In the scientific conference paper "Calibrating neural networks for synthetic speech detection: A likelihood-ratio-based approach", the three scientists from the Fraunhofer IDMT in Ilmenau presented a method with which uncalibrated results from synthetic speech recognition can be specified in so-called likelihood ratios.

Neural networks for synthetic speech detection are normally uncalibrated. This means that the score expressing whether a speech is synthetic or not varies widely according to the file's quality, the amount of times it has been re-uploaded on social networks, the presence of background music, etc.

Now, the method presented by Cuccovillo, Aichroth, and Köllmer for detecting synthesized speech outputs the result as a likelihood ratio. This indicates not only whether the speech appears to be synthesized or not, but also the error rate. This error measure is also used in forensic laboratories for DNA analyses, for example.

The method was developed as part of the news-polygraph and speechtrust+ projects at the Fraunhofer IDMT and combines the institute's expertise in the research areas of media forensics and trustworthy Artificial Intelligence.

This method will soon support forensic laboratories and fact-checkers in reliably detecting audio deepfakes.

Congratulations to our colleagues and many thanks to the AES organizing committee!