Recent audio deepfake detectors separate bona fide speech from synthetic speech, yet it remains unclear which stage of a text-to-speech system supplies the detection evidence. We address this with controlled resynthesis and detector adaptation in an F5-TTS-BigVGAN pipeline. Since vocoder reconstruction of a real mel can itself be separable from the source utterance, we fix the vocoder and trace the larger change in detector separation to acoustic generation. Adversarially fine-tuning the acoustic model, with no detector in its objective, raises EER against fixed detectors at comparable quality. However, adapting a detector only on the tuned model's VCTK outputs lowers its LibriSpeech EER from 19.42% to 7.46% and improves detection of unseen base F5-TTS outputs. These results suggest that acoustic-model updates can reduce the detection evidence available to fixed detectors, while detector adaptation keeps the updated outputs detectable in this pipeline.
Figures & tables
Figure 1: Isolating which TTS stage supplies the detection evidence, with the source corpus and channel condition fixed.
Figure 2: Overview of the experimental framework, with conditions defined in Sec. 2.1 . (a) La and Lf denote the adversarial and feature matching losses. When R itself is evaluated, the original source serves as bona fide. (b) XLS-R SLS is adapted on bona fide utterances paired with their tuned F5-TTS resyntheses.
Corpus
Cond.
SLS
Mamba
AASIST-L
LibriSpeech
R
45.25
47.88
49.88
V
38.00
38.63
50.63
T
15.46 ±0.44
12.38 ±0.78
28.75 ±1.11
TG
17.50 ±0.45
13.46 ±0.69
40.88 ±0.99
FT
19.42 ±1.39
16.04 ±0.81
47.67 ±1.18
VCTK
R
51.26
49.93
49.93
Table 1: EER of fixed detectors per condition. For V, T, TG, and FT, the bona fide trials are 24 kHz resampled sources.
Vocoder
LibriSpeech
VCTK
BigVGAN
38.00
45.94
Vocos
12.00
22.01
Table 2: SLS EER for real-mel reconstruction (V) with the two vocoder configurations the F5-TTS checkpoint supports.
Corpus
Cond.
SLS EER
UTMOS
Spk. sim.
WER
LibriSpeech
T
15.46 ±0.44
3.38 ±0.02
0.724 ±0.002
2.98 ±0.12
FT
19.42 ±1.39
3.41 ±0.01
0.725 ±0.001
2.95 ±0.03
VCTK
T
3.45 ±0.19
3.37 ±0.03
0.754 ±0.002
0.73 ±0.02
FT
4.09 ±0.17
3.35 ±0.03
0.763 ±0.004
0.68 ±0.02
VoxPopuli
T
35.33 ±1.49
3.35 ±0.04
0.747 ±0.000
1.41 ±0.07
FT
40.75 ±0.99
3.45 ±0.03
0.742 ±0.004
1.46 ±0.08
Table 3: SLS EER alongside quality and content measures for T and FT. Higher UTMOS and speaker similarity and lower WER (%) are better.
Eval. set
Cond.
Baseline
LS adapt.
VC adapt.
LibriSpeech
R
45.25
49.50 ±0.38
46.50 ±0.33
V
38.00
40.79 ±0.64
35.46 ±0.88
T
15.46 ±0.44
1.42 ±0.44
5.92 ±0.47
TG
17.50 ±0.45
1.75 ±0.57
6.29 ±0.07
FT
19.42 ±1.39
1.71 ±0.26
7.46 ±0.31
VCTK
R
51.26
50.02 ±0.04
50.12 ±0.21
Table 4: SLS adaptation: EER (lower is better for T/TG/FT). LS/VC: balanced LibriSpeech and VCTK adaptation (3 training seeds). Baseline T/TG/FT summarize generation seeds. ∗ Released checkpoint [ 22 ]
Modern text-to-speech (TTS) systems are rarely deployed as unchanged pre-trained models. They are often adapted through supervised fine-tuning (SFT) or preference optimization such as DPO and GRPO. This raises a practical question for audio deepfake forensics: do fingerprints learned from a foundation generator remain valid after adaptation? We present GenTraceBench, a controlled benchmark spanning five TTS architectures, 16 pre-/post-training variants, and 49,728 utterances generated with fixed texts and speaker prompts. Under a train-on-foundation, test-on-adapted protocol, we evaluate binary detection, closed-set attribution, and open-set verification. DPO and GRPO generally preserve fingerprints, whereas some SFT and pre-training-data changes cause substantial drift; effect sizes vary across three forensic backbones. Repeated training runs confirm the largest W2V-BERT attribution drop, while a data-mixture control with comparable speech quality shows that composition change need not cause drift. In W2V-BERT verification, multi-shot enrollment reduces EER for the SFT condition from 44.4% to 11.0%, whereas the SingNet-only condition remains at or above 45% EER.
Li Wang, Kunyu Feng, Wan Lin +7
The Chinese University of Hong Kong, Shenzhen, China · Huawei Technologies Co., Ltd., China · Amphion Technology Co., Ltd., China
Advances in speech synthesis have made deepfake audio highly realistic. Earlier studies reported 70-80% human detection accuracy, but relied primarily on older synthesizers. We compare human detection for three selected voice synthesis tools released in 2019, 2022, and 2024 with 82 IT professionals, and benchmark humans against six pretrained detectors on the same material. For fully synthetic speech (full spoofs), the F1 score drops from about 90% for RTVC and YourTTS to 48% for ElevenLabs, although listeners were explicitly warned that deepfakes were present. For partial spoofing, where only one sentence of an utterance is altered, strict accuracy falls to 9%, and listeners classify the synthetic sentence as bona fide 77% of the time. Humans and detectors fail in complementary ways, and neither reliably localizes short manipulations. Additionally, listeners increasingly mislabel bona fide speech as fake, eroding trust in unmanipulated audio. These findings show that human perception alone is unreliable for the selected modern and partial-spoof conditions and motivate procedural verification, provenance, watermarking, and segment-level detection.
Audio deepfakes generated by neural text-to-speech and voice-cloning systems threaten speaker verification and public discourse at scale. The core challenge is cross-dataset generalization: detectors trained on one synthesis pipeline collapse on unseen forgeries. We argue that this failure is primarily because of structural synthetic speech artifacts which are multi-timescale trajectory anomalies. Though every existing detector aggregates a fixed-window frame statistics, this misaligns the architecture with the signal. We propose FlowFake, a Liquid Time-Constant (LTC) architecture whose hidden state evolves via a learned ODE, with per-neuron adaptive time constants simultaneously resolving spectral (10ms) and prosodic (2s) cues. At only 34K parameters FlowFake achieves formal BIBO stability and O(dt^4) integration error. On a four-dataset cross domain benchmark (ASVspoof2019-LA, FakeOrReal, InTheWild, MLAAD), FlowFake reaches 75.29% on ASVspoof2019 trained only on FakeOrReal and 79.97% trained only on MLAAD. It outperforms RawGAT-ST and Whisper-DF on every evaluated pair and matching SSL Wav2vec2 (300x larger) at 0.01% of its parameter count. The source code is available on : https://github.com/GhostRider2023/FlowFake