Microsoft patents system to grade how well audio fills in missing data

1. A computer-implemented method comprising: providing as input to one or more packet loss concealment models audio signals having missing portions due to network packet loss and receiving from the one or more packet loss concealment models enhanced audio signals comprising restored portions that replace the missing portions of the audio signals using the one or more packet loss concealment models; obtaining quality labels for the enhanced audio signals output by the one or more packet loss concealment models, wherein the quality labels represent perceived audio quality ratings that characterize how effectively the one or more packet loss concealment models enhanced the degraded audio signals by restoring the missing portions of the audio signals lost due to network packet loss; using the enhanced audio signals and the quality labels, training a quality estimation model through supervised learning to receive as input an audio signal enhanced by a packet loss concealment model and to generate as output a synthetic quality label indicating a quality of the audio signal as enhanced by the packet loss concealment model, the training comprising updating parameters of the quality estimation model based on differences between synthetic quality labels output by the quality estimation model and the quality labels; providing new enhanced audio signals, generated by the one or more packet loss concealment models, as input to the trained quality estimation model and generating as output synthetic quality labels for the new enhanced audio signals, each synthetic quality label and its corresponding enhanced audio signal forming a training data pair; and using the training data pairs to improve packet loss concealment performance in an audio application.