and hopefully a bit of compression
Not sure exactly what you mean there; making something "readable" or "pronouncable" will inevitably expand the space required for it. Maybe you meant "hopefully a bit of redundancy"? It would be good if, even if the user makes a small mistake, the system can detect and perhaps even correct it.
Really it depends very much on how big your UUIDs are and how they are most often communicated. If they need to be communicated over phone or VoIP, you want more audible redundancy. If they need to be entered into mobile devices with numeric keypads, it tends to be difficult to enter alphabetic characters, moreso if they are case-sensitive. If they are written down a lot, you need to worry about characters that look similar (O and 0 and o, for instance). If they need to be memorised, then probably strings of real words are the best (have a look at the PGP Word List).
However I think a great all-round solution is just using numberic digits. They're a lot harder to confuse with each other (both when spoken and written) than some alphabetic characters. Easy to enter on mobile devices, and people aren't too bad at memorising numbers.
And the length of the string is not too bad either. Let's compare base32 with base 10 (decimal). The length of a decimal string is log_10(32)
times the length of the corresponding base32 string, or about 1.5 times as long. Ten characters of base32 correspond to 15 decimal digits.
Not much of a penalty, IMO, seeing as in base 32 it's easy to confuse C and T, or S, F and X (when spoken), and someone speaking with a foreign accent is more likely to cause trouble.