🎭 Dialogs-RU · Expressive Russian Text-to-Speech
A VITS2 speech synthesizer trained on Dialogs, a studio-quality, expressive, conversational Russian speech corpus. Choose one of three studio voices and one of 13 emotional styles, type some Russian, and press Generate.
Runs on CPU — free for everyone to try. 🎉
- 📚 Dataset: langswap/dialogs-ru-emotional-conversations · OpenRAIL
- 🧠 Model weights: frappuccino/dialogs-ru-vits2
- 💻 Training code: github.com/shigabeev/vits2-emotional
- 📄 Paper: Dialogs: a studio-quality expressive conversational Russian speech corpus for dialog assistants (Shigabeev & Latyshev, Langswap)
Voice / Голос
Emotion / Эмоция
0.5 1.8
0 1
Tips
- Numbers, dates, money, units & abbreviations are read out in full —
normalized automatically with
rutextnorm(e.g.7,5%→ «семь целых и пять десятых процента»,№5→ «номер пять»). - Stress matters in Russian. Stress is placed automatically with
ruaccent. To override it, put a+right before the stressed vowel:за́мок→з+амок,замо́к→зам+ок. - Voices: Masha & Sveta are female, Dima is male.
- Styles like laughing, whisper, poem or tongue-twister are rarer in the data, so they are more subtle. Neutral, happy and sad are the strongest.
- This is a research proof-of-concept (UTMOS ≈ 3.36); expect lively, conversational prosody rather than perfectly clean studio audio.
Examples / Примеры
| Russian text / Текст | Voice / Голос | Emotion / Эмоция | Normalize numbers, dates & abbreviations · нормализация (rutextnorm) | Auto stress (ruaccent) · авто-ударение |
|---|
Model & demo released under OpenRAIL · trained on the Dialogs corpus · training code