Integration of Timbrel, Cepstral Domain and Linear Prediction-Based Features for Replay Attack Detection
Amol A. Chaudhari, D. K. Shedge, Vinayak K. Bairagi · International Journal of Electrical and Electronics Engineering · 2023
The automatic speaker verification system is vulnerable to several spoofing attacks. Among these spoofing attacks, detecting replay attacks is challenging as attackers do not need any expertise to mount replay attacks. Many efforts from the research community have focused on anti-spoofing solutions against the reply attack. Such efforts are classified as one focusing on feature extraction and others concentrating on classifiers. This work evaluates the performance of feature extraction schemes CQCC, LFCC, and MFCC. The success of Linear Prediction analysis has been demonstrated in the past. This work evaluates the performance of LPC and LPCC features. The recent work in the literature has focused on using multiple features and combining these features for improved performance. In this work, numerous components of CQCC, MFCC, LFCC, LPC and LPCC are integrated considering various combinations and evaluated. In literature, the success of Timbrel features has been demonstrated for speaker identification. The feature vector formed using various Timbrel features is integrated with cepstral and linear prediction-based features. Finally, Timbrel features zero cross rate are combined with these multiple features. Among all experiments carried out on the ASVspoof 2017 version 2 database, EER 5.44% is achieved for the integration of zero cross rate and LPC on the development set and 17.79% EER is conducted for the integration of zero cross rate, MFCC, CQCC, LFCC, and LPCC features on evaluation set.