A Deep Identity Representation for Noise Robust Spoofing Detection

The issue of the spoofing attacks which may affect automatic speaker verification systems (ASVs) has recently received an increased attention, so that a number of countermeasures have been developed for detecting high technology attacks such as speech synthesis and voice conversion. However, the performance of anti-spoofing systems degrades significantly in noisy conditions. To address this issue, we propose a deep learning framework to extract spoofing identity vectors, as well as the use of soft missing-data masks. The proposed feature extraction employs a convolutional neural network (CNN) plus a recurrent neural network (RNN) in order to provide a single deep feature vector per utterance. Thus, the CNN is treated as a convolutional feature extractor that operates at the frame level. On top of the CNN outputs, the RNN is employed to obtain a single spoofing identity representation of the whole utterance. Experimental evaluation is carried out on both a clean and a noisy version of the ASVSpoof2015 corpus. The experimental results show that our proposals clearly outperforms other methods recently proposed such as the popular CQCC+GMM system or other similar deep feature systems for both seen and unseen noisy conditions.