Actual circuit switched connections, regular voice calls, would've been quite low latency. Later there got to be some delays from digitization; even then it would've been hard to detect in actual conversations.<p>ISDN lines were favored by radio stations <i>forever</i> because of the reliably low latency and jitter they offered.<p>Actual numbers? I wanna guess ~100ms max for a cross country call, pre-digital long distance. Maybe 350ms post digital. IIRC the ISDN hop I had to a local CO was ludicrously low, like 9ms? but i could be wrong. That was years ago.<p>Digital links for internet connections were lower inherent latency but then had layers of other modulation happening on top of them, and the variations there are a different order of thing. I assume you mean bare link latency.<p>I've NFI how many different layers of re-encoding your data might be expected to go through on a standard call today; but the list of transitions would probably be long and amusing.
Circuit switched analog POTS has been dead for a long time, but delays were limited to amplifier delays (almost nothing) and speed of light in wires or radio.<p>Digitial T1 PRI calling may not be fully dead, you might have a miniscule delay for the analog to digital sampling, but the multiplexing on a T1 and higher is one sample per channel at a time. Where trunks meet, there's got to be a buffer to synchronize, but it doesn't need to be more than one sample. At 8000Hz sampling, a couple one sample buffers here and there doesn't amount to much. Of course, there's still amplifiers and the speed of light in wires and radio.<p>For VOIP calls, most audio codecs deal in 20ms packets, and calling software often sends between 1 and 5 audio packets in an IP packet. Plus it takes time to encode and decode. So you've probably got 25 to 110ms sampling delay added there. Then there's a jitter buffer. T1 samples arrive on time, pretty much all the time; UDP packets tend to have inconsistent delay (jitter) so you've got to add an artificial delay of (worst case delay - best case delay). That depends a lot on the properties of the path between the two devices; shared medium networking like wifi, cell networks, cable modems, old unswitched ethernet can have a lot of difference between worst and best case delay. Some software has a fixed length jitter buffer of about 1-2x the IP packet length; others will adapt to the observed arrival time differences, up to some limit.<p>Some (I think most) video calling applications send audio and video as separate streams of IP packets. It's also possible to include audio and video data in the same IP packets. It is possible to reduce audio delay in that case, by using smaller audio packets, reducing the sampling delay. It's technically possible to send small audio packets at higher frequencies for an audio only call, but the overhead of an IP header, UDP (usually) header, and application header is already pretty significant. Sending twenty 1 ms audio packets would use a lot more bandwidth than sending one 20 ms audio packet; I'd guess about 10 times as much. But if you piggyback on video packets, it might not be as costly, and video packets are usually high frequency anyway.
I found a mention of "ITU standards (G.114) for maximum transmission path delay of 50 ms for call transmitted over either the continental US or the Atlantic."