Exploration of Compressed ILPR Features for Replay Attack Detection

Sarfaraz Jelil, Sishir Kalita, S R Mahadeva Prasanna, Rohit Sinha

This paper deals with the problem of detecting replay attacks on speaker verification systems. In literature, apart from the acoustic features, source features have also been successfully used for this task. In existing source features, only the information around glottal closure instants (GCIs) have been utilized. We hypothesize that the feature derived by capturing the temporal dynamics between two GCIs would be more discriminative for such task. Motivated by that, in this work we explore the use of discrete cosine transform compressed integrated linear prediction residual (ILPR) features for discriminating between genuine and replayed signals. A spoof detection system is built using the compressed ILPR feature and a Gaussian mixture model (GMM) classifier. A baseline system is also built using constant-Q cepstral coefficient feature with GMM back-end. These systems are tested on the ASVSpoof 2017 Version 2.0 database. On fusing the systems developed using acoustic and proposed source features an equal error rate of 9.41% is achieved on the evaluation set.

 DOI: 10.21437/Interspeech.2018-1297

Cite as: Jelil, S., Kalita, S., Prasanna, S.R.M., Sinha, R. (2018) Exploration of Compressed ILPR Features for Replay Attack Detection. Proc. Interspeech 2018, 631-635, DOI: 10.21437/Interspeech.2018-1297.

  author={Sarfaraz Jelil and Sishir Kalita and S R Mahadeva Prasanna and Rohit Sinha},
  title={Exploration of Compressed ILPR Features for Replay Attack Detection},
  booktitle={Proc. Interspeech 2018},