This is a small sample of what we can do with audio on the network. The first thing we can do is indeed to encode it a various data rates. Indeed, the quality of the audio depend of the compression ratio. To illustrate that, we first compressed then decompressed a sample of sound using various compression techniques: Indeed, you will notice that our LPC codec is not very good. In fact, we are looking for partners here. Whoever can provide us with more efficient coders is welcomed.

The reason we use this is because we study the transmission of voice over packet networks. Packet networks do nasty things for you. First, they tend to introduce random delays. But we know how to deal with that - Van Jacobson demonstrated the delay adaptation algorithms in his program, "vat". Then, they tend to loose packets - or deliver them too late, which is equivalent. For exemple, this 40 kps transmission was affected by a 20% loss rate . You can observe silences in the middle of the phrases - this is really nasty.

Real men don't leave silences when a packet is lost. They will simply replace it by a copy of the previous packet. With the same lost rate, the replacement strategy results in better understandability but this is far from perfect.

We have observe that most losses are isolated. We hope to cope with the losses by introducing redondancy in the packets, i.e. sending a more compressed version of the previous packet. Under the same conditions which we tested previously, this is the effect of various redondancy techniques:

Sounds promising, right? 20% is really a very high loss rate, yet we manage to cope with it. Well, it is now up to us. We really have to integrate this into IVS.

Acknowledgments

This is brought to you by Christian Huitema