In the 'rnnoise_process_frame function' of RNNoise v0.2, why is the gain output from the current frame used to process the previous frame

[Franklheq]

In line 457 of the src/denoise. c file, in the rnnoise_process_frame function, it can be seen that the previous operations are all processing the current frame.

But the operation after line 478 is to apply the gain obtained by inputting the current frame into the neural network to the previous frame.

I have tried to modify v0.2 to apply the current frame output to the current frame like v0.1, but the effect does indeed deteriorate. I'm not quite sure about the principle behind it, can someone help me answer it?