amanteur
The sampling rate of the processed song has decreased. Is there any way to restore it to the original sampling rate?
The previous image and the next image are the time-frequency image of the original audio and of the separated audio of other, respectively. It is obvious that the sampling rate of the audio has been reduced. Is there any way to keep the original sampling rate of the other class audio?
The effective frequency of drums, vocals, and bass can be degrade because their energy in the high-frequency is already very low
The model can be further optimized. For the vast majority of speech, 95% of the energy is below 10kHz, but for the "other" class, no frequency components should be discarded. Many string instruments (such as guitars) have high-frequency characteristics, and the high-frequency components of string instruments contain rich harmonic information, which is crucial for the perception of timbre, sound quality. Therefore, I believe that a more targeted selection of the reserved frequency band information would be more appropriate.