So after quite a few headaches with various libraries and plenty of time spent recompiling, I've done a few tests using the Raspberry Pi (raspbian armhf) as a potential live stream source.

I'm only interested in AAC+ 64kbit stereo, so my tests are focussed there.

I'm using a Behringer UCA202 class-compliant USB audio interface as the audio source, with input.alsa() and streaming to an icecast2 server.

I've tested with both the %aacplus and the %fdkaac encoders.

%aacplus results in approximately 54-57% CPU usage, with minor audio glitches occuring every 10 seconds so.. these glitches are definitely more frequent (and more audible) when the CPU usage is higher. (see below).

%fdk-aac using mpeg4_he_aac results in approx 51-55% CPU usage, so is a slight improvement.

%fdk-aac using mpeg4_he_aac_v2 results in a fairly solid 42% CPU usage... a definite improvement there, although the audio glitches are still there occasionally... enough that I wouldn't want to use this for anything production just yet. 

Interestingly the CPU usage when streaming an mp3 file using this encoder is 55-60%. Even when doing other tasks to push the CPU higher, the audio is still a lot cleaner and more glitch-free than any of the live stream tests. (although they do appear when CPU goes over around 75-80%)

My suspicion is that these audio glitches are related to the ALSA driver not quite talking with my USB audio interface correctly, although I'm not seeing any 'catch up' errors in the liquidsoap logs...

I've used the guide at http://wiki.linuxaudio.org/wiki/raspberrypi to tweak the rpi as much as possible for audio, but there are still issues. 

Trying to use Jack on the rpi has been a fairly dismal failure. Jackd1 (patched for the rpi) runs, but even using jack_rec to record a short audio sample results in around one second of clean audio, then overlaid with static / white noise. Trying to run without the -s flag on jackd results in many alsa xrun error warnings, so I think the rpi is just failing to keep up with the audio. 

Liquidsoap also segfaults as soon I try to use input.jack, so something is quite broken there.

Attempting to use jackd2 ended with failure to allocate memory and Bus Error's when trying to launch jack.

I've not tried portaudio or pulseaudio yet, but I'm not really sure if either would be any better than ALSA.

I'm planning to try to find a different USB audio interface to try out so testing will continue.