ruslan-stressed / README.md
stilletto's picture
Repack: pair .wav+.txt inside the same tar (WebDataset layout) and drop standalone texts_part_*.tar
e551436 verified
|
Raw
History Blame Contribute Delete
10.1 kB
---
language:
- ru
license: cc-by-4.0
task_categories:
- text-to-speech
pretty_name: RUSLAN with Word Stress Marks
size_categories:
- 10K<n<100K
tags:
- tts
- speech-synthesis
- russian
- stress-marks
- accent
- ruslan
source_datasets:
- ruslan
---
# RUSLAN with Word Stress Marks · RUSLAN с проставленными ударениями
> **English / Русский**
---
## English
### What is this?
A drop-in replacement for the metadata of the **RUSLAN** Russian single-speaker
TTS corpus, with **word-stress marks added to every multi-syllabic Russian
word** in the transcripts. Audio is bundled unchanged.
The motivation is to train Russian TTS models (e.g. Kokoro, Tacotron, VITS,
StyleTTS, XTTS) that pronounce words with **correct lexical stress**.
Vanilla Russian text does not encode stress, and stress placement is one of the
hardest things for a TTS model to learn from raw orthography alone — especially
for **homographs** (`за́мок` "castle" vs `замо́к` "lock", `му́ка` "torment" vs
`мука́` "flour", etc.). Training on stress-annotated text removes that
ambiguity and dramatically improves prosody.
### Source
- **Audio + original transcripts:** the [RUSLAN corpus](https://ruslan-corpus.github.io/)
by Lenar Gabdrakhmanov, Rustem Garaev and Evgenii Razinkov (Interspeech 2019),
released under **CC BY 4.0**.
- **Stress annotation:** added by Anthropic Claude Opus via the Claude Code CLI,
followed by automatic validation (vowel-position check, text-integrity check)
and small targeted fixes for proper-name typos.
### Contents
```
metadata.csv # 22,200 lines, pipe-delimited: filename|text_with_stress
wavs_part_001.tar # samples 000000…007399 (≈2.1 GB) — wav + paired txt
wavs_part_002.tar # samples 007400…014799 (≈3.2 GB) — wav + paired txt
wavs_part_003.tar # samples 014800…022199 (≈4.2 GB) — wav + paired txt
```
The data is laid out in the **WebDataset** convention: each sample consists of
a `.wav` and a `.txt` sharing the same key (`000123_RUSLAN.wav`
`000123_RUSLAN.txt`), both stored together inside the same `.tar` shard. The
HuggingFace dataset viewer pairs them automatically.
`metadata.csv` is provided as a convenience for users who only want the text;
it is the exact concatenation of all `.txt` files (`filename|text` per line).
The audio is split into three uncompressed `.tar` shards instead of being
uploaded as 22,200 individual files (HuggingFace rate-limits per-file commits
on free tier). After download, extract every shard to get a flat folder with
matched `.wav`/`.txt` pairs:
```bash
mkdir -p data
for f in wavs_part_*.tar; do tar -xf "$f" -C data; done
```
Total: 22,200 mono 22 kHz WAV files (≈31 hours of speech) + matching `.txt`
transcripts with stress marks.
### Stress format
- The stress mark is **U+0301 COMBINING ACUTE ACCENT**, placed *immediately
after* the stressed vowel (so a single grapheme cluster `а́`, `е́`, `о́`, etc.).
- **Every** multi-syllabic word is stressed. Single-syllable words (including
pronouns/prepositions like `я́`, `мне́`, `что́`, `на́`) are also marked.
- The letter **`ё`** is implicitly stressed and is **not** combined with U+0301
(we keep `ё`, never `ё́`).
- Numbers, abbreviations and Latin-script tokens are left without stress.
- Homographs are disambiguated from sentence context by the model.
### Loading example
```python
from datasets import load_dataset
ds = load_dataset("stilletto/ruslan-stressed", split="train")
print(ds[0])
# {'filename': '000000_RUSLAN', 'text': 'С тре́вожным чу́вством беру́сь я́ за́ перо́.', 'audio': {...}}
```
Or directly from CSV:
```python
import csv
with open("metadata.csv", encoding="utf-8") as f:
for line in f:
fn, text = line.rstrip("\n").split("|", 1)
# fn → wavs/{fn}.wav
```
### License
**CC BY 4.0**, inherited from the original RUSLAN corpus.
When using this dataset please cite the original RUSLAN paper:
```
@inproceedings{gabdrakhmanov2019ruslan,
title={RUSLAN: Russian Spoken Language Corpus for Speech Synthesis},
author={Gabdrakhmanov, Lenar and Garaev, Rustem and Razinkov, Evgenii},
booktitle={Interspeech 2019},
year={2019}
}
```
### Known limitations
- Stress was placed by an LLM. While ≥99.9 % of lines pass automated validation
(correct vowel position, text integrity), occasional homograph errors are
possible and should be expected at training scale.
