I've looked at the AT&T simultaneous voice and data stuff (VoiceSpan), and from what I can determine from reading between the lines, the voice stuff is not really digitized (a la PCM), rather it is just pushed around somewhere else in the spectrum. The data is apparently modulated using a variant on QAM, and the data rate varies dynamically depending on whether or not you are saying anything.
As I recall the block diagram is ...
We had a discussion about this on this list a while back. I've been following the matter elsewhere and it is certainly true that some of the proprietary stuff and the proposed standard are fully digital packet interleaved digital CLEP coded voice. It would be possible to use the echo cancelling technology of modern modems (which subtract out the transmit signal to leave the receive) to subtract out both the transmit and receive digital signals (QAM by the way for v.32 and v.34) and leave just the residual noise which could be voice at a very low level. Whether one could get an adaquate bit error rate (even with the trellis coding) from the far end data signal given the worst case line loss if the signal was mixed with low level voice I do not know. In any case if the analog voice under QAM data trick was actually used, one could have a reasonable security analog masking type scrambler for free by sending random digital data down the line in both directions (such as data derived from Johnson noise or radioactive decay). There would be no easy way for a third party to filter out the digital data (unless of course the interloper had a four wire tap on the line with good enough directionality to demodulate the data in both directions with a reasonable BER). Dave Emery die@die.com