AISHELL-3

85-hour multi-speaker Mandarin TTS corpus with 218 speakers in clean recording conditions; the standard Chinese multi-speaker TTS baseline.

• Task: tts
• Languages: zh
• Hours: 85
• Domain: read speech
• License: CC BY-NC-ND 4.0
• Homepage: https://www.openslr.org/93
• Paper: https://arxiv.org/abs/2010.11567

Recommendation

Best starting point for multi-speaker Mandarin TTS. Clean studio conditions and a large speaker pool make it suitable for voice cloning research. Non-commercial license — check before production use.

Getting the data

Downloadable via VoxKitchen (aishell3, source: openslr, size: 17.7 GB):

vkit docker download --tag slim aishell3 --root ./data/aishell3

Subsets: data_aishell3.

Suggested processing

A recommended VoxKitchen pipeline ships in the repository at voxkitchen/templates/pipelines/tts-data-prep.yaml — run it with vkit docker run.