Voice drift is a known technical challenge in autoregressive speech generation models, where accumulated errors during step-by-step generation cause the timbre to deviate from the target speaker.