We've been working on a new lowlatency codec for speech and music,
CELT, and are about to public a paper on it. (celt-codec.org).
While CELT wouldn't be useful for narrowbanded voice some of the
components of CELT would be very useful in an AMBE killer.
(Particularly CWRS, the algebraic vector quantizer). CELT itself
could be used for a 10-20khz wide mode (40-60kbit/sec with decent SNR)
to replace FM with something with much greater quality if there was
interest... But a lower bandwidth mode will be far more interesting.