A voice agent that takes a second and a half to answer sounds like a dropped call. At Coru the response arrives in under 800 ms after the customer stops speaking, and that threshold separates a call that ends with a payment commitment from one that ends with the customer hanging up.
The supporting data is recent. An analysis published in January 2026 covering more than a million voice agent calls found that every 100 ms of latency above 800 ms cuts task completion by 4 to 6 %, and that above 1,200 ms abandonment rises by 15 to 25 points (Hamming, 2026). For financial services and healthcare, the same study puts acceptable abandonment below 3 %.
Why the threshold sits at 800 ms
Human conversation has a measured rhythm. Research published in PNAS across ten languages found that the gap between one turn ending and the next beginning is around 200 ms, with cultural variation but almost always under half a second (Stivers et al., cited by Telnyx). A longer pause reads as hesitation, confusion or failure. From 800 ms the silence becomes noticeable; past 1,500 ms, the customer assumes the line dropped and talks over the agent, repeats themselves or hangs up.
In collections the effect is immediate. The customer already comes into the call uncomfortable, and every silence gives them a reason to end it. In origination the loss has another name: the sale that never closes.
Where the milliseconds go
Every voice turn passes through five stages, and each one adds time. These are typical ranges for voice agents on the market; the order of magnitude matters more than the exact figure.
| Stage | What it does | Typical cost | Where time is lost |
|---|---|---|---|
| End-of-speech detection (VAD) | Decides the customer has finished | 100 to 300 ms | Conservative thresholds that wait “just in case” |
| Transcription (STT) | Turns audio into text | 150 to 400 ms | Models that wait for the full sentence instead of streaming |
| Reasoning (LLM) | Decides what to say and which tool to call | 300 to 900 ms | Long prompts, oversized models, external lookups run in sequence |
| Speech synthesis (TTS) | Turns text into voice | 100 to 300 ms | Waiting for the full text before starting to speak |
| Network and telephony | Carries the audio back to the phone | 50 to 200 ms | Chained providers, each in a different region |
At the low end, the total is already close to 800 ms. At the high end, it passes two seconds. That is why a slow agent rarely has one slow component: it usually has five reasonable components chained in sequence, each waiting for the previous one to finish completely.
How to get under 800 ms
The gain comes from changing how the stages are chained, more than from tuning any single one.
- Stream everything. Transcription runs while the customer speaks, the model starts generating from the first transcribed phrase and synthesis starts speaking with the first generated sentence. No stage waits for the previous one to finish.
- Natural interruption (barge-in). Voice activity detection runs in parallel while the agent speaks. If the customer cuts in, the agent stops and listens, instead of finishing an eight-second sentence while the customer repeats that this is not what they asked.
- An owned voice stack. When transcription, model and synthesis are three providers in three regions, the network adds more than any model. Coru runs its own voice layer: the media server and the audio bridge sit on the same private network, and the language model can be swapped without changing the audio path.
- Pronunciation solved before speaking. A rate, an ID number or an amount read wrong forces a repeat, and a repeat costs a full turn. Each agent’s pronunciation dictionary resolves industry acronyms and figures before they reach speech synthesis.
- Latency watched in production. Operations dashboards show voice performance by agent and by channel, with alerts when a voice provider degrades and failover to an alternative voice.
The result in production: under 800 ms from the moment the customer stops speaking to the agent’s first audio, over the bank’s own SIP trunk, on inbound calls and outbound campaigns.
Three questions for any voice vendor
- What is the per-turn latency measured in production, outside the demo, and where can I see it?
- What happens when the customer interrupts mid-sentence?
- How many providers does the audio pass through before it returns to the phone?
If the first answer is an average, the second is that the agent finishes its sentence before listening and the third mentions three or four providers, the call will be shorter than you planned.
Hear how a Coru agent responds. The voice section of ADI Platform shows the full stack, from SIP trunk to conversation turn. See the voice layer or book a demo with your own use case.
Sources: Hamming, Voice Agent Drop-Off Analysis, January 2026 · Telnyx, Voice AI agents compared on latency, 2026