Noise Robust Audio Spoof Detection Using Hybrid Feature Extraction and LCNN

In recent times, Voice Controlled Devices (VCD) have become an important part of our life. However, these devices are vulnerable to replay attacks. Hence, to address the issue, this paper proposes an Automatic Speaker Verification (ASV) system that uses hybrid feature extraction technique at front end and Light Convolutional Neural Network (LCNN) at back end. The proposed hybrid feature extraction techniques involve three different combinations that are; Frequency Domain Linear Prediction (FDLP) with Gammatone Cepstral Coefficients (GTCC), FDLP with Mel Frequency Cepstral Coefficients (MFCC) and GTCC with MFCC. Hence, the work proposes FDLP + GTCC-LCNN, FDLP + MFCC-LCNN and GTCC + MFCC-LCNN based ASV systems. The training and evaluation of all the systems is performed on state-of-the-art Voice Spoofing Detection Corpus (VSDC) that contains original audio samples (0PR), single order replay (1PR) recorded audios, and multi order replay (2PR) recorded audio samples. The obtained results reveal that the proposed FDLP + GTCC-LCNN based system outperforms the other two proposed systems. It gives 0.43 and 0.52% Equal Error Rate (ERR) in 1PR and 2PR attacks, respectively. In order to check the performance of proposed system in noisy environment, Additive White Gaussian Noise (AWGN), Babble and Street Noise are mixed with the audio samples of VSDC dataset at 5, 10 and 15 dB. It can be concluded from attained results that the proposed system gives good results under both clean and noisy multi order replay attack environments.