Robust end-to-end synthetic speech detection with deep neural networks and masking

This thesis proposes a method to improve the robustness of i-vectors for synthetic speech detection under noisy conditions. I-vectors are fixed-length representations commonly used in speaker recognition systems. However, their performance degrades with noise. In order protect the system, using a convolutional neural network (CNN) to generate a noise mask is proposed. This mask suppresses unreliable regions in the speech spectrogram corrupted by noise. The masked spectrogram is then used to extract more robust i-vectors. Experiments use the ASVspoof 2015 dataset with added babble, white, and car noise. The CNN is trained to estimate the signal-to-noise ratio in each spectrogram frame. This generates the noise mask that is applied before i-vector extraction. Results show the proposed masking approach reduces equal error rates by over 50% compared to standard i-vectors from noisy speech. However, performance degrades on car noise which was not seen during CNN training. This highlights the need for more diverse training noise types. In conclusion, the proposed spectrogram masking technique using a CNN can increase robustness of i-vectors for synthetic speech detection in noisy conditions. The noise mask helps suppress unreliable regions to provide improved anti-spoofing performance. However, the mask does not generalize well to unseen noise types. Overall, the study shows potential for deep learning-based masking to improve security of speaker recognition systems against spoofing attacks under noise. But more research is needed into handling diverse noise conditions.