Frequency Domain Linear Prediction Features for Replay Spoofing Attack Detection

Buddhi Wickramasinghe, Saad Irtza, Eliathamby Ambikairajah, Julien Epps


Automatic speaker verification (ASV) systems are vulnerable to various types of spoofing attacks such as speech synthesis, voice conversion and replay attacks. Recent research has highlighted the need for more effective countermeasures for replay attacks, which can be very challenging to detect, however replayed speech has previously shown frequency band-specific differences when compared with genuine speech. In this paper, we propose the use of long-term temporal envelopes of subband signals using a frequency domain linear prediction (FDLP) framework. This flexible framework makes use of temporal envelope information, which has not previously been investigated for replay spoofing detection. Evaluations of the proposed system and its fusion with other subsystems were carried out on the ASVspoof 2017 database. Interestingly, smoother temporal envelopes, based on very long windows of up to 1 second, seem to be most successful and show good prospects for performance improvements via fusion.


 DOI: 10.21437/Interspeech.2018-1574

Cite as: Wickramasinghe, B., Irtza, S., Ambikairajah, E., Epps, J. (2018) Frequency Domain Linear Prediction Features for Replay Spoofing Attack Detection. Proc. Interspeech 2018, 661-665, DOI: 10.21437/Interspeech.2018-1574.


@inproceedings{Wickramasinghe2018,
  author={Buddhi Wickramasinghe and Saad Irtza and Eliathamby Ambikairajah and Julien Epps},
  title={Frequency Domain Linear Prediction Features for Replay Spoofing Attack Detection},
  year=2018,
  booktitle={Proc. Interspeech 2018},
  pages={661--665},
  doi={10.21437/Interspeech.2018-1574},
  url={http://dx.doi.org/10.21437/Interspeech.2018-1574}
}