A Comparison of CQT Spectrogram with STFT-based Acoustic Features in Deep Learning-based Synthetic Speech Detection

Abstract

Automatic Speaker Verification (ASV) systems have proven to bevulnerable to various types of presentation attacks, among whichLogical Access attacks are manufactured using voiceconversion and text-to-speech methods. In recent years, there has beenloads of work concentrating on synthetic speech detection, and with the arrival of deep learning-based methods and their success in various computer science fields, they have been a prevailing tool for this very task too. Most of the deep neural network-based techniques forsynthetic speech detection have employed the acoustic features basedon Short-Term Fourier Transform (STFT), which are extracted from theraw audio signal. However, lately, it has been discovered that the usageof Constant Q Transform's (CQT) spectrogram can be a beneficialasset both for performance improvement and processing power andtime reduction of a deep learning-based synthetic speech detection. In this work, we compare the usage of the CQT spectrogram and some most utilized STFT-based acoustic features. As lateral objectives, we consider improving the model's performance as much as we can using methods such as self-attention and one-class learning. Also, short-duration synthetic speech detection has been one of the lateral goals too. Finally, we see that the CQT spectrogram-based model not only outperforms the STFT-based acoustic feature extraction methods but also reduces the processing time and resources for detecting genuine speech from fake. Also, the CQT spectrogram-based model places wellamong the best works done on the LA subset of the ASVspoof 2019 dataset, especially in terms of Equal Error Rate.

Keywords

• voice spoofing detection
• deep neural networks
• voice biometrics
• deepfake audio detection
• synthetic speech detection