The cpu cost of the hierarchical coding algorithm mainly depends of the compression scheme used following the temporal subband decomposition. In our experiments, we have successfully tested the PCM, ADM6, ADM5, ADM4, and ADM3 compression schemes, which have very low cpu requirements [14]. More efficient compression coder could as well fit into the temporal subband decomposition described above. However, very high compression coding algorithms do not appear to fit well with layered coding over the Internet because the overhead of IP, UDP and RTP[32] headers decrease significantly their bandwidth savings.
Table 1 shows the bandwidth overhead for layered and non-layered audio coding using PCM and ADM4 compression schemes. The IP/UDP/RTP bandwidth headers overhead is shown for three sizes of packets corresponding to 20ms, 40ms and 80 ms of compressed speech.
Table 1: Bandwidth overhead for several audio coding/transmission schemes
Table 2: IP/UDP/RTP header size
Interactivity needed by audio/video conferencing applications is often cited as requiring low packetization intervals of 20ms. However, such a low packetization interval increases the number of packets sent per second, especially if several layers are used. Then the bandwidth requirement of the IP/UDP/RTP headers may be higher than the bandwidth requirement of the actual payload, see Table 1. For example, the bandwidth corresponding to the headers for 6 flows is (for 20ms packetization interval) equals to 106 kbps with IPv4 and 144 kbps with IPv6. Table 2 recalls the size of IPv4, IPv6, UDP and RTP headers.
However, experience with audioconferencing tools in the MBone shows that 40 ms and 80 ms are convenient values especially if redundant techniques [28] are used to minimize the effect of large ``holes'' in the speech. 80 ms appears to be a good compromise between interactivity needs and bandwidth overhead generated by the hierarchical coding.
In some cases (e.g. very low bandwidth links), the IP/UDP/RTP headers may be compressed to reduce the bandwidth overhead [8]. However, such techniques increase the CPU consumption at the routers and they are expected to be used on point to point links, but not on an end-to-end basis on the MBone.