PlaybackSpoof: A Playback-Based Voice Liveness Task under Trusted Audio Acquisition

PlaybackSpoof is a benchmark for playback-based presentation detection under a trusted audio-acquisition assumption. It asks whether speech reached the sensor through an added physical reproduction–propagation–recapture path, rather than only whether the content originated from a human speaker or a spoofing system.

Here, a replay attack is the reproduction of previously recorded genuine speech. A playback attack is broader: it includes replayed genuine speech as well as physically presented synthetic, converted, adversarially modified, or otherwise spoofed audio. This gives the task a generation-agnostic target whose definition does not change with each new speech-generation method.

ASVspoof 5 Track 1 supplies recent bona fide and diverse spoofed source material together with fixed train, development, and evaluation partitions. It is the source corpus used to instantiate PlaybackSpoof, not a conceptual dependency or the identity of the benchmark. PlaybackSpoof preserves the upstream audio, partitions, protocol rows, labels, and metadata while adding physical playback recordings and presentation-channel annotations.

The canonical maximum-retention release has 1,026,608 examples: 95,032 CH, 418,272 CS, 95,032 PH, and 418,272 PS. The clean and playback version of every selected source form an explicit pair. A balanced factorial manifest with 95,032 examples in each category is included under metadata/ as a derived experimental view and does not duplicate audio.

Important interpretation

The trusted-acquisition assumption excludes downstream digital injection: externally supplied attack audio must reach the system through its microphone. Playback detection is therefore a complementary first decision, not a replacement for every speech countermeasure. Conventional content-origin detection remains a distinct problem.

CH is the unchanged clean ASVspoof 5 bona fide source reference. It is not a new direct-live utterance recorded by the target microphone. Consequently, CH-versus-playback experiments measure presentation-path discrimination and provide a proxy for playback-oriented liveness; they must not be described as strict proof of live human presence.

Category	Content origin	Channel	Release label
CH	human	clean source reference	source_reference
CS	spoof	clean source reference	source_reference
PH	human	playback and recapture	playback
PS	spoof	playback and recapture	playback

Playback samples are assigned to exactly one of HH, HL, LH, or LL. The two letters identify the intended playback-device and recording-device quality levels. Exact devices, distances, capture rates, and the documented limitations are in metadata/acquisition-summary.json.

Benchmark tasks

• Presentation detection: CH versus {PH, PS} asks whether the added physical playback path is present, independently of content origin.
• Content-origin detection after playback: PH versus PS asks whether bona fide-origin and spoof-origin content remain distinguishable after both traverse playback.
• Clean content-origin reference: CH versus CS retains the conventional clean bona fide-versus-spoof comparison.

The separation between content origin and presentation path is central to the benchmark. Including both PH and PS prevents playback from being synonymous with the original spoof label.

Because each condition uses distinct equipment, computer, and indoor environment, cross-condition evaluation is an explicit domain-shortcut stress test. A cue tied only to the training chain should fail off-diagonal; transfer indicates that the learned evidence is shared across physical playback domains, subject to the limitations below.

Files and loading

Audio is stored as deterministic WebDataset TAR shards near 1 GiB. Each example is a pair with a shared opaque key:

<sample_id>.flac
<sample_id>.json

The JSON member contains the complete canonical row, including original ASVspoof 5 fields, added presentation labels, pairing identifiers, source partition, and audio SHA-256. The original partition mapping is train → train, development → validation, and evaluation → test.

from datasets import load_dataset

dataset = load_dataset("mariozenmedina/playbackspoof", streaming=True)
first = next(iter(dataset["train"]))

metadata/canonical-samples.tsv is the authoritative release manifest. release-manifest.json lists every shard with sample counts, byte sizes, and SHA-256; SHA256SUMS supports independent verification.

Integrity and selection

The release contains only Track 1 material. Original audio, filenames, source protocol rows, labels, metadata, and partitions were not modified. Playback recordings passed frozen inventory and acoustic-QC checks. Twelve active-source-overlap dropouts and twenty-one confirmed very-low recordings were excluded; non-failure level warnings were retained without waveform normalization.

The maximum-retention physical cohort contains 513,304 selected playback captures and their 513,304 clean source counterparts. Selection and exclusion details are included in metadata/selection-report.json, metadata/cell-quotas.tsv, and metadata/qc-decisions.tsv.

License, ethics, and attribution

The retained ASVspoof 5 database is distributed under the Open Data Commons Attribution License (ODC-By). Its bona fide data is additionally identified upstream as CC BY 4.0. Review LICENSE.txt and upstream/README.txt before redistribution or use. Users must follow applicable data-protection rules and the ASVspoof ethics guidance, and must not use the material for malicious spoofing, unauthorized access, or harmful impersonation.

Please cite ASVspoof 5:

Xin Wang et al. “ASVspoof 5: crowdsourced speech data, deepfakes, and adversarial attacks at scale.” ASVspoof Workshop 2024, pp. 1–8. https://doi.org/10.21437/ASVspoof.2024-1

The PlaybackSpoof paper/citation should be added here when it becomes available.

Acquisition controls and limitations

Every microphone was oriented orthogonally to the loudspeaker cabinet/front plane. The condition protocols retain the speaker-to-microphone distances (30 cm HH, 20 cm HL, 15 cm LH, and 10 cm LL), playback volumes, microphone gains, devices, audio frontends, and native capture rates. Volume and gain were fixed within each condition with no per-utterance adjustment. Operating-system audio enhancements, automatic gain control, echo cancellation, and noise suppression were disabled. Exact condition values and their protocol sources are listed in metadata/acquisition-summary.json.

The four conditions were recorded in four distinct indoor environments, one environment and one computer per condition. Equipment, computer, distance, native rate, frontend, and environment therefore vary jointly across HH, HL, LH, and LL. Off-diagonal cross-condition evaluation controls against success based solely on memorizing the exact equipment or environment used for training, because both change at evaluation; successful transfer supports, but does not prove, more general playback-path evidence. It does not eliminate every shortcut, since generic device, room, level, or channel cues may also transfer. The corpus does not test a fifth unseen domain. Detailed room geometry was not recorded.