Which Acoustic Features Survive Compression? Tracker-Estimated First-Formant Statistics for Codec-Robust Audio DeepfakeDetection

Authors: Yash Sukhdeve, Aaron Storey, Masudul H. Imtiaz

Published: 2026-11-09 21:28:48+00:00

AI Summary

This paper investigates which low-dimensional acoustic features remain informative for audio deepfake detection despite codec degradation, a common issue in real-world scenarios. By screening 36 utterance-level features on ASVspoof datasets across various codec conditions, the authors found that only first-formant (F1) measures demonstrate practical separation and remarkable stability against codec-induced degradation, outperforming a full 36-feature SVM in robustness.

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.


Key findings
Only four first-formant (F1) measures showed large practical separation and significant codec stability, with an EER degradation of only 5.9 pp under codec mismatch compared to 14 pp for the full 36-feature set. These F1 features, highly redundant and representing a single latent dimension, offer a compact, codec-stable, and physically interpretable diagnostic signal for deepfake detection. The F1-SVM demonstrated better generalization on unseen data (In-the-Wild corpus) compared to the all-36 model and even a neural baseline (AASIST).
Approach
The authors screen 36 utterance-level acoustic features (formants, PNCCs, LFCCs, prosody) on the ASVspoof 2019 training data and test their codec robustness on the ASVspoof 2021 LA evaluation set across seven codec conditions. They identify features with large practical effect size using rank-biserial correlation and bootstrap CIs. An RBF-SVM is then used to compare the detection performance and codec stability of a four-feature F1 subset against the full 36-feature set.
Datasets
ASVspoof 2019 LA, ASVspoof 2021 LA, In-the-Wild corpus
Model(s)
RBF-SVM, Logistic Regression, MLP, Random Forest, Decision Tree (for comparison)
Author countries
USA