Yougen/wuw_dataset

Wake-Up-Word (WUW) speech dataset, packed as WebDataset tar shards. The input is a Kaldi-style data directory (wav.scp, text, utt2spk, utt2dur, segments), where multiple utterances share a long recording via the segments file.

To avoid duplicating audio, each tar sample corresponds to one full recording. The utterance-level metadata (id / start / end / text / spk / duration) is stored in a JSON list inside that sample. Downstream consumers slice the decoded waveform by [start*sr : end*sr] themselves.

Layout

data/
  <split>/
    metadata.csv          # flattened utterance-level table
    audio/
      <split>-000.tar
      <split>-001.tar
      ...

Shard counts:

• train: 1423 tar shard(s)

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

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

metadata.csv columns (one row per utterance): key, shard, rec_id, rel_path, wav_format, id, start, end, duration, text, spk

Loading

from datasets import load_dataset

ds = load_dataset("Yougen/wuw_dataset")
ex = ds["train"][0]
wav_array = ex["wav"]["array"]
sr = ex["wav"]["sampling_rate"]
for seg in ex["json"]["segments"]:
    s = int(seg["start"] * sr)
    e = int(seg["end"]   * sr)
    print(seg["id"], seg["text"], wav_array[s:e].shape)

Streaming:

ds = load_dataset("Yougen/wuw_dataset", streaming=True)
for example in ds["train"]:
    print(example["__key__"], len(example["json"]["segments"]))
    break