We found a match
Your institution may have rights to this item. Sign in to continue.
- Title
Implicit processing of linear prediction residual for replay attack detection.
- Authors
Veesa, Suresh; Singh, Madhusudan
- Abstract
This work explores linear prediction (LP) residual based excitation source features for detection of replay speech signals. The LP residual signal is derived from speech using the LP analysis method and mostly contains excitation source information, in implicit form. The LP residual signal is processed through popular mel-frequency cepstral analysis and constant Q cepstral analysis methods implicitly. The obtained excitation source feature representations are referred as residual mel-frequency cepstral coefficients (RMFCC) and residual constant-Q cepstral coefficients (RCQCC). Moreover, the cosine phase of LP residual based analytic signal (i.e., residual phase) is processed using constant Q cepstral analysis. The obtained feature is referred as residual phase constant-Q cepstral coefficients (RPCQCC), representing phase based information. All three LP residual based features along with popular constant-Q cepstral coefficients (CQCC) feature, have been explored in replay detection context. The classical Gaussian mixture model is used as back-end classifier. The ASVspoof 2017 Version 2.0 database is used for conducting replay detection experiments. The standalone EERs for RMFCC, RCQCC and RPCQCC features are 20.89%, 18.51% and 13.82%, respectively. Further, different combinations of features are explored, resulting in an 8.71% EER for the fusion of all three LP residual based source features and the CQCC feature. The proposed method outperforms state-of-the-art methods, demonstrating worthiness of exploring excitation source features for detecting replay attacks.
- Subjects
CEPSTRUM analysis (Mechanics); FORECASTING; GAUSSIAN mixture models; DATABASES
- Publication
International Journal of Speech Technology, 2024, Vol 27, Issue 3, p781
- ISSN
1381-2416
- Publication type
Article
- DOI
10.1007/s10772-024-10125-5