The upper one is definitey the better one.
Spectrograms (like the ones you produced with Spek) are used to display and inspect tendencies over time. For example if some noisy signal is present all the time, it will show up as a straight horizontal line somewhere:
You can determine some general properties, like that 20 Khz cut, but an average frequency spectrum plot for the whole song is really more ideal for that, like this one (har-bal is a handy tool here):
That said, there are some other things, like lossy compression (mp3 etc) that shows up pretty significantly (I guess that was what you were comparing?), and being able to see that over time, confirms that is not a momentary tendency of some sort:
Based on this you could generalize that quality signals are continuous and smooth on the Y axis, i.e. no sudden cut offs. On the X-axis you should see some variation and no straight horizontal lines going through the whole spectrogram. Ofcourse this depends very much on the material (if someone thinks holding down that C key all the way .. it will show up as a line).
Here is the difference plot from your two plots. It shows some action all over the frequencies, which could be due to some variable bitrate (or in general lower sample rate):