Skip to content
Part 3, chapter 1 · Voice Agents · 2026-08-03

How AI Voice Agents Work, Layer by Layer

Speech in, text, a decision, speech out, down a phone line. Every voice agent is those four layers, and the hard part is not any of them individually.

A voice agent is four systems in a loop. Audio arrives, something transcribes it, something decides what to say, something turns that into speech, and the audio goes back. The components are commodities. The loop is not.

LayerJobTypical latency
TelephonyCarry audio to and from the phone network20–100ms
Speech to textTranscribe the caller, streaming100–300ms
Language modelDecide what to say300–800ms
Text to speechSynthesise the reply80–200ms

The number that decides whether it sounds human

In natural conversation the gap between one person stopping and the next starting is around 200ms. Push past roughly 800ms and callers start doing what people do on a bad line — repeating themselves, talking over the agent, or assuming the call dropped.

Add the layers up and the budget is tight before anything goes wrong. That is why the category obsesses over latency, and why platforms advertise time-to-first-token rather than mouth-to-ear: the first number is flattering and the second is what the caller experiences.

Time to first byte is not what the caller hears. Mouth-to-ear includes endpointing — deciding the human has finished, plus the model, plus synthesis, plus the network in both directions. It is routinely three to five times the quoted figure, and no vendor publishes it.

The actually hard problem

Not any single layer. It is knowing when the human has stopped talking.

Wait too long and the agent feels slow and stupid. Cut in too early and it interrupts someone mid-sentence, which is worse, an agent that talks over you reads as broken in a way a slow one does not. There is no correct threshold, because the right answer depends on the caller, the phone, the background noise and what was just asked.

A pause after "my account number is..." means keep waiting. The same pause after "yes" means reply now. Systems that handle this well are using semantic cues, not just silence duration, and this is where the difference between platforms actually shows up.

Barge-in

When the caller starts talking while the agent is speaking, the agent must stop immediately, discard what it was going to say, and process the interruption. Getting this wrong produces the single most frustrating failure in the category: an agent that keeps reading its script while a human is trying to correct it.

Speech to speech, and the trade

Newer models collapse transcription, reasoning and synthesis into one hop, removing conversions and cutting latency meaningfully.

The trade is control. You cannot swap a cheaper component when one ships, you do not get a clean transcript for free, and debugging is harder because there is no intermediate state to inspect. For compliance-heavy deployments that need verbatim records, the assembled pipeline is still the safer choice.

What a minute of all this costs, with each component priced from its vendor's own published rate, is on the cost model.

Questions

How much do AI voice agents cost?
Advertised platform rates run $0.02 to $0.14 a minute, and none of those is a unit cost. Computed all-in on published list rates, most land between $0.065 and $0.125. The cost model shows every component with the date we read it.
What is a good latency for a voice agent?
Under about 800ms mouth-to-ear. Natural conversational gaps are around 200ms, and past 800ms callers start repeating themselves or talking over the agent.
Why do voice agents interrupt people?
Endpointing thresholds tuned too aggressively. The system decides silence means the caller finished, when they were pausing mid-sentence. Good systems use semantic cues rather than silence duration alone.
What is barge-in?
The caller talking while the agent is speaking. The agent must stop immediately and process the interruption. Handling it badly is the most frustrating failure mode in the category.
Is speech-to-speech better than a pipeline?
Lower latency, less control. You cannot swap components, you do not get a free transcript, and debugging is harder. For compliance-heavy use cases the assembled pipeline is usually still correct.
Which layer costs the most?
They are closer than people expect, the four land within about two cents of each other per call minute. Synthesis is the one with real range, from $0.014 to $0.085 depending on the voice.
Can I build this myself?
Yes, with frameworks like Pipecat or LiveKit Agents. What you are taking on is turn-taking and interruption handling, which the hosted platforms have spent years tuning and you will not match in a sprint.

Tools mentioned

Sources

Source interests are labelled. Almost everything published about this subject is written by someone selling into it.

← All chapters