voice ai cost

What does background speech cost a voice agent?

A voice agent is billed for every second it listens. In a room with more than one person, most of what it hears was never meant for it. It still pays to transcribe and reason over all of it.

~86%

of the audio a voice agent processes in a multi-talker setting was never addressed to it. Estimate.

The fix is not a smarter model. It is deciding what deserves the model at all.

Last updated June 30, 2026

See the cost model Request an eval

the billing clock

The meter starts when audio moves, not when someone speaks to you.

The standard voice pipeline runs in four steps: voice activity detection, speech to text, the language model, then speech back. Voice activity detection finds that someone is speaking. It does not find who they are speaking to. Every voice that clears it is forwarded to speech to text, billed by the second, and on to the language model, billed by the token. Nowhere in those four steps does anything ask the one question that decides whether the spend was warranted: was this addressed to me?

A voice pipeline with the billing meter on the wire Two input streams, one addressed and one background, both clear voice activity detection and flow identically through speech to text, billed per second, and the language model, billed per token, then to speech back and a reply. No stage asks whether the audio was addressed to the device. addressed background VAD speech? STT per second $ billed LLM per token $ billed TTS speech back reply no step asks addressed to me?
A bystander, a second customer, a TV in the background. Voice activity detection forwards them all, and the meter runs the same as it does for a real order.

a worked receipt

One drive-thru, priced from public rates.

Every input below is a public list price or a labeled estimate. Change the inputs to match your own deployment and the shape of the answer holds.

Input or output Value Basis
Inputs
Orders per location per day 400 estimate
Active audio per order 1.5 min estimate, conservative high end
Pipeline cost per minute $0.114 public list prices, speech to text plus model plus speech back
Audio not addressed to the agent ~86% estimate, built bottom up
Result, per location
Spend per year, no gate ~$24,966 estimate
Spend per year, with a pre-ASR gate ~$3,495 estimate, only the ~14% addressed audio is processed
Recovered per year ~$21,471 estimate, 86% of spend
Result, at fleet scale
500 location chain, recovered per year ~$10.8M estimate
US drive-thru industry, recoverable ~$1.8B estimate
Spend per location per year, with and without a pre-ASR gate A tall bar of about $24,966 with no gate sits beside a short accent bar of about $3,495 with a pre-ASR gate, on a shared dollar axis. The difference, about $21,471 or 86% of spend, is recovered. $25k $20k $10k $0 ~$24,966 no gate ~$3,495 with a pre-ASR gate recovered ~$21,471 86% of spend
Per location per year, priced from public rates. Estimate.
An estimated $25,000 per location per year, near $10.8M for a 500 location chain, is spent transcribing and reasoning over audio no one directed at the agent.

Receipt assumptions are stated above and adjustable. Dollar figures are estimates built on public per-minute pricing, June 2026. The largest source of uncertainty is orders per day.

send us the audio

The new audio-native models do not remove this cost. They raise it.

Frontier labs have folded speech to text, reasoning, and speech back into one model that takes raw audio and bills per audio token. Audio tokens are priced far above text tokens. On one major realtime API the audio input rate is about thirteen times the text rate. The model bills on every second of audio it receives, addressed or not, because the model itself is now doing the listening. Measured across thousands of production sessions, an active call runs about $0.18 to $0.46 per minute. You are paying the most expensive layer in the system to perform its cheapest task: deciding whether anyone is talking to the device.

You do not invoke the model to do the hearing.

A cheaper model does not fix it. The price per token falls, the bill on every background second remains. Even a model that perfectly ignores background speech still bills every token it receives, because billing happens at ingestion, not at reply.

Cost per minute of a gated path versus an audio-native path, and the audio token premium On a shared baseline, a gated path bills about $0.006 per minute on roughly 14 percent of audio, while an audio-native path bills $0.18 to $0.46 per minute on 100 percent of audio. Below, an audio input token is priced about thirteen times a text token on one public realtime API. $ per minute, on the audio each path processes gated path STT on ~14% of audio ~$0.006 / min audio-native billed on 100% of audio $0.18 to $0.46 / min audio input token rate vs text, one public realtime API text token 1x audio token ~13x
Public list prices, June 2026. The model bills on every second, addressed or not.

Public list prices, June 2026, for streaming speech to text and for audio-native realtime APIs. Figures are sourced and adjustable; see the receipt above for method.

the oldest solution

Your brain hears everything. It does not think about everything.

