Voice AI Latency: What's Fast Enough and How to Cut It

Voice AI latency is the delay between the moment a caller stops speaking and the moment the agent's first audio plays. People leave about 200 milliseconds between turns on average. Benchmarks published in 2026 put default voice agent stacks at roughly 1.3 to 1.7 seconds median, measured from the caller's side. The figures depend heavily on where each vendor starts and stops the clock.

How fast do people take turns?

The benchmark that most people cite comes from a PNAS study by Stivers and colleagues (2009). Across a sample in 10 languages, the mean gap between speakers was +208 milliseconds. The most common gap, the mode, was 0 milliseconds, and the median was +100 milliseconds. Each language's mean sat within about 250 milliseconds of the cross-language mean. The study also found that a 200-millisecond silence was judged a delay in most languages, yet still considered on time.

That study describes people, not machines, but it gives a target. Coval's latency guide offers its own thresholds. It says under 1 second feels natural, 1 to 2 seconds is acceptable, and over 3 seconds is noticeably poor. Its use-case targets are:

• Consumer service: under 2 seconds, acceptable under 3 seconds.
• Real-time conversational products: under 1 second, acceptable under 1.5 seconds.

These are one vendor's opinions, not an industry standard. Hamming's glossary is stricter about the median. It calls a P50 turn latency under 1,600 milliseconds excellent, under 1,800 good and under 2,000 acceptable.

Where the delay comes from

Latency is a sum of stages. The table sets Coval's component estimates beside Hamming's budgets. Both are vendor estimates, so use them as planning ranges, not measurements.

| Stage | Coval estimate | Hamming budget | |---|---|---| | Speech recognition after speech ends | 200 to 500 ms streaming; 500 to 1,500 ms batch | 100 to 300 ms | | Language model | 400 to 800 ms fast; 800 to 1,500 ms standard; 1,500 to 3,000 ms large | 200 to 800 ms | | Speech synthesis to first audio | 200 to 400 ms fast; 400 to 800 ms standard; 800 to 1,500 ms high quality | 100 to 400 ms | | Tool and API calls | 50 to 200 ms per call; webhooks 100 to 500 ms; third-party integrations 200 to 2,000 ms | Not given | | End to end | Not given | 800 ms to 2 s typical |

The telephony leg is usually small by comparison. TECHSY measured carrier round trips with a median of 71 milliseconds for Telnyx and 89 milliseconds for Twilio, in a roundup published by Telnyx. That is a carrier measurement, not an end-to-end call time.

Real benchmark numbers

Speech synthesis is one stage that vendors often publish a single number for. A benchmark run by Coval, and republished by Gradium, which is itself a vendor in the table, measured perceived time to first audio on September 8, 2026. It used 480 samples per model, English only, and excluded the connection handshake.

| Model | P50 (ms) | P95 (ms) | |---|---|---| | Inworld TTS Flash 2 | 75 | 111 | | ElevenLabs Flash v2.5 | 185 | 225 | | Gradium TTS | 214 | 346 | | Cartesia Sonic 3.5 | 269 | 346 | | Deepgram Aura-2 | 290 | 522 | | ElevenLabs Eleven v3 Conversational | 320 | 393 | | Cartesia Sonic 3.6 | 440 | 696 |

Rankings change, so quote this as a snapshot. For native speech-to-speech models, Artificial Analysis lists first-audio times from 0.70 seconds for Grok Voice 2.0 and 0.81 seconds for Grok Voice 3.0, up to 1.34 seconds for GPT-Live-1 and 1.35 seconds for Gemini 3.8 Live at its extended thinking setting.

The platform roundup from Telnyx adds a third view. It combines an Openbenchmarks run, which measured caller-side time to first audio from call recordings, with Tested Media's voice-to-voice medians. Telnyx is itself one of the platforms in the table.

| Platform | Openbenchmarks median (ms) | Openbenchmarks p95 (ms) | Tested Media median (ms) | |---|---|---|---| | Telnyx | 1,296 | 1,856 | Not tested | | ElevenLabs | 1,424 | 1,768 | Not tested | | Bland | 1,520 | 2,248 | 850 | | Vapi | 1,558 | 2,008 | 720 | | Retell | 1,740 | 2,259 | 680 |

The Openbenchmarks endpointing was pinned to 0.1 seconds for Telnyx, Vapi and Retell only, and each platform ran its default model stack. Tested Media does not publish how it detected events, and its medians sit 670 to 1,060 milliseconds below Openbenchmarks' figures for the same three vendors. Read the table as evidence of method sensitivity, not as a ranking.

