A new automatic speech-recognition task lets selected participants opt out of a meeting transcript without leaving the conversation. The proposed system continues transcribing everyone else and marks when an excluded person is speaking, but attempts not to preserve that person’s words.
Researchers attach a lightweight enrollment-conditioned gate to a frozen speech language model. A participant can enroll at inference time, including someone the gate never encountered during training, so the underlying recognition model does not need to be retrained for each new opt-out request. Tests covered English meetings from AMI and Mandarin conversations from AliMeeting.
For excluded speakers, the share of correctly transcribed words or characters fell from 72.3% to 48.2% in English and from 73.6% to 27.3% in Mandarin, while error rates for retained speakers remained roughly stable. Those numbers also expose the current limitation: substantial excluded speech was still recoverable, particularly in English, so this is not yet a strong privacy guarantee. The research establishes a useful interface and evaluation problem, but meeting platforms would need stronger suppression, adversarial testing, and clear signaling before presenting it as dependable consent control.