UdonTTS【VRChatワールド用音声合成システム】
UdonTTS is text-to-speech system based on neural networks and ShaderGPT. It is designed for Unity 2022 and compatible with VRChat Udon. • UdonTTS V3 uses Kokoro model, and supports English and Japanese in multiple voices. The base model and English support adds 60M build size, and Japanese support adds 30M build size. • UdonTTS V2 uses VITS model. English and Japanese are supported in different models, and it's possible to train custom voices. Note V2 and V3 are not compatible. UdonTTSはニューラルネットワークとShaderGPTに基づいた音声合成システムです。Unity2022用に設計されており、VRChat Udonと互換性があります。 • UdonTTS V3はKokoroモデルを使用し、英語と日本語を複数の音声でサポートします。ベースモデルと英語サポートは60Mのビルドサイズを追加し、日本語サポートは30Mのビルドサイズを追加します。 • UdonTTS V2はVITSモデルを使用。英語と日本語を異なるモデルでサポートし、カスタムボイスのトレーニングも可能です。 V2とV3は互換性がありません。
Demo | デモ
• V3 Kokoro (English & Japanese) https://vrchat.com/home/launch?worldId=wrld_93d99bcf-eb5f-40af-8b46-bf3a1c40f3fc • V2 MMS (English) https://vrchat.com/home/launch?worldId=wrld_80adc04c-d7f0-4d13-87d1-8e6a135ce1d2 • V2 VITS (Japanese) https://vrchat.com/home/launch?worldId=wrld_0a0dc744-b406-41d9-8285-05800ea175ab
License | 利用規約
UdonTTS and ShaderGPT are released under MIT license. ShaderGPT is also available at: https://gitlab.com/lox9973/ShaderGPT
License | 利用規約 (V3 Kokoro)
Kokoro is an open-weight TTS model, released by hexgrad under Apache license 2.0. https://huggingface.co/hexgrad/Kokoro-82M Udon scripts contain compiled code from the following projects: • Open JTalk by Nagoya Institute of Technology is released under the Modified BSD license. https://open-jtalk.sourceforge.net • Flite by Carnegie Mellon University is released under a BSD-like license. https://github.com/festvox/flite
License | 利用規約 (V2 MMS)
"Massively Multilingual Speech (MMS): English Text-to-Speech" by Facebook, is licensed under CC-BY-NC 4.0. https://huggingface.co/facebook/mms-tts-eng
License | 利用規約 (V2 VITS)
The VITS model is trained with phi16's "Pipeline˸ Commentary" speech data, and it is licensed under CC-BY-NC 4.0. https://vrchat.com/home/launch?worldId=wrld_92f5cee0-8416-40ea-961f-3ee68ea691cf Udon scripts contain compiled code from the following projects: • Open JTalk by Nagoya Institute of Technology is released under the Modified BSD license. https://open-jtalk.sourceforge.net • Flite by Carnegie Mellon University is released under a BSD-like license. https://github.com/festvox/flite
Credit | クレジット
• "VITS: Conditional Variational Autoencoder with Adversarial Learning for End-to-End Text-to-Speech" by Jaehyeon Kim, Jungil Kong, Juhee Son https://github.com/jaywalnut310/vits • The Japanese VITS model is trained using "VITS Fast Fine-tuning" by Plachtaa: https://github.com/Plachtaa/VITS-fast-fine-tuning • WASMs are converted to Udon using wasm2usharp by raii: https://github.com/raii-x/wasm2usharp
Changelog | 変更履歴
2025-05-11: v3.0 use Kokoro model 2024-10-24: clarify license 2024-10-23: v2.0 add Japanese 2024-06-08: v1.0 initial release