Recent advances in speech synthesis and voice cloning have increased the need for reliable synthetic-speech detection, yet high-accuracy detectors increasingly rely on large pretrained models that are costly to invoke on every recording. Rather than replacing such detectors, we investigate a compact front-end screen that processes all inputs cheaply and forwards only suspicious recordings for more expensive analysis. To enable lightweight screening without a large learned encoder, we exploit spectral traces introduced by speech-generation operations. We analyze how learned upsampling and inverse short-time Fourier transform synthesis can produce predictable spectral artifacts and measure their presence directly in generated waveforms. Because the strength of these artifacts varies across generators, we combine decoder-guided spectral measurements with complementary descriptors of short-time spectral shape and temporal variation in a compact gradient-boosted tree. Across seven speech generators and two human-speech sources, the proposed screen achieves an equal error rate of 0.021% with an estimated model storage of 151 KiB. When used as the first stage of a simulated cascade with a 1.15-billion-parameter detector, it reduces estimated detection energy by 84.4% while operating at a 0.050% synthetic-speech miss rate, demonstrating the potential of decoder-guided acoustic evidence for low-cost front-end screening.
Figures & tables
Figure 1: Upsampling-layer spectra of (a) the OpenVoice V2 tone-color converter and (b) the CosyVoice2 HiFT decoder; crosses mark predicted DC-replica frequencies. (c) Spectral peaks relative to the local background at 6 and 10 kHz across 200 CosyVoice2 recordings.
Figure 2: (a) Background subtraction at 5.25 kHz in a CosyVoice2 recording; the envelope is the background curve estimated from local spectral minima. (b) Mean normalized spectral residuals from 200 LibriSeVoc (Human) and 200 CosyVoice2 (TTS) recordings.
Seen-Source
Cross-Source
System
Size
EER
Libri
Chatt.
Cosy2
Qwen3
Vibe
[ 13 ]
[ 18 ]
[ 4 ]
[ 7 ]
[ 14 ]
RawGAT-ST [ 23 ]
1.68 MiB
0.023
86.1
100.0
100.0
100.0
100.0
AASIST [ 10 ]
1.14 MiB
0.021
90.7
100.0
100.0
100.0
100.0
RawTFNet [ 27 ]
694 KiB
0.046
71.8
100.0
100.0
100.0
100.0
W-MesoNet [ 11 ]
29.2 MiB
0.065
84.8
96.5
97.0
94.0
100.0
Table 1: Model storage and detection performance: seen-source EER (%) and accuracy (%) at frozen seen-source thresholds on unseen human speech and four additional TTS systems.
Held out
CosyV.
Tort.
Fish
Parler
StyleT2
OpenV.
Melo
Test clips
8,309
10,388
2,853
1,944
2,024
7,211
10,388
EER (%)
0.014
0.288
1.046
3.245
3.845
5.065
6.747
Table 2: Leave-one-generator-out EER (%) for proposed detector.
FNR limit (%)
0.000
0.010
0.020
0.050
FPR (%)
0.359
0.058
0.021
0.009
Table 3: Screening operating points selected from the held-out test ROC under the specified empirical FNR limits.
Synthetic speech detection benchmarks now report sub-1% error rates on some in-domain evaluations, yet performance degrades under unseen attacks, channel mismatch, and distribution shift. Based on a three-year effort with Phonexia, a commercial speaker-recognition vendor, we report barriers encountered while building and deploying a detector. Many public benchmarks are not licensed for commercial model development. Real inputs are not four-second clean clips but long, codec-degraded, sometimes partially synthetic recordings. And when a calibrated system returns a log-likelihood ratio of 2.5, no one can tell the customer what it means for their decision. Rather than proposing a new model, we connect these barriers to concrete research and coordination proposals: shared standards for commercially usable datasets, realistic deployment benchmarks, and scores that non-experts can act on. These observations come from one project and should be tested in other settings.
Provenance watermarking is increasingly treated as a safeguard for synthetic speech, whether built directly into speech-generation models such as Chatterbox, provided through dedicated techniques such as AudioSeal, or deployed by commercial platforms such as ElevenLabs. We identify a previously uncharacterized liability: when synthetic speech is watermarked and human speech is not, detectors trained alongside latch onto the watermark as a spurious "watermark => fake" shortcut. This single feature yields three coupled failures: generalization degradation (model performance deteriorates on unseen data), strip-to-evade (a watermarked fake escapes once unwatermarked), and mark-to-frame (watermarking a real voice flags it as fake). In a controlled white-box experiment, a watermark-trained detector shows all three (for example, mark-to-frame lifts Equal Error Rate from 16% to 75%). In a black-box test of a commercial API, we show that adding a watermark to real speech disguises it as fake. However, this shortcut is fixable: retraining with the watermark on both classes decorrelates it and restores clean behavior. We release experiment data as a paired clean-versus-watermarked corpus (WASP).
Recent advancements in text-to-speech technologies enable generating high-fidelity synthetic speech nearly indistinguishable from real human voices. While recent studies show the efficacy of self-supervised learning-based speech encoders for deepfake detection, these models struggle to generalize across unseen speakers. Our quantitative analysis suggests these encoder representations are substantially influenced by speaker information, causing detectors to exploit speaker-specific correlations rather than artifact-related cues. We call this phenomenon speaker entanglement. To mitigate this reliance, we introduce SNAP, a speaker-nulling framework. We estimate a speaker subspace and apply orthogonal projection to suppress speaker-dependent components, isolating synthesis artifacts within the residual features. By reducing speaker entanglement, SNAP encourages detectors to focus on artifact-related patterns, leading to state-of-the-art performance.