I am fairly sure you cannot bypass the amr codec on the device, as the echo cancellation filters in the RNC would distort the datastream if it's not looking enough like speech audio. Anything going down the line needs to be modelled to the range of human audio.

That's basically the reason I haven't really looked too hard at it yet, as it would require some experimenting to find out the modulation sheme that can be passed through without distortion.