Research audio companion

Adversarial Text-to-Speech Synthesis Against Forensic Deepfake Detection

The same words. The same voice.
Three ways to shape the signal.

Jiayan Cui, Davide Salvi, Paolo Bestagini, Stefano Tubaro

Politecnico di Milano

Hear the comparison

A closer listen.

Choose a voice and a prompt.
Switch vocoders to compare the details.

Loading the audio collection…

Headphones recommended. These are the recordings from the original audio collection; no audio processing is applied during playback.

Behind the sound

A vocoder trained
with a detector in mind.

The study explores adversarial fine-tuning of a BigVGAN vocoder: retaining speech synthesis objectives while adding feedback from forensic deepfake detectors.

Compare how the training objective changes across the three variants. AASIST is held out of vocoder fine-tuning to evaluate transfer to another detector.

Research code is being prepared for public release.

V

Baseline vocoder

Speech synthesis training without a detector-specific adversarial objective.

Detector target: none
VR

A single detector

Adversarial fine-tuning against RawNet2 in a white-box setting.

Detector target: RawNet2
VRT

Two detectors together

Adversarial fine-tuning against RawNet2 and TSSD jointly.

Detector targets: RawNet2 + TSSD

About this collection

Four training-set speakers and two unseen voices are included. Gorou and Yae Miko are character voices from Genshin Impact by miHoYo, as identified in the original collection. Reference clips demonstrate the speaker’s voice and contain different text from the synthesized prompts.

Prompt excerpts are reconstructed from the source filenames. Listening examples illustrate the paper’s methods; detector performance is evaluated separately in the study.

Original audio collection