Abstract: A computer-implemented method is disclosed in which a first machine-learning model is used to generate an intermediate estimate of a target signal within input audio. Based on this estimate and the input audio, parameters of an adaptive filter are computed. The adaptive filter is then applied to the input audio to produce a filtered audio signal. A second machine-learning model generates an enhanced audio signal based at least in part on the filtered signal, with both models jointly trained end-to-end to enhance the target signal. Various other aspects are also disclosed.

Abstract: Methods and systems are described for acoustic feedback cancellation through informed beamforming associated with a device. In various examples, systems or methods receive audio signals from a microphone and determine a transfer function between a speaker and the microphone. A feedback covariance may be generated using the determined transfer functions, followed by generation of a total noise variance based on combining the feedback covariance with a noise covariance. Beamforming may then be performed based on the total noise variance to spatially filter and suppress audio signals from feedback paths while preserving signals from target directions. The system may adapt to changing acoustic conditions by updating the feedback covariance matrix and adjusting the beamforming accordingly.