De-identified MFCC Features and Speaker-Disjoint Splits: Speaker Lineage Determination under Physical Voice Disguise Using a Siamese Neural Network

De-identified acoustic features and evaluation splits supporting the manuscript "Speaker Lineage Determination under Physical Voice Disguise Using a Siamese Neural Network for Forensic Applications" (submitted to the Australian Journal of Forensic Sciences). The dataset contains Mel-Frequency Cepstral Coefficient (MFCC) feature files for 454,000 speech-sample pairs derived from 100 speakers, covering normal speech and four physical voice disguise conditions: gagging, masking, nasal pinching, and throat constriction. Each .npz file contains one float32 array (key: "mfcc") of shape 160 × N: two vertically stacked folios of 80 MFCC-derived coefficients each (rows 0–79: normal recording; rows 80–159: disguised recording), where N is the number of time frames and varies with sentence duration. Features are stored at full length; a fixed 68-frame window was applied at load time during model training (see README.md). Labels: label0 = different-speaker pair; label1 = same-speaker pair. The train/validation/test splits are speaker-disjoint: no speaker appears in more than one split. All results reported in the associated manuscript were obtained using these exact splits, enabling full reproduction of the evaluation. Contents:- train_label0.zip — 193,200 training pairs (different-speaker)- train_label1.zip — 170,800 training pairs (same-speaker)- test.zip — 45,000 test pairs (8,400 different-speaker; 36,600 same-speaker)- val.zip — 45,000 validation pairs (8,400 different-speaker; 36,600 same-speaker)- MANIFEST.csv — SHA-256 checksum and byte size of every file- README.md — full documentation, file naming convention, and loading example Raw audio recordings are NOT included: they constitute identifiable biometric data collected under informed consent and cannot be shared under institutional ethics committee restrictions. All speaker identifiers have been replaced with anonymous codes.