FAD: A Chinese Dataset for Fake Audio Detection

Abstract

Fake audio detection is a growing concern and some relevant datasets have been designed for research. But there is no standard public Chinese dataset under additive noise conditions. In this paper, we aim to fill in the gap and design a Chinese fake audio detection dataset (FAD) for studying more generalized detection methods. Twelve mainstream speech generation techniques are used to generate fake audios. To simulate the real-life scenarios, three noise datasets are selected for noisy adding at five different signal noise ratios. FAD dataset can be used not only for fake audio detection, but also for detecting the algorithms of fake utterances for audio forensics. Baseline results are presented with analysis. The results that show fake audio detection methods with generalization remain challenging. The FAD dataset is publicly available 1 .Recently, fake audio detection is not limited to ASV system, but also starts to focus on real-life scenarios. The first audio deep synthesis detection challenge [12] (ADD 2022) focuses on challenging situations, including low-quality fake audio and partially fake audio detection. More datasets are constrcted by deep-learning speech techniques, such as: FoR [13], WaveFake [14], and HAD [15] datasets.These above-mentioned datasets facilitate the progress of the fake audio detection research. However, in practical applications, audios on social media come in many languages with noisy background and the type of fake audio may be unknown to the model. Those various factors greatly influence the performance of the detection models. The generalization of the detection models is still an urgent need to address. Specifically, the generalization includes generalization to unknown types and robustness to noise and other factors. Most datasets focus on the evaluation of the former aspect, 1 https://zenodo.org/record/6635521#.Ysjq4nZBw2x

Voice Deepfake Detection Using the Self-Supervised Pre-Training Model HuBERT

“…Moreover, the fake audios was generated based on 11 Mandarin TTS systems and 2 Mandarin VC systems, and the duration is randomly set in the range between 2 and 10 s, with the sampling rate of 16 kHz [8]. The FAD dataset consists of 12 types mainstream voice deepfake techniques such as STRAIGHT, LPCNet, and HifiGAN to generate fake audios, and real audios from six different corpora such as AISHELL1, AISHELL3, and THCHS-30 [26]. The evaluation set of the FAD dataset contains 14,000 utterances generated by four unknown deepfake methods that were not included in the training and validation sets, which can better detect the robustness and generalization of the model in the face of unknown attacks.…”

Warning: Humans Cannot Reliably Detect Speech Deepfakes

“…2. Out-of-domain: We trained the Mandarin network with FAD (Ma et al, 2022), another Mandarin-language dataset. We used the pre-trained ASVspoof network (Delgado et al, 2021) for English-language evaluation.…”

Does Audio Deepfake Detection Generalize?

“…A potential solution lies in the ML-driven detection of such deepfakes using, for example, binary classifiers to discriminate between genuine/bonafide and AI-generated speech. The field has witnessed a surge in research, from the creation of extensive datasets [10,11,12,13,14,15,16] to the development of new detection models [17,18,19,20,21,22]. Most notably, initiatives such as ASVspoof [23,24,25] which were launched to benchmark competing detection solutions, seemingly show impressive progress; lower and lower state-of-the-art error rates are reported on a regular basis [21,22,26].…”