ayousanz/css10-ljspeech-multilingual

CSS10 + LJSpeech Multilingual Dataset

A unified multilingual speech dataset combining CSS10 (10 languages) and LJSpeech (English) in a consistent LJSpeech format.

Dataset Description

This dataset merges:

• CSS10: A collection of single-speaker speech datasets for 10 languages
• LJSpeech: High-quality English speech dataset (Linda Johnson)

All audio files are provided in a consistent format suitable for TTS training.

Languages and Statistics

Language	Code	Files	Hours	Speaker ID
English	en	13,100	~24.0	LJSpeech
Spanish	es	11,016	~19.2	CSS10_es
Russian	ru	9,599	~16.9	CSS10_ru
French	fr	8,649	~15.2	CSS10_fr
German	de	7,427	~13.1	CSS10_de
Japanese	ja	6,839	~14.9	CSS10_ja
Dutch	nl	6,145	~10.8	CSS10_nl
Finnish	fi	4,755	~8.4	CSS10_fi
Hungarian	hu	4,514	~7.9	CSS10_hu
Chinese	zh	2,971	~6.5	CSS10_zh
Greek	el	1,844	~3.2	CSS10_el

Total: 76,859 utterances, ~140 hours, 11 speakers

File Structure

├── README.md                    # This file
├── metadata.csv                 # Standard LJSpeech format (id|text)
├── metadata_multispeaker.csv    # With speaker info (speaker|id|text)
├── dataset_stats.json          # Dataset statistics
├── speaker_info.json           # Speaker mapping and descriptions
├── audio_durations.csv         # Audio duration information
└── wavs.zip                    # All audio files (77,296 files)

Metadata Formats

1. Standard LJSpeech Format (metadata.csv)

id|text
de|Hanake hatte allen Körperschmuck...
en_LJ001-0001|Printing, in the only sense...

2. Multi-speaker Format (metadata_multispeaker.csv)

speaker|id|text
CSS10_de|de|Hanake hatte allen Körperschmuck...
LJSpeech|en_LJ001-0001|Printing, in the only sense...

Usage

Loading the Dataset

import pandas as pd
import zipfile
from pathlib import Path

# Extract audio files
with zipfile.ZipFile("wavs.zip", "r") as zip_ref:
    zip_ref.extractall(".")

# Load metadata
metadata = pd.read_csv("metadata.csv", sep="|", names=["id", "text"])

# Load multi-speaker metadata
metadata_ms = pd.read_csv("metadata_multispeaker.csv", 
                         sep="|", names=["speaker", "id", "text"])

For TTS Training (Piper)

# Single speaker training (filter by language)
python -m piper_train.preprocess \
    --input-dir . \
    --output-dir output \
    --dataset-format ljspeech \
    --sample-rate 22050

# Multi-speaker training
python -m piper_train.preprocess \
    --input-dir . \
    --output-dir output \
    --dataset-format multispeaker \
    --metadata-file metadata_multispeaker.csv \
    --sample-rate 22050

File Naming Convention

• CSS10 files: {language_code}_{original_id}.wav (e.g., ja_BASIC5000_0001.wav)
• LJSpeech files: en_{original_id}.wav (e.g., en_LJ001-0001.wav)

License

This dataset combines:

• CSS10: CC BY-SA 4.0
• LJSpeech: Public Domain

Please refer to the original datasets for detailed license information.

Citation

If you use this dataset, please cite both original sources:

@misc{css10,
  author = {Kyubyong Park and Thomas Mulc},
  title = {CSS10: A Collection of Single Speaker Speech Datasets for 10 Languages},
  year = {2019},
  publisher = {Interspeech},
}

@misc{ljspeech17,
  author = {Keith Ito and Linda Johnson},
  title = {The LJ Speech Dataset},
  howpublished = {\url{https://keithito.com/LJ-Speech-Dataset/}},
  year = {2017}
}

Acknowledgments

• CSS10 dataset creators and contributors
• Keith Ito for the LJSpeech dataset
• The css10-ljspeech dataset for providing CSS10 in LJSpeech format