zh high-quality speech corpus

This dataset contains complete zh-CN and zh-TW cleaned corpora and one deduplicated shared audio pool. Audio shared by both corpora is stored once.

All audio is FLAC PCM16. Source sample rate and channel count are preserved. Every packaged item passed a source-to-FLAC bit-identical PCM round trip and a separate full-release decode/hash validation.

Corpus rows use portable locators of the form audio/<subset>/<group>/part-N.parquet#row=N.

Download only selected subsets

from huggingface_hub import snapshot_download

snapshot_download(
    repo_id="ACCOUNT/zh",
    repo_type="dataset",
    allow_patterns=[
        "corpus/*",
        "audio/aishell3/*/part-*.parquet",
    ],
)

Or load an individual configuration:

from datasets import load_dataset
corpus = load_dataset("ACCOUNT/zh", "zh-CN")
audio = load_dataset("ACCOUNT/zh", "audio-aishell3")

Release totals

• Unique audio: 34,959,126
• Audio hours: 62,343.041
• Parquet size: 5.481783 TB (4.985653 TiB)

See release.json for exact per-subset counts, hours, bytes, and corpus checksums.