Source-linked AI summary
An Initial Study of Physiological Markers for Deepfake Speech Detection
Vojtěch Staněk, Michal Frič, Kamil Malinka, Anton Firc
TL;DR
Conventional deepfake speech detectors target synthesis artifacts and can generalize poorly to unseen attacks or recording conditions. This paper evaluates 36 handcrafted physiological markers with logistic regression and SVM classifiers, achieving 21.40% EER in-domain with an RBF SVM and 32.77% EER on the excluded MLAAD/M-AILABS pair with logistic regression. The results provide initial evidence that these markers offer an interpretable detection signal, within a study limited by duration and dataset scope.
Problem
Artifact-based deepfake speech detectors can rely on generator-, codec-, or recording-condition-specific patterns, creating a generalization challenge for unseen attacks and conditions.
Source: §I · p. 1Method
The paper evaluates 36 handcrafted respiratory, vocal, and temporal markers with logistic regression and SVM classifiers on pooled English recordings.
Source: §I · p. 1, §IV · p. 3Results
21.40% EER with an RBF SVM on the full EDS-14 in-domain test set, while logistic regression reaches 32.77% EER on MLAAD/M-AILABS excluded from training.
Source: §V · p. 4Takeaways & Limitations
The evaluated physiological markers contain useful distinguishing information, and logistic-regression coefficients indicate which marker values push scores toward bona fide or deepfake speech.
Source: §VI · p. 4Recordings must be at least 14 seconds, partitions may share speakers, utterances, or generators, and the results do not establish generalization to other datasets.
Source: §VI · p. 4
Abstract
from arXiv · showhide
Conventional deepfake speech detectors primarily target synthesis artifacts and often degrade under unseen attacks or recording conditions. Human speech, however, contains physiological markers associated with respiration, phonation, and temporal coordination arising from the speech-production system that synthesizers may not reproduce faithfully. This motivates the hypothesis that deepfake speech can be detected by identifying inconsistencies in these patterns, which we test using 36 handcrafted physiological markers organized into respiratory, vocal, and temporal categories. We evaluate the marker vectors with logistic regression and SVM classifiers on pooled English recordings. The best in-domain system achieves a 21.40% equal error rate (EER), while logistic regression achieves 32.77% on the out-of-domain dataset pair. Since the markers describe specific speech-production properties, the logistic-regression coefficients indicate whether higher marker values push the classifier toward bona fide or deepfake speech. These results provide initial evidence that physiological markers offer a useful, interpretable signal for deepfake speech detection.
I. INTRODUCTION
Conventional deepfake speech detectors target synthesis artifacts but can generalize poorly to unseen attacks or recording conditions. This work tests whether handcrafted physiological markers provide an interpretable detection signal.
I. INTRODUCTION
- Artifact-focused detectors can rely on generator-, codec-, or condition-specific patterns, limiting generalization to unseen attacks or recording conditions.The introduction also frames deepfakes as threats to speaker verification, fraud prevention, and trust in audio evidence.Source: §I · p. 1
- The detector uses 36 handcrafted respiratory, vocal, and temporal markers to identify inconsistencies in natural speech-production patterns.The hypothesis is that synthesizers may not faithfully reproduce these physiological patterns.Source: §I · p. 1
- 21.40% EER in-domain and 32.77% out-of-domain show a measurable standalone signal on pooled English recordings.The introduction reports these results for the proposed detector evaluated on recordings drawn from several modern datasets.Source: §I · p. 1
- Linear-classifier coefficients provide marker-level explanations of which characteristics support bona fide or deepfake decisions.The low-dimensional feature space is designed to make individual speech-production markers interpretable.Source: §I · p. 1
II. RELATED WORK
Prior deepfake detectors emphasize acoustic artifacts, while physiological properties of speech production remain comparatively underexplored in detection. This work investigates handcrafted markers as a complementary signal.
II. RELATED WORK
- Deepfake detectors range from handcrafted spectral and cepstral features to SSL representations with dedicated detection modules, with strong benchmark performance but persistent out-of-domain generalization challenges.The cited examples include MFCCs, CQCCs, SSL extractors,, and modules such as AASIST and MHFA.Source: §II · p. 1
- The work contrasts biologically grounded speech-production characteristics with methods focused on synthesis artifacts.The marker categories include respiratory, vocal, and temporal properties; the supplied table caption identifies their categorization but gives no individual table entries.Source: §II · pp. 1–2, Table I · p. 2
- Physiological speech research covers breathing, phonatory quality, and temporal fluency, but these properties remain underexplored in deepfake detection.Prior breath-focused approaches include event statistics and BreathNet [17], whereas this work examines a broader, interpretable marker space.Source: §II · p. 2
III. PROPOSED METHOD
The proposed detector represents each sufficiently long English recording with 36 handcrafted markers spanning respiratory, vocal, and temporal properties, then classifies the vector with linear or nonlinear models.
