AI-Synthesized Voice Detection Using Neural Vocoder Artifacts

Authors: Chengzhe Sun, Shan Jia, Shuwei Hou, Siwei Lyu

Published: 2023-04-25 18:36:28+00:00

Comment: Paper accepted in CVPRW 2023. Codes and data can be found at https://github.com/csun22/Synthetic-Voice-Detection-Vocoder-Artifacts. arXiv admin note: substantial text overlap with arXiv:2302.09198

AI Summary

This study proposes a novel approach to detect AI-synthesized human voices by identifying artifacts introduced by neural vocoders in audio signals. It introduces a multi-task learning framework for a RawNet2 model, where vocoder identification serves as a pretext task to constrain the feature extractor to focus on vocoder-specific artifacts. This method aims to provide discriminative features for the final binary classifier, achieving high classification performance.

Abstract

Advancements in AI-synthesized human voices have created a growing threat of impersonation and disinformation, making it crucial to develop methods to detect synthetic human voices. This study proposes a new approach to identifying synthetic human voices by detecting artifacts of vocoders in audio signals. Most DeepFake audio synthesis models use a neural vocoder, a neural network that generates waveforms from temporal-frequency representations like mel-spectrograms. By identifying neural vocoder processing in audio, we can determine if a sample is synthesized. To detect synthetic human voices, we introduce a multi-task learning framework for a binary-class RawNet2 model that shares the feature extractor with a vocoder identification module. By treating vocoder identification as a pretext task, we constrain the feature extractor to focus on vocoder artifacts and provide discriminative features for the final binary classifier. Our experiments show that the improved RawNet2 model based on vocoder identification achieves high classification performance on the binary task overall.

Key findings

The proposed multi-task learning RawNet2 model achieved superior performance with the lowest Equal Error Rates (EERs) of 0.13% on LibriSeVoc and 0.19% on WaveFake datasets in intra-dataset evaluation, outperforming other baselines. It also showed good robustness to common post-processing operations like resampling and background noise, though generalization to entirely unseen vocoders in cross-dataset evaluation showed some limitations.

Approach

The authors propose a multi-task learning framework based on a RawNet2 model. This framework combines a binary classification task (real vs. synthetic) with a vocoder identification module as a pretext task, sharing a common feature extractor. This setup forces the feature extractor to learn vocoder-specific artifacts, which are then used for robust synthetic voice detection.

Datasets

LibriSeVoc, WaveFake, ASVspoof 2019

Model(s)

RawNet2 (as the backbone model for feature extraction)

Author countries

USA

← Previous