Japanese Multi-Speaker Speech Dataset in LJSpeech Format

60,233 speech utterances from 20 Japanese speakers, totaling approximately 90.6 hours of audio. The dataset is formatted for LJSpeech compatibility and optimized for training TTS models like Piper. Audio samples have a 22,050 Hz sample rate.

Use Cases

• Train a multi-speaker TTS model like VITS or Piper using 20 distinct Speaker IDs and their associated utterances.
• Analyze speaker-specific characteristics by comparing the distribution of Utterances across the 20 different speakers.
• Fine-tune a text-to-speech model on Japanese audio data formatted in the LJSpeech structure.
• Generate synthetic speech for 20 different voice profiles using the provided Speaker ID and Internal ID mappings.

Strengths

• Contains 60,233 high-quality speech utterances, providing a substantial volume of training data.
• Features 20 distinct speakers, enabling multi-speaker TTS model development.
• Offers approximately 90.6 hours of total audio content for model training.

Limitations

• Speaker distribution is imbalanced, with the top speaker having 4,675 utterances and others having fewer.
• Limited to a single language (Japanese), restricting cross-lingual applications.
• No explicit metadata columns (e.g., transcript text, phoneme alignments) are described in the input, which may limit certain analyses.

Provenance

Source

Derived from the 'moe-speech-plus-ljspeech' dataset.

Collection Method

Extracted the top 20 speakers by utterance count from the source dataset.

Freshness

Last updated on 2025-12-25.