Voicevox Voice Corpus: 577 Hours of Synthetic Japanese Audio

Containing 445,793 synthetic Japanese voice recordings totaling over 577 hours of audio generated via the VOICEVOX engine. Created by ayousanz and updated in May 2024, the data is based on the ITA, Tsukuyomi-chan, and ROHAN text corpora.

Use Cases

• Training text-to-speech (TTS) models using the .wav audio files
• Benchmarking automatic speech recognition (ASR) systems against synthetic Japanese voices
• Acoustic analysis of the ITA, Tsukuyomi-chan, and ROHAN text corpora

Strengths

• 445,793 individual .wav files
• 577 hours and 51 minutes of total audio duration
• Standardized text sources including ITA and Tsukuyomi-chan corpora

Limitations

• Synthetic audio may lack the natural prosody and emotional variance of human speech
• Potential for label noise if the VOICEVOX generation encountered phonetic errors

Provenance

Source

ayousanz via Hugging Face

Collection Method

Synthetic generation using the VOICEVOX engine

Freshness

Last updated May 2024.

Geography

Japan