Spoofing detectors for speaker verification (systems that decide whether a voice sample comes from a real person or from a speech synthesizer or voice converter) are nearly perfect on the attack types they were trained on and fail badly on new ones. On the crowdsourced ASVspoof 5 benchmark, the standard AASIST detector goes from 17.83% equal error rate (EER, the operating point where false accepts equal false rejects) on development data to 29.12% on evaluation data. On one old-fashioned attack, unit-selection concatenative synthesis (attack A12), it reaches 78.9% EER, which is worse than chance (ASVspoof 5). The detector has learned what known generators sound like. It has not learned what real speech must satisfy.
Cosmology faced a similar problem: how to test whether one model explains all the data when no alternative is known in advance. The answer was the posterior predictive consistency test. You fit the model on one part of the data, predict a held-out part, and ask how surprising the actual held-out data are, using p-values calibrated by simulation. The Dark Energy Survey used this to find a mild disagreement between large and small angular scales and to trace an anomaly to one specific redshift-bin combination (Doux et al.). We propose the same test for speech. A model trained only on genuine speech predicts one acoustic "probe" of an utterance from another. Examples are high-frequency detail from low-frequency structure, phase from magnitude, and one time segment from its neighbours. An utterance is flagged when its probes disagree more than genuine speech ever does. The p-values are calibrated separately for each transmission codec.
Core Claim
Unseen attacks will be easier to detect as failures of internal consistency than as the presence of generator fingerprints. This should hold most strongly for attacks that contain no neural vocoder artifacts. Unit-selection synthesis is built by joining fragments of real recordings. Each local stretch is genuine, so a fingerprint detector sees nothing wrong. What fails is agreement across time: a segment does not match what its neighbours predict. A temporal leave-one-segment-out consistency test should therefore expose A12 where AASIST fails.
The second part of the claim is calibration. Raw reconstruction error depends on codec and content, so a single threshold cannot transfer across conditions. A p-value calibrated against genuine speech passed through the same codec can. Under that calibration, false-alarm rates on genuine speech should match their nominal level across all 11 ASVspoof 5 conditions.
Relation to Prior Work
The method comes mainly from Doux et al.. They condition on a data subset, draw predictive realizations of a held-out subset, and calibrate a χ²-like statistic by Monte Carlo. We reuse that machinery but must drop its Gaussian likelihood assumption, which they themselves flag as a limitation. Speech probes are far from Gaussian, so we use a learned conditional generative model.
Handley's suspiciousness statistic motivates one design choice. It subtracts an information term from an evidence ratio so that disagreement is not confused with prior volume. The audio analogue is that low-content utterances, such as monotone or near-silent speech, produce small prediction errors regardless of authenticity. We therefore normalize each utterance's discrepancy by the model's own predictive spread on that utterance. We do not run nested sampling. Handley's runs cost 12 CPU-years, which is irrelevant to a detector that must run in milliseconds. Our predictive model is amortized and needs only a single forward pass, plus a few samples.
The DES Y3 analysis (Abbott et al.) supplies a constraint rather than a method. Consistency tests are only trustworthy in the regimes where the predictive model is reliable, and DES had to cut small scales it could not model. We will likewise exclude probes on which the genuine-speech model is itself poorly calibrated, and report those exclusions.
On the audio side, two genuine-only methods are the closest existing work. FG-LCNN trains a genuine-only autoencoder whose transformed features feed a supervised classifier. It reports 4.07% EER on ASVspoof 2019 LA and shows that training the transformer on spoofed audio works worse. GFL-FAD masks 30% of spectrogram patches, trains reconstruction on genuine speech only, and fuses the features into an AASIST classifier. It reports 0.25% EER and 0.0071 min t-DCF on 2019 LA.
GFL-FAD's masked reconstruction is already a form of "predict one part from another," so the distinction must be stated precisely. It masks random patches, uses reconstruction as an internal feature, and its final decision comes from a supervised classifier trained on spoofs. We use physically structured, disjoint splits (band, magnitude/phase, time). We turn each split into a calibrated test statistic with no spoofed training data. The per-split p-values also localize the source of disagreement, as Doux et al. located their anomalous bin. Whether structure matters more than random masking is an empirical question, and we test it directly.
Relative phase features show that pitch-synchronized phase strongly distinguishes genuine from spoofed speech. They report 0.002% EER on known attacks but 3.726% average on unseen ones. We use that phase representation as one probe, predicted from magnitude rather than classified directly.
FADEL attaches evidential uncertainty to a supervised detector. Its uncertainty correlates with per-attack EER, and it reports 5.60% EER when trained on 2019 LA and tested on 2021 LA, against 6.10% for sharpness-aware minimization. Its uncertainty is still defined relative to the spoof classes seen in training. Our score does not depend on spoof classes at all.
Research Program
The probes are computed after the ASVspoof 5 preprocessing, which normalizes peak amplitude and trims silence with Whisper. This avoids the silence shortcut documented by Chettri et al.. We define three splits. The band split predicts the 4–8 kHz log spectrum from the 0–4 kHz log spectrum and the fundamental-frequency (pitch) track. The phase split predicts pitch-synchronized relative phase from the magnitude spectrogram. The temporal split predicts each roughly 200 ms segment from its neighbours on both sides, leaving one segment out at a time.
For each split, a conditional generative model is trained on bona fide speech only. A conditional normalizing flow is preferred because it gives exact predictive densities. Training data include bona fide audio passed through the ASVspoof 5 codec set, with the codec given as an input to the model.
