Round trip time

Latency & PerformanceRTTalso: ping timealso: round-trip delayalso: network round trip

In one sentence

Round trip time (RTT) is how long a packet takes to reach a server and come back, the network floor beneath every other latency number, and the one component no amount of software optimization can push below.

10 ms per 1,000 km round trip Speed of light in optical fiber, standard network engineering, 2026Last reviewed 31 July 2026

Not to be confused with End-to-end latency, or Jitter.

Definition

Round trip time is how long a single packet takes to travel to a server and return. It sets the network floor that every other latency number is stacked on top of.

No amount of software work goes below it. It is fixed by distance and physics, which is what makes it the one latency component you cannot engineer away.

Round trip time is set by physics and routing, which is what makes it the one latency component you cannot engineer away. Everything else in a voice pipeline is software you can tune. This is the floor underneath all of it.

What adds to a round trip

  • Propagation delay. Light travels through optical fiber at roughly 200,000 kilometers a second, about two thirds of its speed in a vacuum. That works out to around 5 milliseconds per 1,000 kilometers each way, or 10 milliseconds of round trip per 1,000 kilometers.
  • Routing. Traffic rarely takes the geometric shortest path, so the distance actually traveled runs longer than the distance on the map, often by a lot.
  • Queuing and processing at each intermediate hop.
  • Last-mile access, the link into the visitor's own device, which is frequently the single largest contributor on a mobile network.

The geography a voice deployment lives with

  • A visitor in Mumbai reaching a server on the US east coast is covering roughly 13,000 kilometers, which sets a theoretical minimum round trip of around 130 milliseconds before any routing inefficiency, queuing or processing is added.
  • Actual observed round trips on that path run higher than the minimum.
  • Against a sub-second time to first audio budget, that is a substantial fixed cost paid on every single exchange, and it is paid more than once if the pipeline makes several round trips in a turn.

Why it multiplies in a voice pipeline

  • A pipeline that calls recognition, the model and synthesis as separate services in different regions pays the round trip repeatedly, once for each hop.
  • Consolidating those services into one region, or placing them near the visitor, is frequently the largest latency saving available, and it needs no change to any model.

The tension with data residency

  • Residency requirements can force processing into a specific region regardless of where the visitors are.
  • That is a real tradeoff between compliance and responsiveness, and it belongs in qualification rather than in a surprise at implementation.

Cellular access adds delay beyond a fixed-line equivalent and it moves with signal conditions, so a mobile visitor gets a worse round trip than desktop testing suggests.

Common misconception

That network latency is negligible on a modern connection. On a local network it is. Across continents it is a fixed floor of a hundred milliseconds or more, and it is the reason regional deployment is a product decision rather than a detail.

Why it matters commercially

Round trip time decides whether a globally spread audience can get a responsive experience out of a single region. For a business serving several geographies it is an infrastructure choice with a direct consequence for how the product feels, not something the platform team can settle quietly on its own.

In voice specifically

In text the round trip is spent loading something, and a spinner covers it. In speech the round trip is spent as silence after the visitor stops talking, and there is nothing to put over it. The same network floor a web page hides, a voice agent puts on display.

Visual

The floor you cannot optimize below

The floor you cannot optimize belowWhat the visitor paysEffect on a sub-second budgetServer in the visitor'sregionA few milliseconds each exchangeThe network floor is a rounding errorServer across the worldAround 130 ms for a Mumbai visitor reaching a US east coastserver, roughly 13,000 km, before any routing or queuingClose to the whole 200 ms human turn gap, paid on every turnPipeline split acrossregionsThe round trip is paid on every hop, several times a turnThe largest and most avoidable latency in the buildPipeline consolidated inone regionOne short round trip near the visitorOften the biggest saving available, with no model change

Physics sets the floor. Everything else stacks on top.

The distances are great-circle approximations and the round trips are theoretical minimums, before routing, queuing or processing, so real observed times run higher. Light in fiber travels about 200,000 kilometers a second, giving roughly 10 milliseconds of round trip for every 1,000 kilometers. Two things the table leaves out: data residency rules can pin processing to a distant region, trading responsiveness for compliance, and cellular access adds delay on top, so a phone visitor gets a worse round trip than desktop testing shows.

Statistics

Every figure carries its source and year. Vendor numbers are labelled as vendor numbers, and where no reliable figure exists this page says so rather than borrowing one.

