[Paper Review] EmoFake: An Initial Dataset for Emotion Fake Audio Detection

Yan Zhao, Jiangyan Yi+4 more|arXiv (Cornell University)|Nov 10, 2022

Music and Audio Processing4 citations

TL;DR

This paper introduces EmoFake, the first public dataset for emotion fake audio detection, comprising real emotional speech and fake audio generated via seven open-source emotion voice conversion (EVC) models. It proposes GADE, a graph attention network using deep emotion embeddings, which achieves state-of-the-art performance with an EER of 0.51% on target emotion detection, outperforming existing models on the new benchmark.

ABSTRACT

Many datasets have been designed to further the development of fake audio detection, such as datasets of the ASVspoof and ADD challenges. However, these datasets do not consider a situation that the emotion of the audio has been changed from one to another, while other information (e.g. speaker identity and content) remains the same. Changing the emotion of an audio can lead to semantic changes. Speech with tampered semantics may pose threats to people's lives. Therefore, this paper reports our progress in developing such an emotion fake audio detection dataset involving changing emotion state of the origin audio named EmoFake. The fake audio in EmoFake is generated by open source emotion voice conversion models. Furthermore, we proposed a method named Graph Attention networks using Deep Emotion embedding (GADE) for the detection of emotion fake audio. Some benchmark experiments are conducted on this dataset. The results show that our designed dataset poses a challenge to the fake audio detection model trained with the LA dataset of ASVspoof 2019. The proposed GADE shows good performance in the face of emotion fake audio.

Motivation & Objective

• To address the lack of datasets focused on emotion-manipulated fake audio, where speaker identity and content remain unchanged but emotion is altered.
• To develop a benchmark dataset that captures semantic risks from emotion tampering in speech.
• To evaluate the robustness of existing fake audio detection models against emotion-fake audio.
• To propose and validate a novel detection method, GADE, specifically designed for emotion fake audio.
• To provide baseline models trained on EmoFake to accelerate future research in emotion-aware fake audio detection.

Proposed method

• EmoFake is constructed from the ESD dataset, using five male and female speakers, each uttering 350 sentences across five emotions: Neutral, Happy, Angry, Sad, and Surprise.
• Fake audio is generated using seven open-source EVC models: VAW-GAN-CWT, DeepEST, Seq2Seq-EVC, CycleGAN-EVC, CycleTransGAN, EmoCycleGAN, and StarGAN-EVC.
• GADE employs graph attention networks to model inter-utterance relationships and integrates deep emotion embeddings extracted from a pre-trained emotion recognition system.
• The model uses a graph-based architecture to capture long-range dependencies and emotional inconsistencies in audio sequences.
• Emotion embeddings are used as node features in the graph, with attention mechanisms to weigh the importance of different utterances in detecting fake audio.
• The framework is trained end-to-end on the EmoFake dataset, with evaluation on unseen emotion transitions and EVC model types.

Experimental results

Research questions

• RQ1 How does the performance of existing fake audio detection models degrade when tested on emotion-fake audio, where only emotion is altered?
• RQ2 Which emotion voice conversion (EVC) models generate the most undetectable fake audio in terms of fooling standard detection models?
• RQ3 Can a dedicated model like GADE outperform general-purpose fake audio detectors on emotion-fake audio?
• RQ4 How does the choice of target emotion affect the detectability of emotion-fake audio?
• RQ5 To what extent does fine-tuning existing models on the EmoFake dataset improve their robustness to emotion-based manipulation?

Key findings

• The LA dataset from ASVspoof 2019 fails to generalize to EmoFake, with all tested models showing increased EER on emotion-fake audio, indicating a significant performance drop.
• GADE achieves the lowest EER of 0.51% on detecting fake audio with target emotion 'Surprise', and 0.55% for 'Happy', outperforming all baselines.
• LCNN shows better robustness to emotion fake audio than ResNet34 and SENet, which suffer from severe performance degradation across all emotion transitions.
• Fake audio generated by the Seq2Seq-EVC model (S3) is the most difficult to detect, suggesting architectural differences in EVC models affect detectability.
• Baseline models fine-tuned on EmoFake show improved generalization, especially GADE, which maintains strong performance across all emotion transitions.
• The performance of detection models varies significantly by target emotion, with 'Surprise' being the most detectable (EER up to 7.26% for RawNet2), while 'Happy' and 'Surprise' are most challenging for SENet and ResNet34.