For an utterance with probes $D_A$ and $D_B$, the statistic is
$$T = -\frac{1}{|D_B|}\log q_\phi(D_B \mid D_A, c) - \hat{h}_\phi(D_A, c),$$
where $q_\phi$ is the learned genuine-speech predictive density, $c$ is the codec condition, and $\hat{h}_\phi$ is the model's expected per-dimension negative log density, estimated from its own samples. Subtracting $\hat{h}_\phi$ plays the role of Handley's information correction, so that easy-to-predict utterances are not automatically scored as genuine.
The p-value is the rank of $T$ among statistics computed on a held-out calibration set of genuine speech under the same codec. For the temporal split, we take the most extreme segment and correct for the number of segments. The three split p-values are combined by Fisher's method into one detection score. The individual p-values are also reported for diagnosis.
Baselines
The naive reference point is AASIST trained with the official ASVspoof 5 protocol (AASIST; its reported ASVspoof 5 evaluation result is 29.12% pooled EER and 78.9% on A12, per ASVspoof 5). We retrain it ourselves on identical partitions, using the same compute budget as our method.
The strongest existing method on in-domain data is GFL-FAD, with its reported 0.25% EER on 2019 LA. No reported result was found on Lacuna for GFL-FAD on ASVspoof 5 or on 2021 LA. We will therefore reimplement it with its 0.3 mask ratio and an equal hyperparameter-search budget, and run it on all three benchmarks.
The strongest reported cross-dataset baseline is AASIST-FADEL, at 5.60% EER from 2019 LA to 2021 LA (FADEL). We will reproduce it and also extend it to ASVspoof 5.
Three ablations isolate the proposal's ingredients. The first keeps our flow, data and calibration but replaces structured splits with GFL-style random 30% patch masking; this isolates whether structure matters. The second uses raw $T$ without codec-conditioned calibration; this isolates calibration. The third is a one-class Gaussian density on frozen self-supervised speech embeddings (the wav2vec 2.0 encoder that GFL-FAD compares against), fit on bona fide speech; this tests whether any genuine-only anomaly score is enough.
Our method beats the field if either of the following holds:
- The consistency score alone, with no spoofed training data, achieves lower pooled ASVspoof 5 evaluation EER than retrained AASIST and GFL-FAD.
- Its fusion with AASIST lowers cross-condition EER by a margin larger than the seed-to-seed spread.
Either result must also come with A12 EER far below 78.9%. We do not expect to beat 0.25% EER on in-domain 2019 LA. A loss there is acceptable if it is small.
Why not just…?
Why not just use GFL-FAD's reconstruction error as the anomaly score? This is the most serious threat, because random masked reconstruction already predicts held-out parts from visible ones. The random-mask ablation settles it directly. If random masking matches structured splits on A12 and on ASVspoof 5 overall, then the physically structured split adds nothing, and the contribution reduces to calibration. We would report it as such.
Why not just add evidential uncertainty, as FADEL does? FADEL's uncertainty is learned against known spoof classes, and its 2021 LA result is still several times worse than its in-domain performance. Concatenative speech that resembles genuine audio locally should look confidently genuine to it. The per-attack A12 comparison shows whether that happens.
Why not just augment training with codecs and more attacks, or scale up a foundation model? ASVspoof 5 already adds 32 attacks from 18 teams and 11 codec conditions, and baseline error still rose sharply. GFL-FAD also reports outperforming wav2vec 2.0-based detectors. Augmentation still trains toward generator fingerprints, and A12 has none. We include codec-augmented AASIST as the fair test of this alternative.
Why not just a one-class density on self-supervised embeddings? This is simpler and might work. The third ablation measures it under identical calibration. If it matches our method, the conclusion is that calibrated genuine-only scoring is the active ingredient, not cross-probe prediction.
Decisive Checks
We will train on the bona fide audio of ASVspoof 5 (and, separately, 2019 LA) and evaluate on 2019 LA eval, 2021 LA and ASVspoof 5 eval. Metrics are EER and min t-DCF, reported per attack and per codec. Four results decide the proposal.
The first is whether temporal-split p-values bring A12 EER well below 78.9% while band and phase splits do not. This would confirm the junction-discontinuity mechanism rather than a generic anomaly effect.
The second is whether bona fide false-alarm rates at nominal levels of 1% and 5% stay within sampling error across all 11 codec conditions. The uncalibrated ablation should fail this check.
The third is whether per-split p-values localize failures in a sensible way. Neural vocoder attacks should fail mainly the phase split, and concatenative attacks mainly the temporal split. This is the audio counterpart of tracing an anomaly to a single redshift bin.
The fourth concerns the surrogate adversarial filters, Malafide and Malacopula. These were optimized against classifiers, not against a consistency test, so we will report robustness to them separately. We will also run one adaptive attack that targets $T$ directly, since a fixed test can be gamed once it is known.
References
- ASVspoof 5: Design, collection and validation of resources for spoofing, deepfake, and adversarial attack detection using crowdsourced speech
- Dark energy survey internal consistency tests of the joint cosmological probes analysis with posterior predictive distributions
- Curvature tension: Evidence for a closed universe
- Dark Energy Survey Year 3 results: Constraints on extensions to <mml:math xmlns:mml="http://www.w3.org/1998/Math/MathML" display="inline"><mml:mi mathvariant="normal">Λ</mml:mi><mml:mi>CDM</mml:mi></mml:math> with weak lensing and galaxy clustering
- Light Convolutional Neural Network with Feature Genuinization for Detection of Synthetic Speech Attacks
- Genuine-Focused Learning using Mask AutoEncoder for Generalized Fake Audio Detection
- Relative phase information for detecting human speech and spoofed speech
- FADEL: Uncertainty-aware Fake Audio Detection with Evidential Deep Learning
- Ensemble Models for Spoofing Detection in Automatic Speaker Verification
- AASIST: Audio Anti-Spoofing Using Integrated Spectro-Temporal Graph Attention Networks