Where a voice agent's reply time goes

Ask any voice AI vendor about speed and you will get a number. Before trusting it, including ours, ask what the clock was measuring.

The measurement we use

We start the clock at the caller's last voiced audio and stop it at the agent's first audio. That is the silence a person actually sits through. A figure that starts when the transcript arrives leaves out the most variable part of the wait.

Our engine logs this timing for each reply, so we can report medians and slower cases from real traffic on a deployment rather than from a single good demo.

What we have measured so far

In test calls on our own test setup in November 2026, the median for English was about 1.7 seconds. Be careful with that figure. Those calls ran over a WebSocket connection with no phone network involved. A real carrier path adds its own delay, and we have not yet published a phone-network figure or a Hindi figure. We will not quote one until we have measured it.

Where the time goes

Three stages take most of it:

• Knowing the caller has finished. The speech provider has to wait through some silence before it can be confident. On some models and languages this is the largest single part of the delay.
• The model's first words. With a fast model we have seen about 0.6 seconds to the first token. Slower or heavily loaded models take longer, and a long prompt does not help.
• The first audio. Text-to-speech needs to produce enough sound to start playing.

What we changed in the engine

Two choices made the largest difference.

Drafting during the hold. While the speech provider decides whether the caller has finished, the engine already asks the model for a reply and prepares the first audio. Nothing plays until the turn is confirmed. If the caller keeps talking, the draft is thrown away and isn't billed. This took about 700 ms off the median in one of our test sets.

Speaking the opening clause early. The engine doesn't always wait for a whole sentence. Once a long enough opening clause is ready, it can go to speech straight away.

What you control

Your configuration moves this number more than any engine trick:

1. Choose a faster language model for simple calls. Keep the larger ones for calls that need them.
2. Keep prompts tight. Every extra instruction is more for the model to read.
3. Watch your tools. A slow order lookup is silence on the line, however quick the voice pipeline is.
4. Test speech models in your callers' language, because turn detection varies between them.

How to test a vendor's claim

Whoever you are evaluating, including us, make ten calls over the real phone path you will use. Include a long sentence, a short "yes", and a question that needs a tool. Time from your last word to the first sound. Look at the slow calls as well as the median, because callers remember the slow ones.

The order-status workflow is a good test because it needs a real lookup. Costs for faster models are visible on the pricing page, and we are happy to run this test with you.