Voice Personas: Algorithmic Stylization and the Politics of Representation in AI-Mediated Chinese-English Speech Translation

This article critically examines the emergence of stylized, machine-generated "voice personas" within Chinese-English speech translation systems, using the CoVoST dataset as a case study. This study did not involve human participants and used publicly available data from the CoVoST corpus. Employing computational stylometry and qualitative discourse analysis, this study identifies six statistically distinct AI voice clusters—ranging from formal and bureaucratic to conversational and telegraphic—that recur across AI translations, often irrespective of source input variation. These personas are evaluated against human reference translations, revealing significant divergences in lexical choice, syntactic complexity, modality expression, and overall discoursal tone. Moving beyond a purely technical assessment, the article argues that these emergent personas constitute an "algorithmic habitus," reflecting and potentially reinforcing algorithmically encoded ideological preferences and dominant discourse norms. Framed within Venuti's concept of "translator invisibility," feminist translation ethics, and critical AI studies, this study interrogates the complex politics of algorithmic style in cross-cultural communication. The findings highlight how seemingly neutral AI systems actively shape representation, raising crucial questions about agency, bias, and the potential for algorithmic mediation to homogenize cultural expression or perpetuate existing power imbalances. This research contributes to a necessary interdisciplinary dialogue within AI & Society, urging a shift towards more culturally sensitive, ethically accountable, and human-centered approaches to designing and deploying translation technologies.