears as MyVoice.

Inference Pipeline

The Unity runner follows the official Supertonic inference contract through FP16 .sentis model files in StreamingAssets:

1. duration_predictor_fp16.sentis
2. text_encoder_fp16.sentis
3. vector_estimator_fp16.sentis
4. vocoder_fp16.sentis

The generated waveform is converted into a Unity AudioClip and played through the AudioSource on the demo GameObject.

Notes

• Supported languages: en, ko, ja, ar, bg, cs, da, de, el, es, et, fi, fr, hi, hr, hu, id, it, lt, lv, nl, pl, pt, ro, ru, sk, sl, sv, tr, uk, vi
• Supported bundled voices: F1-F5, M1-M5

License

This repository is a Unity integration demo around Supertonic 3 assets. The upstream Supertonic repository identifies its sample code as MIT-licensed, while the Supertonic 3 pre-trained model weights and related model assets are provided under the upstream OpenRAIL-M terms.