The brain is about 2% of body weight and burns about a fifth of your energy. It cannot afford to fully process every sound, so it does not. Cheap, fast filters decide what reaches the expensive machinery. The same design decides what should reach a language model. The parallels below are illustrations, not a claim that any product reproduces neuroscience.

early selection

A filter before meaning

You can follow one voice in a crowded room from its pitch and direction, before your brain has decoded every competing sentence. A pre-ASR gate is that filter. It reads cheap acoustic and contextual cues and decides whether the language model should run at all.

sensory gating

A gate below the expensive part

Sensory signals are filtered before they reach the cortex. A cheap, always-on gate protects the costly machinery from noise it never needed to evaluate. The addressee decision plays that role for the model.

predictive coding

Spend on surprise, not on the expected

The brain suppresses the predictable background and elevates only what is unexpected and relevant. Speech addressed to the device is the signal worth elevating. Ambient talk is the background to drop.

fast, then slow

A fast system that wakes the slow one

A fast, automatic system runs first and decides when the slow, deliberate, costly system is worth waking. Routing all audio into the frontier model wakes the expensive system for every sound in the room.

The brain never sends every sound to its most expensive layer. A voice agent should not either.

Routing every sound to the expensive layer versus gating first, then spending Top: many inputs all converge on one expensive layer that fires on every sound. Bottom: a cheap, always-on gate drops most audio so it is never billed, and passes only the warranted signal on to the expensive layer. ROUTE EVERYTHING TO THE MODEL expensive layer fires on every sound GATE FIRST, THEN SPEND cheap gate expensive layer only when warranted most audio dropped, never billed
The brain filters cheaply before it spends. Illustration, not a biological claim.

the addressee gate

Put one decision in front of the stack. Was this addressed to me?

Selective Auditory Attention (SAA) adds a single decision above the audio layer. For every utterance it answers one question, is this speech addressed to the device, before speech to text billing starts. Audio that was not addressed is dropped. Audio that was addressed passes through to your existing speech to text and language model, unchanged. The decision is hosted and runs in real time, before speech to text. It is addressee detection, not speaker identification. It never asks who is speaking, only whether they are speaking to the device.

It sits above any speech to text and any model, including the audio-native ones, and makes any of them cheaper.

In our research the decision runs ahead of billing and fails closed when the signal is ambiguous (arXiv:2604.08412). Cross-lingual recall is a known limitation under active work.

The addressee gate above the stack, routing addressed audio to the pipeline and dropping the rest Audio from all voices enters the addressee gate. If addressed, it passes to speech to text, the language model, speech back, and a reply. If not addressed, it is dropped at zero cost and never billed. The gate sits above any speech to text and any model. audio in all voices addressee gate addressed? hosted, real time YES STT LLM TTS reply NO dropped $0, never billed works above any speech to text and any model
One decision, before the meter starts. Addressee detection, not speaker identification.
the core argument

The voice agent is paying to listen to a room that was never speaking to it. The fix is one decision, made before the meter starts.

measure it

Measure it on your own audio.

Send a sample of your real acoustic environment. We return the addressee decision rate and the share of audio a gate would have dropped before it reached your stack. Free evaluation kit.

We do not share your information. See our privacy policy.

common questions

What voice AI teams ask about background speech and billing.

Why do voice agents pay for audio that was never addressed to them?

The standard pipeline uses voice activity detection to decide when to send audio onward. It detects that speech is present, not who the speech is for. In a multi-talker setting, bystanders, background media, and side conversations all clear it and are billed by speech to text and the language model the same as a real command. No step in the standard pipeline asks whether the speech was addressed to the device.

How much does background speech cost at scale?

At public streaming speech-to-text rates each unfiltered minute costs a fraction of a cent, but a busy multi-talker location accumulates hundreds of non-addressed minutes a day. The worked receipt on this page estimates about $25,000 per location per year, and about $10.8M for a 500 location chain, on audio no one addressed to the agent. These are estimates built from public pricing.

Do audio-native realtime APIs fix this?

No. They fold speech to text and the model into one service billed per audio token, and audio tokens are priced well above text tokens. The model bills on every second of audio it receives, addressed or not, so routing all audio to it raises the cost of background speech rather than removing it.

Is this speaker identification?

No. SAA answers whether speech was addressed to the device, not who is speaking. It does not perform speaker identification, diarization, or voice biometrics.

Does noise cancellation solve it?

No. Noise suppression cleans the audio signal. It does not decide who the clean speech was for. A denoised side conversation is still billed end to end. The addressee decision is a separate layer.