公开(公告)号：US12136416B1

公开(公告)日：2024-11-05

申请号：US17857384

申请日：2022-07-05

Applicant: Meta Platforms, Inc.

Inventor： Zhe Liu ,  Ke Li ,  Fuchun Peng

IPC: G10L15/16 ,  G10L15/06 ,  G10L15/30 ,  G10L15/32

Abstract: In one embodiment, a method includes accessing a decoded hypothesis corresponding to an utterance, computing a predicted probability of observing each token in the decoded hypothesis by having a local first machine-learning model process the decoded hypothesis, computing a confidence score for each token in the decoded hypothesis by having a second machine-learning model process the decoded hypothesis, where the confidence score indicates a degree of confidence for the token to be observed at its position, calculating a loss for the computed predicted probabilities of observing tokens in the decoded hypothesis based on the computed confidence scores, and updating parameters of the local first machine-learning model based on the calculated loss.