Contents
Yes, and often sharply. Ordinary compression, the MP3 or phone codec a clip has already passed through by the time you hear it, strips the fine, mostly high-frequency detail a voice detector relies on. It degrades detection in both directions: an AI clip can slip through as real, and a genuine recording can flag as AI. The more a clip has been re-encoded, forwarded, or screen-recorded, the less a detector score means.
Why does compression hurt a voice detector so much?
Because the evidence lives exactly where lossy codecs cut. When a model synthesizes speech, the vocoder stage leaves traces that concentrate in the higher frequencies: Frank and Schönherr (2021), analyzing six generator architectures for the WaveFake dataset, found the generators deviate from real audio “specifically among the higher frequencies.” A lossy codec like MP3 works by discarding detail the ear is least likely to miss, much of it in that same high-frequency band. So compression and detection are pulling on the same thread, and compression pulls harder: it can erase the vocoder fingerprint before the detector ever sees it; how a detector reads that fingerprint is set out in how do AI voice detectors work. A detector reading a heavily compressed clip is often working from the acoustic equivalent of a low-resolution copy.
How much accuracy is actually lost?
Enough to matter. Negroni, Cuccovillo, Bestagini and colleagues (ICASSP 2026) measured how far a detector’s equal error rate, the point where false accepts and false rejects balance, rises under different codecs:
| Codec | Detector equal error rate |
|---|---|
| M4A | 21.8% |
| EnCodec | 29.2% |
| MP3 | 40.9% |
MP3, the most common web and phone codec, is the worst, pushing error above 40 percent. A number that high is close to a coin toss, which means a “human” verdict on a compressed MP3 is weak evidence that the voice is real. And this sits inside a wider collapse: tested on real audio collected outside the lab, detectors that read in the low single digits on a benchmark climb into the 35 to 55 percent range (Yi, Wang, Tao et al., 2023), with compression one of the conditions doing the damage alongside unfamiliar generators.
Does it only make fakes harder to catch?
No, and this is the part people miss. Compression fails in both directions. A compressed AI clip can pass as real because its fingerprint was stripped, and a compressed real clip can flag as AI because the detector is reading the wrong thing. Negroni and colleagues found that many models “base their decisions on unvoiced timesteps, frequency bands where no speech occurs, or even background noise, rather than the actual speech,” so a detector keying on compression artifacts rather than the voice can push a genuine recording toward a false “AI” label. A low score and a high score on a heavily compressed clip are both unreliable, for the same underlying reason.
Do some detectors survive compression better than others?
Somewhat. The tools that lean hardest on fine handcrafted spectral detail lose the most, because that detail is exactly what compression removes, while detectors built on large self-supervised speech models hold up better. Frank and Schönherr (2021) noticed an early version of this, reporting that “more traditional models proved to be more robust” than some neural ones in their tests, and later work confirms that learned representations degrade less under compression than handcrafted cepstral features. Training a detector on compressed audio also helps it cope with compressed audio. But “better” is not “immune”: every family loses ground, and a codec has usually already touched the clip before any of them see it.
What should you do about it?
Test the original file, not a re-encode. Whenever you can, run the untouched upload or voice note rather than a re-shared copy, a screenshot, or a clip pulled from a forwarded video, because each of those adds another layer of compression. Read a clean “not AI” result on a heavily compressed clip as carrying little information rather than as a clearance, the same way a re-saved screenshot defeats an image detector. And corroborate: run a second method, look for provenance where it exists, and weigh the source, rather than leaning on one score from degraded audio, the first-pass procedure set out in how to tell if audio is AI-generated.
Sources
- Frank, Schönherr (2021). WaveFake: A Data Set to Facilitate Audio Deepfake Detection. NeurIPS 2021 Datasets and Benchmarks.
- Negroni, Cuccovillo, Bestagini et al. (2026). Multi-Task Transformer for Explainable Speech Deepfake Detection via Formant Modeling. ICASSP 2026.
- Yi, Wang, Tao et al. (2023). Audio Deepfake Detection: A Survey.