Audio Spectrogram Transformer for Synthetic Speech Detection via Speech Formant Analysis

In this paper, we address the challenge of synthetic speech detection, which has become increasingly important due to the latest advancements in text-to-speech and voice conversion technologies. We propose a novel multi-task neural network architecture, designed to be interpretable and specifically tailored for audio signals. The architecture includes a feature bottleneck, used to autoencode the input spectrogram, predict the fundamental frequency (f0) trajectory, and classify the speech as synthetic or natural. Hence, the synthesis detection can be considered a byproduct of attending to the energy distribution among vocal formants, providing a clear understanding of which characteristics of the input signal influence the final outcome. Our evaluation on the ASVspoof 2019 LA partition indicates better performance than the current state of the art, with an AUC score of 0.900.