We are the first to identify neural vocoders as a source of features to expose synthetic human voices.
Here are the differences shown by the six vocoders compared to the original audio:
image
We provide LibriSeVoC as a dataset of self-vocoding samples created with six state-of-the-art vocoders to highlight and exploit the vocoder artifacts.
The composition of the data set is shown in the following table:
image
The source of our dataset ground truth comes from LibriTTS. Therefore, we follow the naming logic of LibriTTS. For example, 27_123349_000006_000000.wav, 27 is the reader's ID, and 123349 is the ID of the chapter.
Deepfake Detection
We propose a new approach to detecting synthetic human voices by exposing signal artifacts left by neural vocoders and modifying and improving the RawNet2 baseline by adding multi-loss, lowering the error rate from 6.10% to 4.54% on the ASVspoof Dataset.
This is the framework of the proposed synthesized voice detection method:
