Baraa Said العربية Message

Research · With Mohammadkhair Awwad · Birzeit

Can classical speech features still catch AI-generated voices?

Research testing whether classical speech features can still catch AI-generated voices. Across 69 detection systems, generalisation to unseen attacks and codecs, not the choice of features, turned out to be the real bottleneck.

All work
Three spectrograms from the study: genuine speech, a detected fake and a fake the detector missed
Genuine speech, a caught fake and a missed fake
detection systems compared
69
EER of a log-mel LCNN on 13 unseen attacks
5.95%
silence and duration statistics that topped the cross-corpus test
5

The question

Cloning a voice is cheap now. Can the classic hand-crafted speech features (MFCC, LFCC, CQCC) with simple classifiers still tell a real voice from a synthetic one, or does detection need deep learning?

How we tested it

With Mohammadkhair Awwad at Birzeit University, we trained detectors on the ASVspoof 2019 logical-access data and compared 69 systems: hand-crafted feature sets with classical classifiers such as SVMs, and a log-mel LCNN. Each one was tested on attacks it had never seen, on audio passed through codecs it had never heard, and on a different corpus altogether.

What we found

  • The log-mel LCNN reached 5.95% equal error rate on thirteen unseen attacks.
  • On the cross-corpus test, five simple silence and duration statistics were the best of all 69 systems. Detectors had learned shortcuts in the training data rather than what makes a voice synthetic.
  • The real bottleneck is generalisation, not the choice of features.

Equal error rate (EER) is the point where a detector wrongly rejects as many real voices as it wrongly accepts fake ones; lower is better.

Questions about Catching AI-generated voices

Can AI-generated voices be detected?

In this study, detectors did well on attacks similar to their training data: a log-mel LCNN reached 5.95% EER on thirteen unseen attacks. But on a different corpus, five simple silence and duration statistics beat all 69 systems, showing that generalisation is the real problem.

What is equal error rate (EER)?

The operating point where a detector's false-rejection rate equals its false-acceptance rate. A lower EER means a better detector.

Who worked on this research?

Baraa Said and Mohammadkhair Awwad at Birzeit University.

More work

Questions?

Hiring for an AI or full-stack role, or need a system built? One message is enough.