samples/README.md

VIPBench reviewer-friendly sample subset

This directory holds a 5-speaker slice of the full release so reviewers can inspect the dataset without downloading the full 5.8 GB bundle. Bundle size: **115 MB** (audio + subsetted embeddings).

The format is identical to the full release; analyses that work on data/ work on samples/ by changing one path.

How this sample was created

Selection criterion: one speaker per sociophonetic group, gender-balanced as far as possible at 5 cells (3M + 2F). Within each of the 5 sociophonetic groups, the first speaker (by speaker ID order in data/speakers.csv) of the chosen target gender was selected. The 5 chosen speakers are listed in the table below; their integer codes (1-5) cover the full 5-group stratification scheme of the full benchmark.

For each chosen speaker, the sample includes:

• Their reference audio clip (1 file per speaker, e.g. M01R.wav).
• All 98 comparison clips paired with that speaker as the reference (covers stimulus types 1, 2, 3, and Type 6 morphs anchored on this speaker; Types 4 and 5 different-speaker pairs that use one of the 5 speakers as a comparison are also retained because their reference field is one of the 5).
• All listener judgments on those 490 pairs (6,401 judgments).

Filtering logic (reproducible from the full release):

1. samples/speakers.csv = rows of data/speakers.csv where id ∈ {M01, F06, M11, F16, M21}.
2. samples/stimuli.csv = rows of data/stimuli.csv where reference ∈ {M01, F06, M11, F16, M21}. Yields 490 rows.
3. samples/participant_responses.csv = rows of data/participant_responses.csv where stimuli_id appears in (2). Yields 6,401 rows.
4. samples/audio/reference/ and samples/audio/comparison/ = audio files corresponding to the IDs in (1) and (2).
5. samples/embeddings/<model>.npz = same 10 main + 5 layer-bundle embeddings as the full release, but with each .npz reduced to the 495 keys (5 references + 490 comparisons) corresponding to the sample. Embedding dimensions are unchanged.

The sample was constructed by deterministic filtering of the released CSVs and embedding files (no re-extraction; code/run_all_extractions.sh is not used here). All sampled audio and embeddings are bit-identical to their counterparts in data/.

What's included

Item	Count
Speakers	5 (M01, F06, M11, F16, M21; one per sociophonetic group, 3M+2F)
Reference audio (R.wav)	5
Comparison audio	490 (98 per speaker, all stimulus types)
Listener judgments	6,401
Pre-extracted embeddings	10 models, subsetted to 495 keys each
Per-layer SSL embeddings	5 models, subsetted to 495 keys each
Stimulus types covered	All 6

Layout

samples/
  README.md                        # this file
  speakers.csv                     # 5 rows (subset of data/speakers.csv)
  stimuli.csv                      # 490 rows (subset of data/stimuli.csv)
  participant_responses.csv        # 6,401 rows (subset of data/participant_responses.csv)
  audio/
    reference/                     # 5 *R.wav (symlinks to ../../../exp_2)
    comparison/                    # 490 *.wav (symlinks to ../../../output)
  embeddings/
    rawnet3.npz, ecapa_tdnn.npz, ... (10 models)
    layers/wav2vec2.npz, ... (5 SSL models)

Quickstart

import numpy as np, pandas as pd
from sklearn.metrics.pairwise import cosine_similarity

stim = pd.read_csv('samples/stimuli.csv')
emb = dict(np.load('samples/embeddings/ecapa_tdnn.npz'))

cos = []
for _, row in stim.iterrows():
    ref, cmp = emb[f'{row.reference}R'], emb[row.id]
    cos.append(cosine_similarity([ref], [cmp])[0][0])
stim['cos_ecapa'] = cos

# P(same) target
stim['p_same'] = stim['same_vote'] / stim['num_response']

# Pearson r
print(stim[['cos_ecapa', 'p_same']].corr())

Speaker subset

ID	Group	Gender	Age bracket
M01	1 (New York City English)	M	1 (under 45)
F06	2 (Southern American English)	F	1 (under 45)
M11	3 (African American English)	M	1 (under 45)
F16	4 (Latino English)	F	1 (under 45)
M21	5 (Asian American English)	M	1 (under 45)

See docs/data_dictionary.md for the full integer-to-group mapping and column schemas.

License

CC-BY-NC 4.0 for audio + judgments + embeddings; same as the full release. See ../LICENSE.