LlamaPartialSpoof: An LLM-Driven Fake Speech Dataset Simulating Disinformation Generation

TL;DR

This paper introduces LlamaPartialSpoof, a comprehensive fake speech dataset created with LLMs and voice cloning to evaluate and improve countermeasure systems against diverse disinformation tactics.

Contribution

It presents a new large-scale dataset with both fully and partially fake speech, highlighting vulnerabilities in current detection systems and emphasizing the need for more robust defenses.

Findings

Current detection systems have high error rates, with the best at 24.49% EER.

Vulnerabilities include biases toward specific TTS models and concatenation methods.

Detection systems struggle to generalize to unseen fake speech scenarios.

Abstract

Previous fake speech datasets were constructed from a defender's perspective to develop countermeasure (CM) systems without considering diverse motivations of attackers. To better align with real-life scenarios, we created LlamaPartialSpoof, a 130-hour dataset that contains both fully and partially fake speech, using a large language model (LLM) and voice cloning technologies to evaluate the robustness of CMs. By examining valuable information for both attackers and defenders, we identify several key vulnerabilities in current CM systems, which can be exploited to enhance attack success rates, including biases toward certain text-to-speech models or concatenation methods. Our experimental results indicate that the current fake speech detection system struggle to generalize to unseen scenarios, achieving a best performance of 24.49% equal error rate.