ygyuan/kws_dataset

Keyword-Spotting (KWS) speech dataset, packed as WebDataset tar shards. The input is a Kaldi-style data directory (wav.scp, text, utt2spk, utt2dur, segments), where each utterance is packed as a single tar sample.

Layout

data/
  <split>/
    metadata.csv
    audio/
      <split>-000.tar
      <split>-001.tar
      ...

Shard counts:

• train: 2784 tar shard(s)

Inside each tar, every sample is a pair sharing a unique key:

<key>.wav       # raw audio bytes (original format preserved)
<key>.json      # {"id":..., "rel_path":..., "wav_format":"wav",
                #   "duration":..., "text":"<keyword>", "spk":...,
                #   "rec_id":..., "start":..., "end":...}

metadata.csv columns: key, shard, id, rel_path, wav_format, duration, text, spk, rec_id, start, end

Loading

from datasets import load_dataset

ds = load_dataset("ygyuan/kws_dataset")
print(ds)
print(ds["train"][0])
# sample keys: 'wav' (decoded audio), 'json' (metadata), '__key__', '__url__'

For streaming (no full download needed):

ds = load_dataset("ygyuan/kws_dataset", streaming=True)
for example in ds["train"]:
    print(example["__key__"], example["json"]["text"])
    break