III. PROPOSED METHOD
- Recordings must exceed 14 seconds because several respiratory and temporal markers require enough context for reliable estimation.The analysis is restricted to English speech recordings.Source: §III · p. 2
- Respiratory markers measure breathing behavior and its coordination with speech, including breath frequency, duration, acoustics, inhalation, and speech-breath timing.The method notes that most TTS systems do not explicitly model breathing, potentially yielding missing or artificially placed breath sounds.Source: §III · p. 2
- Vocal markers characterize phonation and vocal stability through F0 statistics, jitter, shimmer, and voice harmonics.These features reflect biomechanical vocal properties associated with voiced speech production.Source: §III · p. 2
- Temporal markers describe macro-level rhythm, fluency, and articulation timing through speech rate, articulation rate, and pause statistics.The cited method text also notes that natural pauses and disfluencies associated with cognitive planning are typically not replicated by synthesizers.Source: §III · p. 2
- The pipeline converts each recording into a 36-dimensional marker vector and applies a linear classifier whose coefficients support marker-level interpretation.The overview groups inputs into respiratory, vocal, and temporal markers.Source: §III · p. 2, Figure 1 · p. 3
- Marker extraction combines breath-interval detection, voice analysis, and timing descriptors; LR and RBF SVM classifiers use training-set imputation and standardization.The fitted preprocessing is reused unchanged for validation, test, and out-of-domain evaluation; LR coefficients represent log-odds changes per standard-deviation increase in a marker, holding others fixed.Source: §III · p. 3
IV. EXPERIMENTAL SETUP
The evaluation uses a pooled English corpus filtered for duration and tests both recordings from the pooled sources and a dataset pair excluded from training. The protocols do not guarantee speaker- or generator-disjoint partitions.
IV. EXPERIMENTAL SETUP
- EDS-14 pools English bona fide and deepfake recordings from multiple public datasets, retaining recordings at least 14 seconds long.The dataset combines sources to increase diversity in speakers, recording conditions, and synthesis methods.Source: §IV · p. 3, Table II · p. 4
- The in-domain protocol splits pooled recordings 70/15/15 into training, validation, and test sets, but speakers, utterances, or generators may overlap across partitions.Some component datasets have incomplete metadata, so the partitions are distinct by recording file rather than guaranteed by speaker or generator.Source: §IV · p. 3
- The out-of-domain protocol excludes MLAAD/M-AILABS from training, but overlap in speakers, utterances, or generators remains possible.Accordingly, the result measures evaluation on a dataset excluded from training, not guaranteed unseen speakers or generators.Source: §IV · p. 4
V. RESULTS
The standalone physiological-marker detector distinguishes bona fide from deepfake recordings in the evaluated settings, with performance varying by classifier and dataset composition.
V. RESULTS
- 21.40% EER for RBF SVM versus 27.60% for LR on the full EDS-14 in-domain test set.This is the best result on the full in-domain corpus.Source: §V · p. 4
- 29.47% EER for LR versus 41.25% for RBF SVM on the reduced in-domain test set excluding MLAAD/M-AILABS.Because training and evaluation composition both differ from the full in-domain setting, this comparison does not isolate the contribution of MLAAD/M-AILABS.Source: §V · p. 4
- 32.77% EER for LR versus 37.83% for RBF SVM on MLAAD/M-AILABS, which was excluded from training.Relative to the reduced in-domain evaluation, LR deteriorates by 3.30 percentage points, while RBF SVM improves by 3.42 percentage points.Source: §V · p. 4
- LR is more stable across the evaluated settings, while RBF SVM achieves the best result only on the full in-domain corpus.Lower EER indicates better performance.Source: §V · p. 4
A. Intrinsic Marker-Level Explainability
The LR model makes marker-level contributions inspectable through coefficients in a semantically meaningful feature space, but those coefficients do not establish causal physiological differences.
A. Intrinsic Marker-Level Explainability
- The LR score is a weighted combination of physiological markers, unlike scores from conventional SSL-based systems’ high-dimensional learned representations.The feature space is semantically meaningful, allowing individual marker contributions to be inspected directly.Source: A. Intrinsic Marker-Level Explainability · p. 4
- A positive LR coefficient means a higher-than-average standardized marker value pushes the score toward deepfake, while a negative coefficient pushes it toward bona fide.Lower-than-average values have the opposite effect.Source: A. Intrinsic Marker-Level Explainability · p. 4
- The strongest coefficients span respiratory, vocal, and temporal markers, but correlations among markers and dataset differences preclude causal physiological interpretations.The coefficients show how the model separates classes in the evaluated datasets, not causal physiological differences between them.Source: A. Intrinsic Marker-Level Explainability · p. 4
VI. DISCUSSION AND CONCLUSION
The study finds a useful standalone detection signal in physiological markers and an interpretable LR model, while emphasizing that results are preliminary and bounded by dataset and recording constraints.
VI. DISCUSSION AND CONCLUSION
- 21.40% EER is the best full in-domain result, achieved by RBF SVM; LR is more stable across the evaluated settings.The authors report that the markers contain useful information for distinguishing bona fide from deepfake speech in the evaluated datasets.Source: §VI · p. 4
- LR coefficients show which marker values push scores toward bona fide or deepfake, without establishing causal physiological differences or relationships that hold in other datasets.Table IV ranks the most influential markers by coefficient magnitude and uses red/green to indicate support for deepfake/bona fide, respectively.Source: §VI · p. 4, Table IV · p. 5
- Recordings must last at least 14 seconds, and overlapping speakers, source utterances, or generators plus single-class datasets limit generalization claims.The authors characterize the results as an initial test, not evidence of generalization to other datasets.Source: §VI · p. 4
- Jitter and shimmer remain sensitive to recording quality and extraction errors.Source: §VI · p. 5
- The approach is an initial step rather than a replacement for neural detectors; future work should test speaker identity, duration, cross-corpus robustness, and marker extraction.The authors also suggest combining markers with state-of-the-art SSL-based detectors as a way to potentially improve performance while preserving marker-level interpretability.Source: §VI · p. 5