ko cleaned speech corpus

This refresh provides a converged full train manifest, a frozen eval manifest, and a token-coverage subset of approximately 5000 hours. The validated FLAC payloads are reused without decoding or transcoding. Completely referenced Parquet shards are hard-linked locally; partial shards are filtered into exact row-only Parquet files.

Corpus overview

Variant	Rows	Hours
train	20,137,680	35976.329
coverage_5000h	2,919,596	4992.635
eval	1,198	1.997

All audio locators are portable and have the form audio/<subset>/<group>/part-N.parquet#row=N.

The release audit records exact input/output hashes, rejected audio mappings, language-tag validation, and zero train/eval audio-identity overlap.

from datasets import load_dataset
ds = load_dataset("YuKuanFu/ko", "ko-5000h", split="train")

Packaging exclusions

Rows that could not be represented by the validated FLAC pool are omitted and recorded here rather than silently replaced:

Variant	Rows	Hours	Reasons
train	40,733	41.434033	all_zero_audio: 35, duration_mismatch: 40,698
coverage_5000h	7,338	7.364521	all_zero_audio: 5, duration_mismatch: 7,333
eval	1	0.001478	duration_mismatch: 1