E4rris/LibriPhrase_evalset

LibriPhrase Eval Set

Official test/evaluation dataset for LibriPhrase, the short-phrase benchmark for open-vocabulary keyword spotting introduced in:

Shin, J., Lee, J., Ko, S., Han, J., & Berg-Kirkpatrick, T. (2022). Learning Audio-Text Agreement for Open-vocabulary Keyword Spotting. Interspeech 2022. arXiv:2206.15400 — https://arxiv.org/abs/2206.15400

This upload is a verbatim copy of the official test data distributed by the authors on Google Drive: https://drive.google.com/drive/folders/1pv3wVYzhSwDsL1dIQk93zmFBZg458ujE (source repo: https://github.com/gusrud1103/LibriPhrase). All credit belongs to the original authors; this dataset is re-hosted on the Hub for convenience and reproducibility. Please cite the paper above if you use it.

Contents

The repo preserves the original Google Drive top-level structure:

E4rris/LibriPhrase_evalset/
├── train-other-500/                 # 136,461 short-phrase wav clips
│   └── train-other-500/
│       └── <speaker-id>/<chapter-id>/<speaker>-<chapter>-<utterance>_<n>.wav
└── evaluation_set/                  # episode manifests (anchor / comparison pairs)
    ├── libriphrase_diffspk_all_1word.csv   # 118,557 pairs
    ├── libriphrase_diffspk_all_2word.csv   #  70,335 pairs
    ├── libriphrase_diffspk_all_3word.csv   #  12,609 pairs
    └── libriphrase_diffspk_all_4word.csv   #   1,512 pairs

The train-other-500/ branch is generated from the LibriSpeech train-other-500 subset. Files follow the LibriSpeech layout <speaker>/<chapter>/<speaker>-<chapter>-<utterance>_<phrase-index>.wav, where the _<phrase-index> suffix identifies which short phrase (1–4 words) was segmented from that utterance.

CSV schema

Each CSV row is one anchor/comparison pair. Columns:

column	description
anchor	path to the anchor wav (the enrolled keyword audio)
anchor_spk	LibriSpeech speaker id of the anchor
anchor_text	text transcript of the anchor phrase
anchor_dur	anchor duration in seconds
comparison	path to the comparison wav
comparison_spk	LibriSpeech speaker id of the comparison
comparison_text	text transcript of the comparison phrase
comparison_dur	comparison duration in seconds
type	pair type, e.g. diffspk_easyneg, diffspk_hardneg, etc.
target	1 if the comparison is a positive match for the anchor, 0 otherwise
class	class label used in the paper's evaluation protocol

The diffspk_* prefix means the anchor and comparison are spoken by different speakers. Easy / hard negatives (easyneg / hardneg) are split by Levenshtein distance between the anchor and comparison transcripts, giving the LibriPhrase Easy (LPE) and LibriPhrase Hard (LPH) subsets from the paper.

Intended use

Evaluating open-vocabulary keyword spotting / query-by-example systems on short (1–4 word) phrases. Each CSV is a slice by phrase length; a full evaluation typically aggregates results across all four. Positive/negative pairs and the easy/hard split let you report AUC / EER consistently with the paper.

Citation

@inproceedings{shin2022libriphrase,
  title     = {Learning Audio-Text Agreement for Open-vocabulary Keyword Spotting},
  author    = {Shin, Ju-Hee and Lee, Joonhyun and Ko, Seunghyun and Han, Jounyeol and Berg-Kirkpatrick, Taylor},
  booktitle = {Proceedings of Interspeech},
  year      = {2022}
}