I have a requirement and a theoretical implementation, but I'm not sure it's right. A person a, input a voice through the microphone, through the short-time Fourier transform, will generate a sample frequency, through deep learning can analyze the characteristics of this person's speech frequency fa

A person B, input a voice through the microphone, also through the short-time Fourier transform, generate a sample frequency, also through deep learning to analyze the person's characteristic frequency Fb.

Since the Fourier frequency domain is composed of several sine or cosine functions, I remove the speaking frequency Fb of B from the frequency domain of B, and then combine it with the Fa frequency of a, so as to achieve the effect of B imitating the sound of A.