Multi-channel speech enhancement model training method, speech enhancement method and device

The invention provides a multi-channel speech enhancement model training method, a multi-channel speech enhancement method and a multi-channel speech enhancement device. The training method comprises the following steps: simulating and generating a multi-channel speech training sample according to a main lobe position of a beam and an expected main lobe shape; the multi-channel voice positive sample comprises a voice signal generated inside the main lobe, and a voice signal of the multi-channel voice negative sample is generated outside the main lobe; and performing phase alignment on the multi-channel speech training sample according to the target sound source direction, extracting spatial features and speech spectrum features, inputting the spatial features and the speech spectrum features into the neural network model, setting a label of a positive sample as a main intra-lobe signal, setting a label of a negative sample as 0, and training to obtain a speech enhancement model. According to the method, the e