Novel Variable Length Energy Separation Algorithm Using Instantaneous Amplitude Features for Replay Detection

Voice-based speaker authentication or Automatic Speaker Verification (ASV) system is now becoming practical reality after several decades of research. However, still this technology is very much susceptible to various spoofing attacks. Among various spoofing attacks, replay is the most challenging attack. In this paper, we propose a novel feature set based on our recently introduced Variable length Energy Separation Algorithm (VESA) during INTERSPEECH 2017. The key idea of this paper is to capture the Instantaneous Amplitude (IA) obtained from the instantaneous energy fluctuations. The replay speech is affected by acoustic environment and distortions of intermediate device. Thus, the noise added in replayed speech is important to detect. The Amplitude Modulations (AM) are more susceptible to noise and multipath interferences that may result due to replay mechanism. The experiments are performed on various dependency index (DI) and lower EER of 6.12% and 11.94% is found on dev and eval set, respectively, of ASV Spoof 2017 Challenge database. Furthermore, we compare our results with CQCC, LFCC, MFCC and VESA-IFCC feature sets. The score-level fusion VESA-IFCC and proposed feature set further reduced the EER to 0.19% and 7.11% on dev and eval set, respectively.