Agha Ali Raza

Publications/Speech Processing, Recognition & Audio Security

Conference · 2021

Fake Audio Detection in Resource-Constrained Settings using Microfeatures

Hira Dhamyal, Ayesha Ali, Ihsan Ayyub Qazi, Agha Ali Raza

Interspeech 2021, Brno, Czechia · CORE A

Summary

Detecting fake audio usually takes heavy neural networks, which is impractical on phones and embedded devices and for low-resource languages. We study two tiny, physically meaningful cues: voicing onset time (the brief delay before the vocal cords start vibrating after certain consonants) and coarticulation (how neighboring sounds blend into each other). On the ASVSpoof2019 dataset, synthesized speech shows longer and more variable voicing onset times and less vowel movement than real speech. Combined, these microfeatures detect fakes with a 23.5% equal error rate, making them useful for rapid pre-screening and forensics and as extra inputs to larger detectors. This is the first study of these cues for fake audio detection.

Cite this paper

Hira Dhamyal, Ayesha Ali, Ihsan Ayyub Qazi, Agha Ali Raza. Fake Audio Detection in Resource-Constrained Settings using Microfeatures. Interspeech 2021, Brno, Czechia.

@inproceedings{dhamyal2021fake,
  author = {Hira Dhamyal and Ayesha Ali and Ihsan Ayyub Qazi and Agha Ali Raza},
  title = {{Fake Audio Detection in Resource-Constrained Settings using Microfeatures}},
  booktitle = {Interspeech 2021, Brno, Czechia},
  year = {2021},
  url = {https://aghaaliraza.com/publications/2021--INTERSPEECH--Fake_Audio_Detection_in_Resource-constrained_Settings_using_Microfeatures.pdf}
}

More on this theme

All 12 works on this theme →