From Bob Tyler (Solutions Architect, Impatica Inc.) I usually recommend audio at 32khz (and not necessarily 11,025hz) as I find I get better 'cleaner' post-compression results. However, in some cases this might adversely impact on the finished impaticised file sizes - though with todays increased bandwidths this is likely to be less of an issue.
There really are two "families" of sample rates:
1) the 8K Digital Audio family, which is evenly divisible by 8.
2) the 11K CD Audio family, which is based around 44.1 kHz, 16-bit, Stereo Compact Discs and is NOT evenly divisible by 8.
So the 11,025hz, 22,050hz and 44,100hz sample rates all fall into Group 2.
Early on we used to record our narration in professional sound studios. They generally wanted to deliver "CD Quality" audio and tests with 44, 22 and 11K audio source files were indistinguishable in audio quality and .imp file size -- so 11K became a standard that the studios weren't too unhappy delivering and didn't chew up a lot of disk storage space.
Tests with Group 1 (the 8K family) gave similar results: The translations from all sample rates were indistinguishable in audio quality and final .imp file size -- but there was an "edge" or crispness to the translations from the 11K family that seemed to suit our voices better than the 8K family examples.
So, yes the 8K family translations do sound a little smoother than the 11K family samples -- presumably because the 8K family is evenly divisible by 8 whereas the 11K family is not -- but I (and many others) liked the way that the "edge" to the 11K translations helped the final audio punch though a little better.
Female voices seemed to sound cleaner when translated from 11K family sources as well.
However, in theory, the best downsampling should occur when everything is evenly divisible -- hence the recommendation for recording (or exporting) at 8,000 Hz in our documentation.
The assumption is that third party production tools, in general, will do a better job of downsampling than the simple downsampling routines that we use in our Impatica translators and, since we end up converting everything to 8K Sun .AU format anyway, why not just record and produce original 8K audio files for use with Impatica? Makes sense to a software engineer at any rate.
However, from an audio engineering standpoint, 8K samples aren't any good for recording /anything/ over 4 kHz in frequency.
The problem is that, with 8K recordings, it is impossible to apply any sort of useful EQ or audio sweetening in post-production and 8K simply is not appropriate for female (and many male) voices.
Additionally, Adobe Premiere (5.1), Camtasia Recorder and many other software tools export totally useless, completely unacceptable 8K samples -- so that does leave one looking for a higher frequency, more professional base sample rate that does not require too much storage space.
So the suggestion is to use a frequency from the 8k 'family' of sample rates in order to achieve optimum quality audio. Ideally use 32khz where filesize is less of an issue and 16khz otherwise.