Dialogs is a 20.6-hour studio-quality corpus of expressive, conversational Russian speech, designed for dialog-oriented and emotional text-to-speech. Unlike existing Russian corpora — mostly single-speaker read speech or large but low-quality web-mined audio — Dialogs was recorded by professional theatre actors performing scripted dialogs face-to-face, capturing natural turn-taking, timing, and expressive variation.

Presented at Interspeech 2026 — I. Shigabeev and I. Latyshev (Langswap).

Highlights

🎙️ 20.6 hours, 11,796 utterances, avg. 6.3 s/utterance
🗣️ 3 professional actors — Masha (F, 9.9 h), Sveta (F, 4.4 h), Dima (M, 6.2 h)
🎭 12 style/emotion labels (neutral, happy, surprise, sad, disgust, angry, tongue-twister, poem, whisper, arrogance, laughing, fear)
🔊 Studio quality: 44.1 kHz, 16-bit WAV, ~20 dBA ambient noise
📊 Ready-made splits: train / validation / test = 11,428 / 180 / 188 (19.9 h / 0.30 h / 0.37 h); test/dev stratified (5 per speaker per emotion)
✅ Open & commercial-friendly under a responsible-use license

Contents & layout

preview.parquet — a small sample (5 utterances per style, 60 total) with audio embedded. This is what powers the dataset viewer above, so you can listen before downloading. It is not the full corpus.
Full corpus — all 11,796 utterances live in the split files train.csv / val.csv / test.csv (and metadata.csv for everything), with the audio in wavs/.

The CSVs are pipe-delimited (|). Each row:

field	description
`audio_path`	path to the WAV file (`wavs/...`)
`speaker_id`	`M` (Masha), `S` (Sveta), `D` (Dima)
`text`	transcript
`emotion`	style/emotion label (English)
`emotion_on_russian`	style/emotion label (Russian)
`accent_text`	transcript with stress marks (`+` before stressed vowel)
`duration`	seconds

Usage

Preview (60 samples, audio included):

from datasets import load_dataset
preview = load_dataset("langswap/dialogs-ru-emotional-conversations", split="preview")
print(preview[0])  # has a decoded `audio` column

Full corpus — download the repo and read the split files + wavs/:

import pandas as pd
from huggingface_hub import snapshot_download
local = snapshot_download("langswap/dialogs-ru-emotional-conversations", repo_type="dataset")
train = pd.read_csv(f"{local}/train.csv", sep="|")   # audio at f"{local}/{train.audio_path}"

Intended use

Training and evaluation of expressive, conversational, and emotional TTS systems. A VITS2 baseline trained on this corpus is described in the paper. We recommend mixing Dialogs with larger Russian corpora for production-grade results, as per-speaker data is limited (4.4–9.9 h).

License

Dialogs is released under OpenRAIL — an open, royalty-free Responsible-AI license. You're free to download, use, modify, redistribute, and use it commercially, subject only to the standard OpenRAIL use restrictions (no unlawful use, no harming minors, no harmful disinformation, no harassment or defamation, etc.). Full terms: LICENSE.md. The voice actors gave written informed consent for open release and lawful use, including commercial.

If Dialogs is useful to you, please cite the paper (below).

Commercial use, services & support

Commercial use is free under OpenRAIL — you don't need our permission, and we hope you build great things with it.

Commercial inquiries are highly welcome. For $1000 we provide a commercial agreement that includes indemnification rights and other commercial services — a signed license, clean-rights warranty, support, and custom or exclusive expressive-speech data collection to your spec. Email shigabeevilya@gmail.com.

Donations are very welcome and help us record and release more open data — get in touch at the same address.

Citation

@inproceedings{shigabeev2026dialogs,
  title     = {Dialogs: A Studio-Quality Expressive Conversational Russian Speech Corpus for Dialog Assistants},
  author    = {Shigabeev, Ilya and Latyshev, Ilya},
  booktitle = {Proc. Interspeech},
  year      = {2026}
}