CPU consumption should be minimal, for the most part it resides in the network libraries / ip-stack. murmur doesn't do decode-multiplex-reencode. The audio-streams are forwarded/copied to the appropriate clients verbatim.
RAM working set are limited to input/output buffers for the clients (which would be small since one of the project goals is optimizing for low latency), user metadata, some state and acl's.
According to the mumble/murmur FAQ:
Worst case scenario: Number of users × Number of talking users × 60 kbit/s. With less aggressive quality settings, it's ~20 kbit/s, and the bare minimum is 12kbit/s. Note that Mumble is geared towards social gaming; its quality enables people to talk naturally to each other instead of just barking short commands, so the amount of "users talking at the same time" can be somewhat higher than expected.
This means that a server with 20 players and 2 players talking at once requires 0.8-2.4 Mbit/s, depending on quality settings. In the server's .ini file, you can specify the maximum allowed bitrate for users as well as the maximum number of clients to allow.
Disk IO and storage space needed, is going to be minimal, maintaining userdatabase and acl's.
If you want a more authoritative answer you could check out the contact section of the FAQ, this sounds like something that possibly should be included in the FAQ.