Pfann-style audio dataset, packed as WebDataset tar shards.
Inside each tar, every sample is a pair sharing a unique key:
. # raw audio bytes (ext == wav / mp3 / flac / ... as in source)
.json # {"audio_id":..., "subset":..., "rel_path":..., "duration":..., "sample_rate":..., "channels":...}
metadata.csv columns:
key, shard, audio_id, subset, rel_path, duration, sample_rate, channels
from datasets import… See the full description on the dataset page:
https://huggingface.co/datasets/Yougen/pfann_testset.
