picheny/ccu-hf-data

CCU Multilingual Speech Dataset

Dataset Summary

This dataset contains segmented speech audio and transcripts for multiple languages: Korean (ko), Chinese (zh), Hebrew (he), and Turkish (tr).It was inspired by data collected for the DARPA CCU [(Cross-Cultural Understanding Program)] (https://www.darpa.mil/research/programs/computational-cultural-understanding) and meant to resemble that data.

Data is organized by language/split folders such as:

• ko_train
• ko_dev
• ko_test (and similarly for zh, he, tr)

Each split includes:

• <split>.tsv (metadata / transcript table)
• audio files under files_000/, files_001/, ... (for large splits)

Languages

• ko (Korean)
• zh (Chinese)
• he (Hebrew)
• tr (Turkish)

Data Fields

TSV columns are:

• audio path (relative path to wav file)
• transcript / text

Splits

Current split naming convention:

• <lang>_train
• <lang>_dev
• <lang>_test

Example:

• ko_train/train.tsv
• ko_dev/dev.tsv
• ko_test/test.tsv

Attribution

When using or redistributing items from this dataset, please credit each item as:

"[title]" by [uploader] ([uploader_id]), [webpage_url], licensed under Creative Commons Attribution (reuse allowed) (or: license unconfirmed, selected via YouTube CC filter). Audio extracted, segmented, and transcripts cleaned for speech research.

See ATTRIBUTION.txt for additional information and per-language breakdown.

Licensing

This release uses a filtered subset from the original data collection process intended to include Creative Commons licensed content. The license details for the data are in the files license_lang_filtered.csv

Dataset Creation

Audio was segmented from source media and exported as WAV for ASR use. Large splits may store audio in multiple subdirectories (files_000, files_001, ...) to satisfy repository file-count limits.

Intended Uses

• ASR model training
• ASR evaluation across languages
• Multilingual speech research
• Cultural understanding research

Limitations

• Transcript quality may vary by language/source.
• Source-domain bias is possible.
• Users should independently verify licensing and suitability for downstream use.

Citation

If you use this dataset, please cite this dataset page and acknowledge the original source creators listed in attribution files.