Echo cancellation method and system based on heterogeneous multi-scale convolution recurrent network

Abstract

The application discloses an echo cancellation method and system based on a heterogeneous multi-scale convolutional recurrent network, which first constructs a correlation encoder based on the extraction of relevant feature patterns in multiple frames of context, then constructs a spectral encoder according to the extraction of spectral patterns in a single frame, and then adopts the correlation encoder and the spectral encoder to form a heterogeneous multi-scale encoder and obtain multiple paths of encoded features; the application realizes the function of extracting different features from the input mixed signal spectrum and the reference signal spectrum by using the heterogeneous dual-path encoder, adopts correlation coding to adapt to the time delay between the reference signal and the mixed signal, and extracts correlation feature expressions from multiple frames of spectrum, simultaneously adopts spectral coding to extract spectral pattern expressions from the current frame in order to strengthen the spectral resolution of the current frame, and realizes the function of further adopting a channel attention feature soft selection mechanism to fuse the correlation feature coding and the spectral coding for the dual-path features, and is suitable for being widely promoted and used.

Images