Talk Like You: Imitating How You Speak in Real-Time Talking Head Generation

In plain English

Most talking-head systems get the lip shapes roughly right but strip out the quirks that make a particular person recognisable. TalkLikeYou models an individual's speaking habits in motion space and uses single-step Flow Matching to stay real-time, letting a user pick a habit from a preset or copy it from a reference video. It also proposes a metric, PLAD, for how accurately a habit was imitated — useful if you are trying to measure whether a cloned presenter actually resembles the person.