Skip to content

DATASET.md

Latest commit

 

History

History
70 lines (51 loc) · 2.87 KB

DATASET.md

File metadata and controls

70 lines (51 loc) · 2.87 KB
Raw

EDS-14: English-Deepfake-Speech-14

EDS-14 is the long-recording English speech corpus used in the biological-marker deepfake detection experiments described in the accompanying paper.

Only recordings with duration ≥ 14 seconds are included. This threshold is required because several respiratory and temporal markers are unstable on very short clips.

Files in this repository

File Description
data/eds14/eds14_recordings.csv Index of all 82,204 core recordings (six public sources).
data/eds14/eds14_sources_summary.csv Per-source counts for the six core sources.
data/eds14/manifests/ood_speechfake_en_15s.csv Full manifest for the SpeechFake OOD protocol.
data/eds14/manifests/ood_mlaad_m_ailabs.csv Full manifest for the MLAAD/M-AILABS OOD protocol.

sample_id format

{DatasetName}-{BF|DF}-{index:06d}

Examples: ASVspoof5-BF-000001, SpeechFake-DF-012345, M-AILABS-BF-000042.

Columns in eds14_recordings.csv

Column Meaning
sample_id Readable unique ID.
source Original public corpus (ASVspoof5, MLAAD, M-AILABS, …).
file_path Path relative to the EDS-14 audio root.
label 0 (bona fide) or 1 (spoof).
split Split from the MLAAD/M-AILABS OOD reference manifest (train, val, test, ood_mlaad, ood_m_ailabs).
duration_s Measured audio duration in seconds.

OOD evaluation protocols (thesis)

Two out-of-distribution protocols were used in the thesis. Each manifest is under data/eds14/manifests/:

Manifest OOD hold-out In-domain training pool
ood_speechfake_en_15s.csv Entire SpeechFake subset (ood_speechfake_en_15s, 48,879 rows) ASVspoof5, SCDF, In-the-Wild, MLAAD, M-AILABS
ood_mlaad_m_ailabs.csv MLAAD + M-AILABS together (ood_mlaad__m_ailabs, 2,008 rows) ASVspoof5, SCDF, In-the-Wild, SpeechFake

Manifest columns: sample_id, file_path, label, split — compatible with dp_biomarkers.data.manifest.load_manifest_csv.

Core corpus (82,204 files)

EDS-14 folder Public source Bona fide Spoof Notes
speechfake_en_14s SpeechFake 1,947 46,932 English subset, ≥ 14 s
asvspoof5 ASVspoof5 9,208 20,121 English clips, ≥ 14 s
scdf SCDF 0 1,373 Spoof-only source
mlaad_fake_en14s MLAAD 0 1,159 English MLAAD spoofs
m_ailabs_ge_14s M-AILABS 849 0 English bonafide
in_the_wild_14s In-the-Wild 194 421 Filtered to ≥ 14 s

Expected on-disk layout

EDS14_ROOT/
|-- speechfake_en_14s/
|-- asvspoof5/
|-- scdf/
|-- mlaad_fake_en14s/
|-- m_ailabs_ge_14s/
`-- in_the_wild_14s/

Resolve paths as EDS14_ROOT / file_path from the metadata CSVs.