Abstract: In this paper, we present a neutral network based adversarial method to generate deepfake audios to fool anti-spoofing systems built in various ways. Firstly, we build an any-to-many voice conversion (VC) system to convert source speech with arbitrary language content into target speaker’s fake speech. Then the converted speech generated from VC is post-processed in time-domain to improve the deception ability. The experimental results show that our system has adversarial ability against anti-spoofing detectors with a little compromise in audio quality and speaker similarity. This system ranks top in track 3.1 in the ADD 2022 challenge, showing that our method could also gain perfect generalization ability against different detectors.
For more information, refer to the paper "Time Domain Adversarial Voice Conversion for ADD 2022 Challenge"
Audio samples
This section displays the audios generated by our system. Here, The reference audios, speech samples generated by our voice conversion model and deepfake audios modified by our adversarial post-processing system of 10 target speakers are listed. All speakers are selected from the AIShell-3.