EZ-Ar-VC: Robust Arabic Voice Anonymization with Generalization across Dialects and Code-Switching

TL;DR

This paper introduces EZ-Ar-VC, a robust Arabic voice anonymization model trained on diverse Arabic speech, achieving strong performance in naturalness, intelligibility, and pronunciation preservation. It outperforms zero-shot VC systems by generalizing across Arabic dialects and handling code-switching, demonstrating the necessity of language-specific adaptation for effective voice anonymization.

Why It Matters

Voice anonymization is critical for privacy in speech applications, especially in multilingual and code-switching contexts like Arabic. This work enables more reliable and inclusive privacy-preserving technologies for low-resource and phonologically complex languages.

Key Findings

• Zero-shot VC systems fail to preserve pronunciation in Arabic due to phonological differences.
• EZ-Ar-VC achieves high speaker dissimilarity and naturalness while generalizing across dialects and code-switching.
• Target-language adaptation is essential for robust voice anonymization in low-resource languages.