I went and visited the project of Prof. A. Kondoz in Surrey who got the furthest with a
DoV modem in the public domain and he explained their modeling and system. Essentially
they got up to about 1 Kbit with 2-8% BER on real-world single-operator (!) GSM-networks
running AMR full rate. Kondoz has developed a LPC10 variant for codec that could work with
that bandwith, as expected sounding like robots talking through rain-drains. His main
speciality is voice codecs.
The computational effort for the modem and the codec was in the 1.5 GHz range, so on the
verge of being doable on todays mobile hardware. The project did afaik not progress much
further cause the students who did it graduated and the funding apparently dried up (or
the project was taken over by a .gov contractor for development of a clandestine comms
system, depending on which rumor you believe). Performance over multiple operators (read:
international calls) was only good enough for slow message transfer (low hundreds of
bit/sec), which is not very attractive.
The crucial trick Kondoz used was to adapt the symbol table for the DOV-modem dynamically
to the AMR voice codec codebook changes. If you want to bring that work further, the first
step is of course inserting the frames directly into the baseband (this skipping one codec
step), then building a feedback loop that passes information on the AMR codec state on the
receiving end to the sender, so it can adapt the modem symbol table accordingly. That
combined may yield the crucial datarate improvement to make it work accross operators and
with better understandability. At CryptoPhone we stopped further research into the project
due too high cost and risk of failure (all others I know off canceled as well).
Let me know when you succeed so we can license the tech, until then we just use IP ;-)
Greetings from Berlin,
Frank Rieger