Stylometric Bias and Machine-Readable Voice in AI Speech Translation

This study explores how stylistic regularities emerge in AI-generated speech-to-text translations by examining the English output of the CoVoST corpus, a large-scale multilingual dataset of voice-translated speech. Without relying on human reference translations, we adopt a stylometric and machine learning approach to identify internal consistency bias—manifested as reduced lexical diversity, flattened syntactic structures, and repetitive discourse patterns—in English translations of Chinese spoken input. Stylometric features such as mean type-token ratio (MTTR), parse tree depth, modal verb frequency, and discourse marker usage are extracted from 16,899 English sentences produced by an automatic speech translation system. Using unsupervised clustering and SHAP-enhanced classifiers, we uncover latent stylistic archetypes and domain-invariant regularities that signal translationese unique to speech-based neural MT systems. The results demonstrate that CoVoST's AI outputs exhibit a distinctive translation style: one that prioritizes simplicity and syntactic regularity at the expense of natural variation. This study contributes to Translation Studies by reorienting the concept of translationese away from human-vs-machine binaries, toward intra-AI stylometric drift. Our findings offer both empirical and theoretical insight into how speech-to-text AI systems reconfigure interlingual stylistic norms.