ASVspoof 5 Playback-Controlled Extension

This repository is the working source tree for a playback-controlled extension of ASVspoof 5. The intended dataset studies speech liveness under a presentation-constrained threat model: an attacker cannot inject digital audio and must reproduce a signal through a loudspeaker for microphone recapture.

The project contains the prepared, unchanged ASVspoof 5 Track 1 source material and a versioned plan for assigning every retained source to one playback and recapture condition. Physical acquisition has not started.

Intended task

The extension separates content origin from presentation channel:

• CH: clean human speech, used as the live class;
• PH: bona fide source speech after playback and recapture;
• PS: spoof source speech after playback and recapture.

The main liveness task is CH versus {PH, PS}. Original ASVspoof 5 annotations remain authoritative and unchanged; the capture plan adds the channel-related fields needed for the future derived dataset.

Prepared source snapshot

Only Track 1 material is retained:

Partition	Directory	Protocol rows / FLAC files
Train	flac_T/	182,357
Development	flac_D/	140,950
Evaluation	flac_E_eval/	680,774
Total		1,004,081

Track 2-only enrollment audio and Track 2 protocols were excluded. The original partition names and FLAC filenames were not changed.

Repository layout

flac_T/                    ASVspoof 5 training audio
flac_D/                    ASVspoof 5 Track 1 development audio
flac_E_eval/               ASVspoof 5 Track 1 evaluation audio
original-asvspoof5/        upstream documentation and provenance
  protocols/               unchanged Track 1 protocols and codec table
protocols/capture-plan/     capture plan, acquisition guide, schema, and job shards
config/                     acquisition configuration template
scripts/                    capture-plan generator and acquisition executors
LICENSE.txt                upstream license text

The root documentation describes this extension. Upstream ASVspoof 5 documentation is preserved under original-asvspoof5/ so future release files can be developed without overwriting the source records.

Playback capture plan

The distributable capture plan is indexed by protocols/capture-plan/capture-plan.json. It selects all 1,004,081 retained Track 1 sources and assigns each to exactly one of HH, HL, LH, or LL. Assignments are deterministic, stratified, and keep all codec variants sharing one CODEC_SEED in the same condition.

Condition	Capture jobs
HH	251,026
HL	251,015
LH	251,019
LL	251,021

The plan contains 188,819 PH jobs from bona fide sources and 815,262 PS jobs from spoof sources. Each bona fide source remains the unchanged CH reference as well as supplying its PH capture job; no duplicate clean-audio job is created.

Job rows are stored as twelve streamable TSV shards by source partition and playback condition. The index records the labels, destination paths, equipment quality mapping, allocation algorithm, counts, and a SHA-256 for every shard. See protocols/capture-plan/README.md for the recorder contract and field semantics.

Native Windows Python scripts now implement a disposable stratified pilot and the resumable definitive recorder for one condition at a time. Device binding, content-based alignment, FLAC generation, validation, failure diagnostics and the SQLite progress ledger are documented in protocols/capture-plan/ACQUISITION.md.

Integrity and provenance

All 18 upstream audio archives matched the MD5 checks published by ASVspoof 5 before extraction. The source Git revision, initial import commit, exclusions, and exact counts are recorded in original-asvspoof5/PROVENANCE.md.

License and attribution

The retained source material remains subject to the ASVspoof 5 licensing, attribution, ethics, and citation requirements. See LICENSE.txt and original-asvspoof5/README.txt.

The exact equipment bindings and fixed setup must still be approved through pilots for HH, HL, LH, and LL. No physical recaptured audio or final derived release protocol has been produced yet.