FROM NATURAL TO SYNTHETIC: THE RISE OF VOICE CLONING AND THE FATE OF VOCAL AUTHENTICITY

Abstract

This article aims to analyses voice cloning techniques through an exploratory study of the multilingual zero-shot XTTS-v2 model in the context of Brazilian Portuguese. Thirty synthetic samples were produced based on three reference conditions of approximately 8, 14, and 28 seconds, and subsequently compared with human recordings using acoustic-prosodic descriptors. The reference condition of 28 s showed the lowest mean distance and dispersion in the exploratory index, suggesting greater stability under this condition. While variation was observed across the generated samples and the index does not directly measure vocal naturalness or voice identity, the results still indicate that short reference samples can already produce synthetic speech that approximates acoustic and prosodic properties of the original voice. The findings reinforce the value of multidimensional analysis of vocal identity, contributing to discussing the limitations and impacts of voice cloning and to studies on synthetic speech detection in Brazilian Portuguese.