Voice Agent Testing with Realistic Background Noise

Mai Medhat

CEO & Co-founder, Tuner

One of our most requested features is here. You can now add background noise to Tuner call simulations and test how your voice agent handles calls from a café, a busy street, or an office.

Our customers wanted to answer a simple question: my agent works in a quiet room, but can it still do its job when the caller is somewhere noisy?

Nobody calls from a quiet room

We've all heard it: Voice agents work in demos but struggle in production. There are plenty of reasons for that. One that's hard to test is where your customer is calling from.

When we test an agent, we usually sit in a quiet room and speak clearly. Real customers call from the street, a coffee shop, or a car with kids screaming in the backseat. They're trying to book an appointment or change a delivery address while getting on with their day.

Your agent needs to handle those calls too. And until you test those conditions, you don't know how well it will.

Same scenario, different surroundings

With background noise in Tuner, you can take a scenario you already test and run it in different environments.

  • Choose the environment: A cafe, street, office or car, each a real recording rather than synthetic hiss.

  • Set the intensity: Low, medium, or high, so you can turn up the pressure until something breaks.

  • Keep it present throughout: The sound continues while the caller speaks and during pauses, the way a real call sounds.

  • Use any language: It works with every language or accent Tuner supports, not just English.

Start with a task your agent handles well in a quiet setting. Add background noise, then turn it up and see what changes.

Does the agent still understand the caller? Does it give them time to finish? Can it complete the same task?

The failure you can't hear

Imagine a customer calling to change their delivery address. They say "fifteen," but with traffic in the background, the agent hears "fifty." Speech recognition doesn't crash, it just guesses wrong, and the agent trusts the guess. It updates the address and confirms the change.

The conversation sounds fine. The call is marked as completed. But the delivery is going to the wrong place.

This is the failure worth worrying about. Not the call that falls apart, that one is obvious and you'll catch it. The dangerous one is the call that sounds successful, passes your checks, and quietly does the wrong thing. You don't find out until the package is at the wrong house and the customer is on the phone again.

That's what testing with background noise helps you catch before launch: the misheard number that still gets confirmed, the agent that acts confidently on the wrong detail. It's why it matters to look beyond whether an agent can hold a conversation. Did it actually do what the customer needed, correctly?

Noise can cause more than misheard words. It also affects the parts of your pipeline that decide when to talk (VAD, endpointing, and barge-in). An agent might interrupt someone before they finish, take too long to respond, or keep asking the caller to repeat themselves.

How it handles those moments matters. If an address or payment amount is unclear, asking a specific question and confirming the answer can prevent a mistake. Repeating "sorry, I didn't catch that" three times doesn't help the caller move forward.

Find it, fix it, run it again

After running a simulation, listen to the call and check the transcript. Use Tuner's traces and evaluation results to understand where things went wrong.

Did the agent miss a number? Did it start speaking too soon? Did it hear the right words but take the wrong action?

Those are different problems, and they need different fixes. Once you've made a change, run the scenario again with the same environment and noise level to check whether it helped.

The goal is to build an agent that can handle everyday conditions and ask for help when it's unsure.

Why we built it

Background noise was one of our most requested features because teams wanted to test the kinds of calls they expected their agents to handle every day.

It's now live in Tuner. You can include it in your tests before launch and run those tests again whenever you update your agent.

Take a scenario your agent already passes. Add the background noise your customers are likely to call from. See whether it still gets the job done.

Try background noise in Tuner →

Tuner helps voice AI teams test, monitor, and debug calls before issues reach users.