Smoothed Teager Energy Cepstral Feature for Replay Attack Detection on Voice Assistants

Voice Controlled Systems (VCS) are easy and convenient to use for smart applications. However, it brings security issue for replay and hidden voice commands. To that effect, recently developed ReMASC database focuses on replay spoof attacks targeted for different acoustic environments along with single and multi-channel microphone array with different recording devices. This paper presents Spoof Speech Detection (SSD) system using smoothed Teager Energy Cepstral Coefficients (STECC) and Gaussian Mixture Model (GMM) as classifier. The smooth Teager energy profile resolves the spurious peaks and spikes of subband filtered signal. Teager Energy Operator (TEO) also has noise suppression capability that further helps to detect the replay signal. We performed experiments on ReMASC database. The experiments are classified into three tasks, namely, mismatch training, environment-independent, and environment-dependent. In addition, we also focused on the impact of different recording devices with single input channel. For environment-dependent task, STECC feature set performed better than other features. With impact of different devices, proposed STECC outperforms with a difference of 10 % compared to the other systems.