Research on voice authenticity
We study how synthetic speech detection holds up outside the lab: on voices, generators and recording conditions it has never seen.
What we work on
Detectors trained on yesterday's speech generators often fail on today's. We study how detection behaves on synthesis systems that were never seen during training.
Disputed audio is rarely a studio recording. We measure how phone lines and common audio codecs change the results, and where they stop being reliable.
Noise, room echo, bandwidth limits and edits, often combined. We test on third-party benchmarks built to include them, because clean benchmarks overstate what works in the field.
How we report results
- Out-of-domain first. We report numbers measured on data and generators that were not used for training, not only in-domain scores.
- Speaker-matched evaluation. Genuine and synthetic samples come from the same speakers, so a result reflects whether the voice is synthetic, not who is speaking or which dataset the file came from.
- We report where it breaks. Conditions that degrade results are reported next to those that do not.
- No absolute verdict. An analysis returns a score, and “uncertain” is a valid outcome. Outside the conditions we have measured, we consider abstaining better than guessing.
Current status
- ICASSP 2027. A full paper on the generalization of audio deepfake detectors to new conditions is under review. Under review
- 30+ provisional patent applications filed. On voice authenticity and voice signal analysis. Filed
We do not publish detection rates as product claims. Research figures depend on the data, the generators and the recording channel, and only mean something next to their protocol.
Academic collaboration
We welcome contact from researchers working on speech anti-spoofing, audio forensics and evaluation methodology.