Latency Problems

What normal response time looks like on Xebit, what adds latency, and how to find and reduce it.

Showing instructions for Xebit VoiceCloud.Showing instructions for Xebit VoiceKit, on your own infrastructure. Switch with the selector at the top. Updated 22 October 2026.

Latency is the time between the caller finishing a sentence and the agent starting to answer. On the default speech to speech path, Xebit responds in under 500 ms. If you run a cascade with your own speech to text and text to speech components, response time depends on the components you bring. When latency is consistently higher than your baseline, something in the path is adding time, and Observability shows you which part.

Measure it first

  1. Sign in to the Xebit VoiceCloud dashboard at voice.xebit.ai.
  2. Sign in to the Xebit VoiceKit admin console on your own host.
  3. Open Observability and look at response time per turn for the affected agent. See Observability.
  4. Compare live calls with a test call from the agent's settings. If tests are fast and live calls are slow, look at load, the network or your systems rather than the agent's instructions.

What causes high latency

CauseWhat it means
Long or complex instructionsMore to process on every turn
Wrong language settingThe agent works harder to understand the caller
Slow workflow actionsThe agent is waiting for one of your APIs
Cascade componentsEach speech to text, model and text to speech step adds its own time
Provider degradationA model or speech provider is slow or timing out
LoadMore concurrent calls than the capacity behind them
NetworkDelay on the path between your carrier or SBC and Xebit

How to reduce latency

Simplify the instructions

Keep instructions short, clear and direct, and avoid deeply nested conditions. Open the agent's settings to review them. See System Prompt.

Check the language setting

Make sure the agent's language matches what your callers speak. See Language Issues.

Speed up workflow actions

Find slow actions in the call's event log and fix the API behind them, or fetch the data before the call. Give the agent a filler phrase for lookups that cannot be avoided. See Delayed Responses.

Check capacity

If a large campaign pushes past your concurrency, calls wait for a free slot. See Rate Limits and talk to us if you need more.

On Xebit VoiceKit, latency depends on your own GPUs or model provider and on node capacity. Check per node metrics in Observability for saturated nodes or GPUs, and your model provider's rate limits. Add capacity, or lower campaign pacing until you do.

Watch for provider incidents

When a model or speech provider degrades, it shows up as an Observability incident, with the agents affected and when it started. The SRE agent writes the root cause analysis and proposes a fix, such as moving to another provider, for an engineer to approve.

Still need help?

Email hello@xebit.ai or use the contact page, with the call ID of a call with high latency.

Support is included in your Xebit VoiceKit licence.

Next steps