---
title: Dialogs-RU · Expressive Russian TTS
emoji: 🎭
colorFrom: indigo
colorTo: pink
sdk: gradio
sdk_version: 5.49.1
python_version: "3.10"
app_file: app.py
pinned: false
license: openrail
short_description: Expressive Russian TTS — 3 voices, 13 emotions (VITS2)
tags:
  - text-to-speech
  - russian
  - vits2
  - expressive-tts
  - emotional-tts
---

# 🎭 Dialogs-RU · Expressive Russian Text-to-Speech

A free, CPU-friendly demo of a **VITS2** speech synthesizer for **Russian**, trained on the
**[Dialogs](https://huggingface.co/datasets/langswap/dialogs-ru-emotional-conversations)** corpus —
a studio-quality, expressive, *conversational* Russian speech dataset.

Pick one of **3 studio voices** (Masha 👩, Sveta 👩, Dima 👨) and one of **13 emotional styles**
(neutral, happy, surprise, arrogance, yawn, fear, laughing, whisper, disgust, angry, sad, tongue-twister, poem),
type some Russian text, and the model speaks it.

The text front-end runs in two steps:
**[`rutextnorm`](https://github.com/shigabeev/russian_tts_normalization)** spells out numbers, dates, money,
units and abbreviations (`7,5%` → «семь целых и пять десятых процента`), then
**[`ruaccent`](https://github.com/Den4ikAI/ruaccent)** places the stress automatically.
You can also mark stress yourself with a `+` before the stressed vowel (e.g. `прив+ет`).

## Links
- 📚 **Dataset:** [langswap/dialogs-ru-emotional-conversations](https://huggingface.co/datasets/langswap/dialogs-ru-emotional-conversations) (OpenRAIL)
- 🧠 **Model weights:** [frappuccino/dialogs-ru-vits2](https://huggingface.co/frappuccino/dialogs-ru-vits2)
- 💻 **Training code:** [github.com/shigabeev/vits2-emotional](https://github.com/shigabeev/vits2-emotional)
- 📄 **Paper:** *Dialogs: a studio-quality expressive conversational Russian speech corpus for dialog assistants* — Ilya Shigabeev & Ilya Latyshev (Langswap)

## Notes
This is a research **proof-of-concept** (UTMOS ≈ 3.36 on held-out sentences). It demonstrates that the
Dialogs corpus supports training expressive, conversational Russian TTS even from a modest per-speaker
data budget. Expect lively, dialog-style prosody rather than perfectly clean studio audio. The strongest
styles are *neutral*, *happy* and *sad* (most data); rarer styles such as *whisper* or *poem* are subtler.

License: **OpenRAIL**.