H.323 Annex C describes a mechanism for carrying RTP in AAL5 frames, but it's got some problems with header overhead, even without IP. Francois Audet at Nortel has done some analysis and made some packetization recommendations to much improve the situation, but not really fix it.
"Avoiding the (now really significant) overhead of AAL5 is best done using AAL2 (the new one). The remaining work is defining a mapping of IP/IP header."
We have recently proposed that the ideas of header compression for RTP/UDP/IP be used when carrying RTP payloads over ATM using AAL5. If you "model" an ATM VC as a point to point virtual link, then you can carry the compressed header in the AAL5 trailer (a sequence number is needed, which we propose should be in the U-U field; everything else is already there, with the VC providing the function of the context ID). The rest of the trailer (length, CRC) provides useful and necessary functionality.
This way, you get reasonable efficiency when carrying real-time data over ATM using AAL5. In particular, with 5 msec G.711 voice, a cell carries all the needed information, including the AAL5 trailer.
We use a "header extension" bit in the U-U field to allow us to communicate uncompressed (partial or full) headers.