Join us live on Zoom >
( Register to receive a reminder )
Wednesday, Feb 18, 2026 04:30pm - 05:30pm ET

Towards Robust and Trustworthy AI Speech Models: What You Read Isn't What You Hear

Feb 18, 2026

Abstract

Deepfake voice technology is rapidly advancing, but how well do current detection systems handle differences in language and writing style? Most existing work focuses on robustness to acoustic variations such as background noise or compression, while largely overlooking how linguistic variation shapes both deepfake generation and detection. Yet language matters: psycholinguistic features such as sentence structure, complexity, and word choice influence how models synthesize speech, which in turn affects how detectors score and flag audio. In this talk,  we will ask questions such as: "If we change the way a person writes, while keeping their voice the same, will a deepfake detector still reach the same decision?" and "Are some text-to-speech and voice cloning models more vulnerable to shifts in writing style than others?" We will then discuss implications for designing robust deepfake voice detectors and for advancing more trustworthy speech AI in an era of increasingly synthetic media.