The ITU-T G.114 Standard
The most widely cited benchmark for acceptable call latency comes from ITU-T Recommendation G.114, the International Telecommunication Union standard covering one-way transmission time for voice calls. It defines three bands for one-way, end-to-end delay: under 150 milliseconds is acceptable for most user applications, 150 to 400 milliseconds remains acceptable but quality depends on how the network is planned and provisioned, and anything above 400 milliseconds is generally unacceptable for network planning purposes. G.114 has been in place for decades and predates modern VoIP, yet it is still the baseline reference telecom engineers and voice platforms cite when evaluating whether a connection is usable for real-time conversation.
One-Way Delay vs. Round-Trip Delay
G.114 measures one-way, mouth-to-ear delay: the time it takes a spoken word to travel from one person's mouth to the other person's ear. That is a different number from round-trip latency, which is what many network tools actually report, and it is easy to confuse the two when reading a spec sheet. A call sitting right at the edge of G.114's 150ms threshold in one direction can still feel smooth, because the standard already accounts for the natural back-and-forth rhythm of conversation. But if a system reports 150ms round-trip, the one-way component is roughly half that, so understanding which number a report describes is necessary before judging whether a call setup meets the standard.
Why VoIP Calls Are More Prone to Latency Than PSTN
Traditional PSTN calls run over a fixed, dedicated circuit, so delay is minimal and predictable for the entire duration of the call. VoIP calls travel over shared, best-effort IP networks, and that difference introduces several additive sources of delay that a circuit-switched line never had to deal with. Codec processing adds encode and decode time, and higher-compression codecs designed to save bandwidth take longer to process than simpler ones. Jitter, the variance in how quickly packets arrive relative to one another, requires a jitter buffer to smooth playback, and that buffer itself adds delay on top of transmission time. Packet loss is another factor with no PSTN equivalent: IP packets can be dropped in transit, and loss rates above roughly 20% degrade call quality sharply rather than gracefully. Finally, VoIP traffic can take variable routing paths across the internet, introducing non-deterministic delay that a dedicated PSTN circuit simply does not experience. Each of these sources is small on its own, but they stack, which is why a VoIP call can drift into G.114's 150-400ms band, or beyond it, even when the underlying internet connection looks fine on a basic speed test.
What This Means for Voice AI Agents
G.114's 150-400ms band was written for human-to-human calls, where transmission delay is the only variable. Voice AI agents have a narrower effective budget, because network transmission latency is only part of the delay a caller perceives. On top of that transmission time, an AI agent has to run speech-to-text, feed the transcribed input to a reasoning model, generate a response, and convert that response back to speech, and all of that processing has to complete inside the same window a human caller would judge as a natural pause rather than a hang. That is a meaningfully harder constraint than the general VoIP case this standard was built for. Readers who want the technical detail on how that processing pipeline is compressed to stay inside a natural conversational rhythm can see our breakdown in Sub-100ms responses: inside our voice architecture, which covers the response-generation side of this problem rather than the general transmission standard.
Key takeaways
- ITU-T Recommendation G.114, the international standard for one-way transmission time, sets three bands: under 150ms is acceptable for most applications, 150-400ms is acceptable but depends on network planning, and above 400ms is generally unacceptable.
- These thresholds measure one-way, mouth-to-ear delay, not round-trip time, so a call that feels laggy in conversation may already be double the one-way figure.
- VoIP calls are more prone to noticeable latency than traditional PSTN calls because delay compounds from codec processing, jitter buffers, packet loss, and variable network routing, none of which exist on a dedicated circuit-switched line.
- Voice AI agents work inside an even tighter budget than G.114 allows, since speech-to-text, model reasoning, and text-to-speech all have to complete inside the same delay window a human caller perceives as natural.
