Last post by saratoga -
Spectrograms are useful for looking at analog equipment, but for a perceptual encoder they can't tell you anything. The "denoise" you see is just your imagination. If you were making changes based on that, then your changes are random. That is a bad way to do things.
If you couldn't ABX the original file then what you are doing is pointless; you can't improve on something that doesn't have any audible flaws. You need to find something that can be improved first.
If you want to learn about audio codecs, I suggest starting at lower bitrates, maybe 100 kbps, and then finding files where you can hear a problem. Then you can try tuning. However, this is going to be immensely more difficult then you're assuming, and will require you to have a much deeper understanding of how perceptual encoders actually work.
If spectograms do not mean anything, I wonder why there are still exist.
At a glance, it can be seen some denoise in both encoded samples, especially at high frequencies which anyway bring only annoying HF hiss in ears.
I lowered the maskers of bass, alto and treble to not lose any particular sound from the original sample, which may be covered by stronger maskers.
The original sample (44,100 Hz 16bit WAV) was resampled to 48,000 Hz 24bit and the encoded samples to (48,000 Hz 24bit MP3) and look in spectograms almost identical and sounds almost identical, even on the whatever Hi FI audio system.
I put sfb21=4 to cut any noise >20,000 Hz and to save bits for a better quality encoding at lower frequencies. My goal is to make good quality mp3's for my android. The 4.8X WAV/mp3 dimension ratio it's pretty OK.