New Approach to Detect Replay Attack for Speaker Verification System Using High Frequency Features and ELM Based BiLSTM

Abstract

Replay attack is vulnerable to automatic speaker verification system, where the frauds get the access by replaying the pre-recorded speech utterances of the genuine speakers. In this proposed work, we mainly concentrated on high frequency band and classification part. This paper shows the importance of higher frequency band (6 kHz to 8 kHz). The huge difference between genuine and spoofed speech spectrum is also explained which is caused due to imperfection occurred by using multiple anti-aliasing filters. Alongside, Constant-Q Cepstral Coefficients (CQCC) technique is used to extract magnitude discrimination power features set to detect the replayed spoof attack for speaker verification. Further the ELM based BiLSTM is proposed to improve the system performance. The proposed framework shows better results of Equal Error Rate (EER) to 05.26% for development set and 8.44% for evaluation set.