moe-speech-5speakers-ljspeech

Japanese multi-speaker TTS dataset in LJSpeech format.

Overview

• Speakers: 5
• Total utterances: 19,617
• Format: LJSpeech (wavs/ + metadata.csv)
• Sample rate: 22050 Hz
• Language: Japanese

Speaker Statistics

Speaker ID	Utterances
940de876	4,675
2cf01874	4,632
bbd90363	3,747
1a5a3db8	3,295
4e2f4ba6	3,268

Files

• metadata.csv - Full metadata (id|speaker|text)
• metadata_{speaker_id}.csv - Per-speaker metadata
• wavs_{speaker_id}.tar.gz - Per-speaker audio files

Metadata Format

{file_id}|{speaker_id}|{text}

Example:

940de876_0001|940de876|こんにちは、今日は良い天気ですね。

Usage

Download specific speaker

from huggingface_hub import hf_hub_download

# Download single speaker
hf_hub_download(
    repo_id="midralab/moe-speech-5speakers-ljspeech",
    filename="wavs_940de876.tar.gz",
    repo_type="dataset"
)

Extract files

tar -xzf wavs_940de876.tar.gz -C wavs/

Reconstruct LJSpeech structure

# Create wavs directory
mkdir -p wavs

# Extract all speakers
for f in wavs_*.tar.gz; do
  tar -xzf "$f" -C wavs/
done

Source

This dataset is a subset of moe-speech-plus, containing the top 5 speakers by utterance count.

License

See original moe-speech dataset license.