Multi-scale network with shared cross-attention for audio–visual correlation learning