Thus spake "Howard, W. Lee" <L.Howard@stanleyassociates.com>

That's interesting. . . where's the intersection of the packet size curve and the latency curve?

Many equipment vendors allow you to specify the number of ms of data to include in each packet while others require you to specify byes; I'll assume the former here since the latter is just a linear relation. Toll-quality voice requires a one-way latency of under ~125ms including any processing inside the endpoints. Increasing the packet size inherently adds delay on the transmit side. Then you have the obvious network latency. Finally, the receive side will have a buffer to smooth out jitter in the network; most vendors' equipment is now adaptive, so the jitter buffer might be anywhere from 10-50ms. To keep under budget, at least one of these factors must be minimized. Unfortunately, the public Internet has substantial jitter and high coast-to-coast latency, so often the only factor under your control is the transmit buffer. OTOH, if you're going across a network with decent QoS or within the same general area of the country, you can afford a larger transmit buffer without risking the "walkie talkie" effect.

I mean, where would you set it, and can you offset some of that with fragmentation and intervleaving?

F&I is a technique for reducing jitter on slow, congested links like the last mile to a customer. It's often combined with a priority queue, since the latter is not enough on such links (but is on faster ones). Neither has much to do with the (tiny) sizes of voice packets. S