Light travels through optical fiber at roughly 200,000 kilometers a second, about two thirds of its vacuum speed, which gives approximately 5 milliseconds per 1,000 kilometers one way and 10 milliseconds round trip.

10 ms per 1,000 km round tripindustry range

Speed of light in optical fiber, standard network engineering, 2026 - A physical fact rather than a vendor figure. It is the number that makes intercontinental distance a fixed latency cost you cannot buy your way out of.

A Mumbai to US east coast path of roughly 13,000 kilometers therefore has a theoretical minimum round trip of around 130 milliseconds, before any routing inefficiency, queuing or processing.

About 130 msindustry range

Derived from fiber propagation over roughly 13,000 km, 2026 - A floor, not a measurement. Real observed round trips on that route run higher, and it is quoted as the minimum the physics allows.

Human conversational turn gaps average around 200 milliseconds, so an intercontinental network floor of roughly 130 milliseconds consumes most of a natural turn gap before any processing has even begun.

About 200 msindustry range

Conversational turn-taking baseline, 2026 - A reference point rather than a target. It is what makes the 130 millisecond floor alarming rather than academic.

Sub-second time to first audio is the working design target for conversational voice, and a 130 millisecond network floor spends a meaningful slice of it before any processing runs.

Under 1,000 msindustry range

Industry working threshold, 2026 - A shared working convention, not a published standard. No body sets it and no vendor is audited against it.

Actual observed round trip times vary by route, carrier, time of day and access technology, so published figures for any specific path are indicative rather than guaranteed.

-no reliable figure

Which is why the only round trip you can rely on is one you measure yourself, on the route your visitors actually use.

There is no reliable single figure for how much latency a mobile network adds, because it varies with the cellular generation, the signal and congestion.

-no reliable figure

It is real and it is variable, and quoting a single mobile penalty would be inventing precision that does not exist.

Examples

In practice

A deployment serving Indian visitors routes recognition and synthesis through a European region and the language model through a North American one, so every turn pays an intercontinental round trip twice. Consolidating all three into an Asian region strips several hundred milliseconds off the wait without touching a single model or a line of application logic. The largest latency improvement in the project turned out to be a configuration change.

The everyday version

Round trip time is how long it takes a signal to reach a computer somewhere else in the world and come back. It is set by physics, so no clever software makes it faster. If your customers are in India and your system runs in America, every exchange pays that cost, on the way there and on the way back.

Usage

Who says it

  • Network and infrastructure engineers, routinely, and with specific routes in mind.
  • It drives architecture decisions about which regions to deploy in.
  • Rarely a buyer term, though it sits underneath any question about regional availability.

Where it turns up

  • Next to point of presence locations, regional availability, data residency and latency commitments.
  • A spec that promises a latency target without saying which region serves which audience is promising something it cannot hold for a distant visitor.

Common misuse

  • Assuming network latency is negligible because it is negligible on the office network.
  • Testing latency from a location close to the servers and generalizing the result to everyone.
  • Overlooking that a multi-region pipeline pays the round trip more than once per turn.

Questions people ask

What is round trip time?

Round trip time is how long a single packet takes to travel to a server and come back. It is the network floor beneath every other latency number in a voice system, set by distance and physics, and no amount of software optimization goes below it.

Why can't software make round trip time faster?

Because it is governed by the speed of light in fiber and the route the traffic takes, not by code. Light travels through optical fiber at roughly 200,000 kilometers a second, which works out to about 10 milliseconds of round trip for every 1,000 kilometers. You can shorten the distance by moving the server closer to the visitor, but you cannot make the signal travel faster than that.

How much latency does distance add to a voice agent?

A Mumbai visitor reaching a US east coast server covers roughly 13,000 kilometers, which is a theoretical minimum round trip of around 130 milliseconds before any routing, queuing or processing. Against a sub-second time to first audio budget, and a human turn gap that averages about 200 milliseconds, that fixed cost is paid on every exchange and it is substantial.

Does running my voice agent in one region hurt visitors elsewhere?

Yes, and the further away they are, the more it costs them. A pipeline that calls recognition, the model and synthesis in different regions pays the round trip on every hop. Consolidating those services into one region near the audience is frequently the largest latency saving available, and it needs no model change. The catch is data residency, which can force processing into a specific region regardless of where visitors are.

Share this definition

Last reviewed 31 July 2026. Written and reviewed by Dhruv Dholakia, founder of AsqVox.