InterSpeech 2021

Fake audio generation has undergone remarkable improvement with the advancement in deep neural network models. This has made it increasingly important to develop lightweight yet robust mechanisms for detecting fake audios, especially for resource-constrained settings such as on edge devices and embedded controllers as well as with low-resource languages. In this paper, we analyze two microfeatures: Voicing Onset Time (VOT) and coarticulation, to classify bonafide and synthesized audios. Using the ASVSpoof2019 LA dataset, we find that on average, VOT is higher in synthesized speech compared to bonafide speech and exhibits higher variance for multiple occurrences of the same stop consonants. Further, we observe that vowels in CVC form in bonafide speech have greater F1/F2 movement compared to similarly constrained vowels in synthesized speech. We also analyse the predictive power of VOT and coarticulation for detecting bonafide and synthesized speech and achieve equal error rates of 25.2% using VOT, 39.3% using coarticulation, and 23.5% using a fusion of both models. This is the first study analysing VOT and coarticulation as features for fake audio detection. We suggest these microfeatures as standalone features for speaker-dependent forensics, voice-biometrics, and for rapid pre-screening of suspicious audios, and as additional features in bigger feature sets for computationally intensive classifiers.