PhonemeFake: A Phonetic DeepFake Dataset with Fine-Grained Manipulations

PhonemeFake Synthesis Pipeline

Bilevel Detection Architecture

Put the dataset into the dataset1 folder and run the pipe.sh.

Below image shows the spectograms of the original (top) and DF (bottom) samples for the example audio which vocalizes the transcript: Government approves new policy to support small businesses amid economic slowdown. The word approves converted to rejects.

Reproducing Script

The reproduce.sh script is used to run evaluations on different datasets using different models and specified versions of the data. This script automates the process of running the evaluation and computing the scores for the models and datasets.

Please refer to the ./reproduce/ folder for more details on benchmark models used, scripts and references.

Usage

To run the script, provide detection model, dataset and version as arguments. For example:

./reproduce.sh <detection model> <dataset> <version>

Arguments

• <detection model>: The detection model to be used. Possible values are conf-3, mamba
• <dataset>: The dataset to be evaluated. Possible values are ASVspoof2021DF, ITW, WaveFake
• <version>: The version of the dataset to be evaluated. Possible values are original or phonemeFake.

Example Commands

• To run the evaluation on the ASVspoof2021DF dataset with the original version using the mamba detection model (RawBMamba):

  ./reproduce.sh mamba ASVspoof2021DF original

• To run the evaluation on the ITW dataset with the phonemeFake version using the conf-3 detection model (SCL Deep Fake Detection's conf-3):

  ./reproduce.sh conf-3 ITW phonemeFake

Results

Dataset Directories

The dataset directories are defined at the beginning of the script, please do not forget to update the directories.