Free Stock Music provides free music for download and even has a tag for music on hold. Beatsuite is a great resource for royalty-free hold music, and they also offer professional voiceover services for on-hold messages. Easy on Hold is another great site, and they even break down how their system works with all types of phones, including VoIP!
Although most clients select just one voice talent to read their entire script, some choose two voices for a more interesting finished product. The most common choice is to have both a male and female read the script in an alternating format.

Before you start, check with your IT support team or phone system vendor to see if there are any special requirements you need to be aware of. Things like: How long can the recording be? What type of audio file does it need to be? Is there a maximum file size?
For those still using analogue or digital phone systems, voice on ISDN circuits has a bandwidth of around 4000Hz. The audio from an analogue phone is sampled at 8000Hz, and because of the nyquist theorem, must be low-pass filtered to 4000Hz. So any frequencies in the audio beyond 4000Hz are lost. Interestingly, the human ear and mind are good at recognising human speech, so this is considered "toll-quality" voice, i.e. good enough to pay for. This is the reason why certain vocalised letters are difficult or impossible to tell apart on the phone. For example, the differences between B, P, T, C, etc. are almost entirely above 4000Hz.