Hugging Face's logo

Datasets:
WTForbes
/
nan_new

Modalities:
Text
Formats:
parquet
Size:
1M - 10M
Libraries:
Datasets
Dask
Polars
Dataset card Data Studio Files
xet
 Community
1

You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

nan_new — cleaned Taiwanese Hokkien ASR corpus

This release contains a cleaned full training set and a frozen evaluation set. Audio is embedded in Parquet and can be streamed directly by the Hugging Face datasets library.

Splits

  • train: 3,209,619 rows, 10475.161 hours
  • eval: 709 rows, 1.976 hours

The train split contains 3,209,619 cleaned base rows. No external code-switch corpus is included in this release. The frozen eval audio is physically separated from train. Audio-ID and audio-locator overlap between train and eval are both zero.

Exact transcript strings may occur in both splits when they belong to distinct recordings; common short utterances are not treated as sample leakage.

See release.json for input hashes, cleaning counts, shard checksums, and the base Hub revision used to build this release.
Downloads last month
38
Total file size:
738 GB