Why vendor and measured numbers differ

Four things usually explain the gap.

• Start and stop points. ElevenLabs lists Flash v2.5 at about 75 milliseconds, with a footnote excluding application and network latency. Coval measured the same model at a 185-millisecond median, using a clock that includes leading silence.
• Recordings or self-reports. The Telnyx roundup found that self-reported platform latency ran about 490 milliseconds below figures derived from call recordings.
• Model settings. Thinking modes change speed inside one family, as the Gemini figures above show.
• Price is not speed. Retell lists the cost of each language model per minute but gives no latency for them. Do not infer speed from price. The API comparison lists prices, not speeds, for each provider.

How to cut voice agent latency

1. Stream everything. Pipecat's stages run in parallel, so the reply can start while the rest is still being generated.
2. Tune end-of-turn detection. OpenAI's silence_duration_ms sets how much silence ends a turn, and its documentation says shorter values detect turn ends faster. LiveKit's OpenAI telephony example uses 400 milliseconds as an example, not a default. Shorter values also raise the risk of splitting a transcript early, which is why OpenAI offers semantic_vad.
3. Consider eager turns, and their cost. Deepgram Flux can emit an early end-of-turn so the language model starts sooner. Its documentation warns that this can increase LLM API calls by 50 to 70 percent because of speculative responses.
4. Shorten context and pick a smaller model. OpenAI says smaller context windows reduce the input sent each turn, and that keeping history static helps caching. Its cost guide gives no latency figures, so measure the effect yourself.
5. Remove a hop where you can. Gemini and OpenAI offer audio-to-audio paths that skip separate recognition and synthesis. Microsoft says chaining separate modules "can lead to increased engineering complexity and end-user perceived latency."

Latency is not the only goal

Faster turn-taking can cause false endpoints, where the agent answers before the caller has finished. LiveKit makes its turn-detector model the default and reserves VAD-only mode for minimal latency or for languages the detector does not cover. Eager processing also costs money, as the Deepgram warning shows. Interruption handling matters as much as speed, because a caller who talks over an agent expects it to stop.

Pre-recorded prompts skip synthesis while the caller waits

This point is general reasoning, not a sourced fact. Any line that does not change between calls can be generated ahead of time as an audio file. Then no synthesis happens while the caller waits, and the latency budget is spent only on the variable parts of the call. OndaLabs can produce those files with the text to speech studio, and the pricing page shows the credit costs. OndaLabs makes no speed claim for its generation and is not a real-time service.

Frequently asked questions

What is a good latency for a voice AI agent?

There is no single agreed number. Coval says under 1 second feels natural for real-time conversation and under 2 seconds suits consumer service, while Hamming's glossary treats a P50 turn latency under 1,600 milliseconds as excellent. Pick a target from your use case, and measure it the way a caller experiences it.

How fast does a conversation feel natural?

Human speakers leave about 200 milliseconds between turns on average, with a mean of +208 milliseconds across the 10-language study. The same study judged a 200-millisecond silence a delay in most languages, but still on time. Voice agents usually take far longer, so most of the design work goes into closing that gap.

What is time to first audio?

Time to first audio, or TTFA, is how long a system takes to produce its first audio output. Artificial Analysis defines it for speech models as the seconds needed to generate the first token of audio output. Vendors and benchmarks start the clock at different points, so compare figures only within the same test.

Why do vendor latency numbers differ from what I measure?

Vendors start and stop the clock in different places, and some exclude network and application time. Benchmarks also differ in how they detect events, whether they use recordings or self-reports, and which model settings they use. Measure end to end from the caller's side, on your own calls, with the same settings for every provider.

Does telephony add latency?

Some, but the carrier leg is usually small. TECHSY measured median carrier round trips of 71 milliseconds for Telnyx and 89 milliseconds for Twilio. That covers the carrier leg only, so measure the whole path on your own calls.