AI-Synthesized Voice Detection Using Neural Vocoder Artifacts

View on arXiv ← Back to list

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

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

AI Summary

This research proposes a novel approach to detect AI-synthesized voices by identifying neural vocoder artifacts in audio signals. A multi-task learning framework, using a RawNet2 model with a vocoder identification module, is introduced to improve detection accuracy.

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 improved RawNet2 model achieved low Equal Error Rates (EERs) in detecting synthetic speech on LibriSeVoc and WaveFake datasets. The model also showed good performance on the ASVspoof 2019 dataset and demonstrated robustness against resampling and background noise, although cross-dataset generalization was limited.

Approach

The authors use a multi-task learning framework with a RawNet2 model. This model shares a feature extractor with a vocoder identification module, treating vocoder identification as a pretext task. This constrains the feature extractor to focus on vocoder artifacts, improving the binary classification of real and synthetic speech.

Datasets

LibriSeVoc (newly created dataset with self-vocoded samples from six state-of-the-art vocoders), WaveFake, ASVspoof 2019

Model(s)

RawNet2

Author countries

USA

← Previous