Heterogeneity over Homogeneity: Investigating Multilingual Speech Pre-Trained Models for Detecting Audio Deepfake

View on arXiv ← Back to list

Authors: Orchid Chetia Phukan, Gautam Siddharth Kashyap, Arun Balaji Buduru, Rajesh Sharma

Published: 2024-03-31 21:48:50+00:00

AI Summary

This research investigates the effectiveness of multilingual speech pre-trained models (PTMs) for audio deepfake detection. The study finds that multilingual PTMs outperform monolingual PTMs and propose a novel fusion framework, MiO, achieving state-of-the-art performance on two datasets and comparable performance on a third.

Abstract

In this work, we investigate multilingual speech Pre-Trained models (PTMs) for Audio deepfake detection (ADD). We hypothesize that multilingual PTMs trained on large-scale diverse multilingual data gain knowledge about diverse pitches, accents, and tones, during their pre-training phase and making them more robust to variations. As a result, they will be more effective for detecting audio deepfakes. To validate our hypothesis, we extract representations from state-of-the-art (SOTA) PTMs including monolingual, multilingual as well as PTMs trained for speaker and emotion recognition, and evaluated them on ASVSpoof 2019 (ASV), In-the-Wild (ITW), and DECRO benchmark databases. We show that representations from multilingual PTMs, with simple downstream networks, attain the best performance for ADD compared to other PTM representations, which validates our hypothesis. We also explore the possibility of fusion of selected PTM representations for further improvements in ADD, and we propose a framework, MiO (Merge into One) for this purpose. With MiO, we achieve SOTA performance on ASV and ITW and comparable performance on DECRO with current SOTA works.

Key findings

Multilingual PTMs achieved the best performance for audio deepfake detection across three benchmark datasets. The proposed MiO fusion framework further improved results, achieving state-of-the-art performance on ASV and ITW and comparable performance on DECRO. Cross-corpus evaluation showed better generalization for models trained on multilingual PTMs.

Approach

The authors extract representations from various pre-trained models (including multilingual, monolingual, and those trained for speaker/emotion recognition). These representations are fed into simple downstream networks (FCN and CNN) for deepfake detection. A fusion framework, MiO, combines representations for improved performance.

Datasets

ASVSpoof 2019 (ASV), In-the-Wild (ITW), DECRO

Model(s)

XLS-R, Whisper, MMS, Unispeech-SAT, WavLM, Wav2vec2, x-vector, XLSR_emo, FCN, CNN

Author countries

India, Estonia

← Previous