SEAR: Spoofing Evidence-Grounded Audio Reasoning

SEAR is an audio question-answering benchmark for testing whether audio language models can identify and quantify signal-level acoustic anomalies and use them as evidence for audio deepfake detection. Instead of evaluating only a final authenticity verdict, SEAR separates deepfake detection, forgery-cue identification, acoustic measurement, and forensic rationale generation.

SEAR contains four complementary tasks covering acoustic evidence identification and quantification, forensic rationale grounding, and deepfake detection.

Tasks

Config	Task	Format	Evaluation target
t1	Binary classification	One multiple-choice question per audio	Deepfake detection
t2	Forgery cue identification	One multiple-choice question per audio	Acoustic evidence identification
t3	Acoustic feature measurement	Zero or more multiple-choice questions per audio	Acoustic evidence quantification
t4	Forensic rationale generation	One open-ended question per audio	Forensic rationale grounding

T3 contains one question for each applicable reference finding. If no selected feature is anomalous, the audio record is retained with an empty qa_pairs list and task_status.T3 = "not_applicable"; no artificial measurement question is introduced. T4 asks the model to summarize signal-level acoustic evidence in natural language. Its reference rationale verbalizes the verified feature names, measured values, and deviation directions without revealing the authenticity label or attack identity.

Scale

Source partition	Audio	T1	T2	T3 questions	T4
ASVspoof 2019 LA train	25,380	25,380	25,380	96,318	25,380
ASVspoof 2019 LA dev	24,844	24,844	24,844	138,549	24,844
ASVspoof 2019 LA eval	71,237	71,237	71,237	639,204	71,237
ASVspoof 2021 LA eval	101,921	101,921	101,921	353,664	101,921
Total	223,382	223,382	223,382	1,227,735	223,382

The complete benchmark therefore contains 1,897,881 AQA items over 223,382 audio recordings.

Data format

Each JSONL row represents one audio recording. Questions for the same recording remain grouped in qa_pairs, allowing an ALM runner to load the audio once and answer all applicable questions.

{
  "partition": "19LA_dev",
  "file_id": "LA_D_1000752",
  "audio_path": "data/LA/ASVspoof2019_LA_dev/flac/LA_D_1000752.flac",
  "label_offline": "spoof",
  "attack_metadata_offline": "A05",
  "qa_pairs": [
    {
      "qid": "LA_D_1000752_T2",
      "task": "forgery_cue_identification",
      "new_task": "T2",
      "question_type": "multiple_choice",
      "question": "Which acoustic finding best matches the deterministic analysis of this audio?",
      "options": {"A": "...", "B": "...", "C": "...", "D": "..."},
      "answer": "C",
      "answer_text": "...",
      "metadata": {"target_feature": "voiced_ratio", "no_anomaly_target": false}
    }
  ]
}

The paths are repository-relative identifiers for alignment with locally obtained source audio. This repository does not redistribute ASVspoof audio.

Loading

Replace YOUR_NAMESPACE/SEAR with this dataset repository's identifier:

from datasets import load_dataset

t1 = load_dataset("YOUR_NAMESPACE/SEAR", "t1")
t2 = load_dataset("YOUR_NAMESPACE/SEAR", "t2")
t3 = load_dataset("YOUR_NAMESPACE/SEAR", "t3")
t4 = load_dataset("YOUR_NAMESPACE/SEAR", "t4")

example = t2["19LA_eval"][0]
print(example["audio_path"], example["qa_pairs"])

Users must obtain ASVspoof 2019 LA and ASVspoof 2021 LA from their official distribution channels and resolve audio_path relative to their local copy.

Evaluation protocol and leakage prevention

label_offline, attack_metadata_offline, answer, and answer_text are evaluation-side fields. They must not be included in the model prompt. In particular:

• do not reveal authenticity labels or attack identities during inference;
• do not derive normalization statistics from development or evaluation data;
• use bona-fide training speech only when constructing a deployable reference prior;
• use development data only for model selection or threshold selection;
• reserve evaluation partitions for final reporting.

The included reference evidence is an offline benchmark target constructed with dataset labels and available attack metadata. It is oracle supervision for evaluation, not the output of a deployable detector.

Construction and validation

SEAR uses 35 utterance-level descriptors spanning cepstral, spectral, prosodic, temporal, and energy characteristics. Candidate evidence was statistically screened globally and, when annotations were available, attack-wise. T2 presents a verified finding among controlled distractors and a neutral no-anomaly option. T3 masks the measured value of a specified finding and asks the model to recover it from four candidates.

All 223,382 audio records passed automated checks for partition and audio-path alignment, task cardinality, answer mapping, option uniqueness, feature/value traceability, value masking, and exclusion of attack identifiers from model-visible text. File counts and SHA-256 checksums are provided in manifest.json.

For T4, references are generated solely from the verified structured findings and checked for feature coverage, exact measured values, deviation directions, sentence count, and forbidden class or attack inferences before release.

Intended use and limitations

SEAR is intended for evaluating evidence-grounded acoustic reasoning in ALMs and for controlled research on audio deepfake detection. The reference findings are statistical descriptors rather than exhaustive causal explanations of spoofing. Performance may vary under unseen recording conditions, languages, codecs, attacks, and demographic groups. The source benchmarks and their collection processes may introduce additional biases.

Source data and terms

This release provides AQA annotations and source-audio identifiers only. Users are responsible for obtaining the corresponding ASVspoof data and complying with its terms of use. No license is asserted here over the underlying audio recordings.

Citation

The paper citation will be added upon publication. Until then, please cite the SEAR dataset repository and the corresponding ASVspoof source benchmarks.