Dataset Card for ReazonSpeech

Dataset Summary

This dataset contains a diverse set of natural Japanese speech, collected from terrestrial television streams. It contains more than 35000 hours of audio.

Paper: ReazonSpeech: A Free and Massive Corpus for Japanese ASR

Disclaimer

TO USE THIS DATASET, YOU MUST AGREE THAT YOU WILL USE THE DATASET SOLELY FOR THE PURPOSE OF JAPANESE COPYRIGHT ACT ARTICLE 30-4.

Dataset Format

Audio files are available in FLAC format, sampled at 16000 hz. Each audio file is accompanied with a transcription.

{
    'name': '000/0000000000000.flac',
    'audio': {
        'path': '/path/to/000/0000000000000.flac',
        'array': array([ 0.01000000,  ...], dtype=float32),
        'sampling_rate': 16000
    },
    'transcription': '今日のニュースをお伝えします。'
}

We provide 5 different dataset sizes. Here is the list of available sizes and their approximate recording hours.

Name	Size	Hours
tiny	600MB	8.5 hours
small	6GB	100 hours
medium	65GB	1000 hours
large	330GB	5000 hours
all	2.3TB	35000 hours

You can access this dataset through Hugging Face datasets library.

from datasets import load_dataset
ds = load_dataset("reazon-research/reazonspeech", "all", trust_remote_code=True)

Access the older versions

If you want to access the older versions of ReazonSpeech corpus, you can use the following tags.

Name	Size	Hours
small-v1	350MB	5 hours
medium-v1	22GB	300 hours
all-v1	1TB	19000 hours

License

CDLA-Sharing-1.0

TO USE THIS DATASET, YOU MUST AGREE THAT YOU WILL USE THE DATASET SOLELY FOR THE PURPOSE OF JAPANESE COPYRIGHT ACT ARTICLE 30-4.