- Audio is unchanged from the original RUSLAN release; no extra cleanup or
re-segmentation has been done.
---
## Русский
### Что это?
Это датасет на базе русского однодикторного TTS-корпуса **RUSLAN**, в котором
к транскрипциям **добавлены ударения для всех многосложных русских слов**.
Аудио прикладывается без изменений.
Цель — обучение русских TTS-моделей (Kokoro, Tacotron, VITS, StyleTTS, XTTS
и т. п.), которые произносят слова с **правильным ударением**. В обычном
русском тексте ударение никак не кодируется, и для модели это одна из самых
сложных вещей, особенно для **омографов** (`за́мок` vs `замо́к`, `му́ка` vs
`мука́` и т. д.). Обучение на размеченном тексте снимает эту неоднозначность
и заметно улучшает просодию.
### Источник данных
- **Аудио и оригинальные транскрипции** — [корпус RUSLAN](https://ruslan-corpus.github.io/)
(Ленар Габдрахманов, Рустем Гараев, Евгений Разинков, Interspeech 2019),
лицензия **CC BY 4.0**.
- **Разметка ударений** — выполнена моделью Anthropic Claude Opus через
Claude Code CLI с последующей автоматической валидацией (проверка позиции
ударения относительно гласной, проверка целостности текста) и точечной
правкой опечаток в именах собственных.
### Состав
```
metadata.csv # 22 200 строк, формат filename|текст_с_ударениями (разделитель |)
wavs_part_001.tar # сэмплы 000000…007399 (≈2,1 ГБ) — wav + парный txt
wavs_part_002.tar # сэмплы 007400…014799 (≈3,2 ГБ) — wav + парный txt
wavs_part_003.tar # сэмплы 014800…022199 (≈4,2 ГБ) — wav + парный txt
```
Данные оформлены по соглашению **WebDataset**: один сэмпл состоит из `.wav` и
`.txt` с одинаковым ключом (`000123_RUSLAN.wav``000123_RUSLAN.txt`), оба
лежат **внутри одного и того же tar-шарда**. HF dataset viewer склеивает их
автоматически и показывает текст рядом с аудио.
`metadata.csv` приложен как удобство для тех, кому нужен только текст; он
содержит ровно те же строки, что и `.txt` файлы (`filename|текст` в строку).
Аудио сложено в три tar-шарда (без сжатия) вместо 22 200 отдельных файлов —
бесплатный тариф HuggingFace ограничивает коммиты пофайлово. После скачивания
распакуйте все шарды в общую папку:
```bash
mkdir -p data
for f in wavs_part_*.tar; do tar -xf "$f" -C data; done
```
Итого: 22 200 моно WAV-файлов 22 кГц (≈31 час речи) + соответствующие `.txt`
транскрипции с ударениями.
### Формат ударений
- Ударение — символ **U+0301 COMBINING ACUTE ACCENT**, ставится *сразу
после* ударной гласной (один графемный кластер: `а́`, `е́`, `о́` и т. д.).
- Ударение проставлено **во всех** многосложных словах. Односложные слова
(включая местоимения и предлоги: `я́`, `мне́`, `что́`, `на́`) тоже размечены.
- Буква **`ё`** ударная по определению, U+0301 после неё **не** ставится
(пишем `ё`, не `ё́`).
- Числа, аббревиатуры и латиница оставлены без ударений.
- Омографы расставлены моделью по контексту предложения.
### Лицензия
**CC BY 4.0**, унаследована от оригинального RUSLAN.
При использовании ссылайтесь на оригинальную статью:
```
@inproceedings{gabdrakhmanov2019ruslan,
title={RUSLAN: Russian Spoken Language Corpus for Speech Synthesis},
author={Gabdrakhmanov, Lenar and Garaev, Rustem and Razinkov, Evgenii},
booktitle={Interspeech 2019},
year={2019}
}
```
### Известные ограничения
- Ударения проставлены LLM. Несмотря на то что ≥99,9 % строк проходят
автоматическую валидацию (корректная позиция, целостность текста), при
обучении на полном объёме возможны единичные ошибки на омографах.
- Аудио не подвергалось переразметке или дополнительной очистке относительно
исходного релиза RUSLAN.