SavindyaPrathibhani123/partialspoof-raw-3

PartialSpoof v1.2 — Sampled Subset

Stratified subset of the PartialSpoof v1.2 database for partial spoof detection and segment-level localization research.

Split counts (actual, verified)

Split	Bonafide	Spoof	Total	Sampling strategy
train	2,580	8,557	11,137	All bonafide; spoof stratified by spoof-frame ratio (target was 9,000 — actual yield 8,557 due to category pool limits)
dev	2,548	2,548	5,096	All bonafide; random spoof
eval	7,355	7,355	14,710	All bonafide; random spoof

Repository structure

train/bonafide.zip
train/spoofed.zip
dev/bonafide.zip
dev/spoofed.zip
eval/bonafide.zip
eval/spoofed.zip
labels/train_labels.csv
labels/dev_labels.csv
labels/eval_labels.csv
segment_labels/train/train_seglab_0.16.npy   (+ 6 other resolutions)
segment_labels/dev/dev_seglab_0.16.npy
segment_labels/eval/eval_seglab_0.16.npy
manifests/train_manifest.json
manifests/dev_manifest.json
manifests/eval_manifest.json

Label format

label_int: 1 = bonafide, 0 = spoof

utt_id naming: bonafide utterances use ASVspoof-style IDs (LA_T_*, LA_D_*, LA_E_*); spoof (partial) utterances use CON_* IDs.

Segment labels: numpy defaultdict keyed by utt_id, covering ALL utterances (bonafide and spoof). Values are frame-level label arrays ('1'=bonafide frame, '0'=spoof frame) at 7 temporal resolutions (0.01s to 0.64s). Bonafide utterances have uniformly '1' arrays; spoof utterances have mixed '0'/'1' patterns reflecting the partial-spoof regions.

Load with:

import numpy as np
data = np.load('train_seglab_0.16.npy', allow_pickle=True).item()
frames = data['CON_T_0000029']  # array of '0'/'1' strings

Verification

All splits verified: archive integrity, label/CSV/manifest consistency, segment label coverage (100%), audio decodability, and genuine partial-spoof frame patterns (94-96% of spoof utterances show mixed 0/1 frames; 100% of bonafide utterances are uniformly bonafide-labeled).

Source

Original dataset: PartialSpoof v1.2, Zenodo record 5766198, CC-BY 4.0. Sampling seed: 42.