Effectiveness of Speech Demodulation-Based Features for Replay Detection

Replay attack presents a great threat to Automatic Speaker Verification (ASV) system. The speech can be modeled as amplitude and frequency modulated (AM-FM) signals. In this paper, we explore speech demodulation-based features using Hilbert transform (HT) and Teager Energy Operator (TEO) for replay detection. In particular, we propose features, namely, FIT-based Instantaneous Amplitude (IA) and Instantaneous Frequency (IF) Cosine Coefficients (i.e., HT-IACC and HT-IFCC) and Energy Separation Algorithm (ESA)-based features (i.e., ESA-IACC and ESA-IFCC). For adapting instantaneous energy w.r.t given sampling frequency, ESA requires 3 samples whereas FIT requires relatively large number of samples and thus, ESA gives high time resolution.The experiments were performed on ASV spoof 2017 Challenge database for replay spoof speech detection (SSD).The experimental results shows that ESA-based features gave lower EER. In addition, linearly spaced Gabor filterbank gave lower EER than Butterworth filterbank. To explore possible complementary information using amplitude and frequency, we have used score-level fusion of IA and IF. With HT-based feature set, the score-level fusion gave EER of 5.24 % (dev) and 10.03 % (eval), whereas ESA-based feature set reduced the EER to 2.01 % (dev) and 9.64 % (eval).