Which Acoustic Features Survive Compression? Tracker-Estimated First-Formant Statistics for Codec-Robust Audio DeepfakeDetection
Organizations: Department of Electrical and Computer Engineering, Clarkson University, Potsdam, NY 13699, USA · Department of Computer Science, Clarkson University, Potsdam, NY 13699, USA
Abstract
Audio deepfake detectors that work well on clean lab recordings often degrade after telephone or platform compression. This paper asks a specific question: which low-dimensional acoustic descriptors stay informative after codec degradation? We screen 36 utterance-level features(formants, PNCCs, LFCCs, prosody) on ASVspoof 2019 data and test on the official ASVspoof 2021 LA set across seven codec conditions. Of 33 statistically significant descriptors, only four show large practical separation (|r| >0.5, bootstrap 95% CIs), all first-formant (F1) measures, highly redundant (PC1: 79% variance). The payoff is stability: the 36-feature SVM degrades 14 pp under codec mismatch (9.5%-23.5% EER), while the four-feature RBF-SVM degrades only 5.9 pp (17.0%-22.9%). The F1 subset varies 3.4 pp across conditions versus 13.1 pp for the full set. F1 differences persist after controlling for speaker identity (p greater than 10^10, mixed-effects model). The results suggest that physically interpretable F1 statistics offer a compact, codec-stable diagnostic signal that may complement higher accuracy neural detectors.
Figures & tables
| Panel A: Attack conditions | ||
| ID | Type | Description |
| Training (A01–A06, seen) | ||
| A01 | TTS | Neural waveform model |
| A02 | TTS | Vocoder-based TTS |
| A03 | TTS | Vocoder-based TTS (different) |
| A04 | TTS | Waveform concatenation |
| Family | Descriptors (code names) | |
|---|---|---|
| Formant | 8 | F1_mean , F1_std , F1_bw , F2_mean , F2_std , F2_bw , F1_F2_ratio , formant_disp |
| PNCC | 13 | PNCC_0 – PNCC_12 |
| LFCC | 8 | LFCC_0 – LFCC_7 |
| Prosody | 7 | F0_mean , F0_std , F0_max , voiced_ratio , energy_mean , energy_std , duration |
| Total | 36 |
| Stage | Question | Statistic / tool | Observed outcome | Decision implication |
|---|---|---|---|---|
| Distribution check | Are the feature distributions uniformly well modeled by a Gaussian assumption? | Skew inspection across all 36 descriptors | 33 descriptors are near-normal and 3 are clearly skewed | Use one nonparametric test for all features to avoid assumption mismatch |
| Significance screen | Do bona fide and spoofed utterances differ at all on clean data? | Mann–Whitney with Benjamini–Hochberg correction | 33 / 36 descriptors are significant after FDR correction | Significance alone is not selective enough |
| Effect-size filter | Which descriptors separate the classes by a practically meaningful margin? | Rank-biserial with bootstrap 95% CI (10K resamples) | Only 4 descriptors have large effect ( ) | Retain only the four F1 descriptors for deeper analysis |
| Attack consistency | Do the shortlisted features generalize across multiple spoofing systems? | Kruskal–Wallis across A01–A06 | All four shortlisted features differ strongly across the six attacks ( ) | The signal is not tied to a single easy training attack |
| Redundancy analysis | Are the shortlisted descriptors complementary or mostly duplicative? | Pearson correlation and PCA (standardized features) | Strong positive correlations; first principal component explains 79% of the variance | The four “winners” largely track one dominant F1-related direction |
| Per-codec probe | Does the signal survive different codec conditions? | Per-condition EER on official ASVspoof 2021 LA eval | F1 subset: 3.4 pp range. All-36: 13.1 pp range | F1 features are codec-condition stable |
| Family | Total | Large-effect | Survivor rate |
|---|---|---|---|
| Formants | 8 | 4 | 50% |
| PNCC | 13 | 0 | 0% |
| LFCC | 8 | 0 | 0% |
| Prosody | 7 | 0 | 0% |
| Feature | Real | Fake | Gap | [95% CI] | EER |
|---|---|---|---|---|---|
| F1 mean | 113 | .71 [.70,.72] | 19.5% | ||
| F1 bw | 154 | .70 [.69,.71] | 21.6% | ||
| F1/F2 | .050 | .67 [.65,.68] | 23.0% | ||
| F1 var | 46 | .64 [.63,.65] | 17.9% |
| Condition | AASIST | F1 (4) | All 36 |
|---|---|---|---|
| None (unseen attacks only) | 23.7 | 21.6 | 18.0 |
| G.722 (wideband) | 26.0 | 22.4 | 18.2 |
| Opus (wideband) | 36.8 | 21.8 | 20.8 |
| G.711 A-law | 23.9 | 23.1 | 25.0 |
| G.711 -law | 24.5 | 25.0 | 28.1 |
| GSM | 35.8 | 22.3 | 24.1 |
| System | Dim | 2019 LA eval | 2021 LA | ITW | |
|---|---|---|---|---|---|
| EER | t-DCF | EER | EER | ||
| AASIST ‡ [ 10 ] | raw | 1.73 | 0.056 | 30.7 | 37.6 |
| RawNet2 † [ 25 ] | raw | 5.13 | 0.118 | — | — |
| LFCC-GMM † [ 27 ] | 60 | 8.09 | 0.212 | — | — |
| All-36 SVM | 36 | 16.2 | 0.337 | 23.5 | 50.0 |
| F1-SVM | 4 | 19.8 | 0.388 | 22.9 | 34.7 |
Audio Deepfake Detection
Audio deepfake detection identifies synthetic, cloned, or manipulated speech and other audio to support authentication and media forensics.
See all 210 papers