How to Make GoHighLevel Voice AI Sound Human: Interruption Sensitivity, Backchanneling and Noise Cancellation

You build a Voice AI receptionist, write a decent prompt, connect the calendar, and then the first test call goes like this. You start to answer and it keeps talking over you. You pause to think and there is dead air, as if nobody is listening. A car goes past your window and the agent replies to a question you never asked. The words are fine. It just does not feel like talking to a person.

In my experience the prompt is usually not the problem. It is three settings that most people never open, because HighLevel hides them under advanced settings to keep the builder simple. This guide covers what each one does according to HighLevel's own help docs, where to find it, how I would set it for a typical small business receptionist, and how to test it properly. I also cover the other settings and prompt habits that make a voice agent sound more natural, and the gotchas that catch people out. Every fact about the settings comes from HighLevel's help centre, its Ideas Board and its official YouTube channel, all linked at the bottom.

Why GoHighLevel voice agents sound robotic

A phone call is a lot less tidy than a chat. People interrupt, trail off, say "erm", go quiet while they look for a reference number, and call from cars and kitchens. A voice agent that handles all of that badly sounds robotic even when every word it says is correct. These are the problems I hear most often on test calls:

• Talking over the caller. The agent keeps going when the caller tries to speak, or it cuts the caller off. Both feel rude.
• Silent listening. Real people make small sounds while they listen. An agent that stays completely silent while you explain something feels like a dropped line.
• Reacting to noise. Background chatter, a radio or traffic gets heard as speech, so the agent answers things nobody said or loses its place.
• Replies written for a screen. Long sentences, lists and symbols read fine in a chat but sound stiff when spoken out loud.
• Awkward silence handling. The agent either jumps in too fast when the caller pauses or leaves a long gap before it reacts.

The first three are exactly what the three settings in this guide are for. The last two are about your prompt and a couple of other settings, which I cover further down.

Where the settings live

HighLevel's guide to creating, testing and deploying a Voice AI agent explains that advanced configuration is hidden by default. At the bottom of the agent's configuration section you turn on Show advanced settings to see Call Settings, Agent Behavior, Transcription & Speech, Voice Settings and System Prompts. The same guide lists what sits in each group:

HighLevel's official video on noise cancellation and backchanneling shows Backchanneling under Agent Behavior and Noise Cancellation under Voice Settings, both inside advanced settings. Your browser remembers whether advanced settings are shown, so once you turn them on they stay on.

Step by step: make your voice agent sound human

Here is the short version in under a minute, then the full steps below.

More short GoHighLevel fixes like this are on my YouTube channel, Code With Urooj.

Step 1: Open the agent and show advanced settings

Go to AI Agents, Voice AI and open the agent you want to improve. If it is already taking live calls, think about cloning it first and testing your changes on the copy. Scroll to the bottom of the configuration section and turn on Show advanced settings.

Step 2: If backchanneling or noise cancellation is missing, check Labs

HighLevel's noise cancellation and backchanneling help article says to switch on Voice AI Noise Cancellation & Backchanneling in Settings, Labs before you configure them on an agent. If you can already see both options in your agent, skip this step.

Step 3: Set Interruption Sensitivity

Open Agent Behavior and find Interruption Sensitivity. This controls how readily the agent stops talking when the caller starts speaking. HighLevel added it after a long-running Ideas Board request about the agent and callers talking over each other, and a HighLevel team member confirmed on that Ideas Board post that Interruption Sensitivity was added and can be tweaked.

HighLevel's help docs do not publish the range or the default, so check the control in your own account. In my own agent it is a number, and the receptionist in the video above is set to 0.75. In my testing, a higher number makes the agent stop talking sooner when the caller speaks.

• Raise it if callers complain the agent talks over them, or if it carries on reading out a long answer after the caller has said "no, sorry, I meant...".
• Lower it if the agent keeps stopping mid-sentence because of a cough, a "hmm" or noise in the background, and never finishes its answer.
• Test tip: on a test call, interrupt the agent halfway through a long answer, then make a small noise like "mm" without meaning to interrupt. You want it to stop for the first and keep going for the second.

Step 4: Turn on Backchanneling

Still in Agent Behavior, enable Backchanneling. HighLevel describes it as short, natural listener cues inserted during brief pauses to keep the conversation flowing, and says it only adds them in short pauses so it does not talk over the caller. There are two options:

• Backchannel Frequency: how often the cues happen, from rare to frequent. HighLevel's own recommendation is to start at a middle value.
• Backchannel Words: the list of short phrases the agent can use, such as "yeah", "okay" and "uh-huh". You can edit the list to suit the brand, language and culture.

My advice is to keep the word list short and neutral. For a UK business, "okay", "right" and "mm-hmm" usually sound more natural than "awesome". For a law firm or a clinic, leave out anything that sounds too casual. HighLevel's help article also says to make sure the voice you picked supports the language of your words.

Test tip: explain something long on a test call, like a two-sentence description of a problem with your boiler, with small natural pauses. The agent should make the odd listening sound, not one in every gap.

Step 5: Choose a Noise Cancellation mode

Open Voice Settings and find Noise Cancellation. HighLevel's help article lists two modes, and HighLevel's video also shows the option to leave noise as it is:

• Remove Noise: filters common ambient sounds such as air conditioning hum, keyboard taps and traffic.
• Remove Noise + Background Speech: also suppresses overlapping voices, for callers in shared or busy places.

HighLevel's advice is to start with Remove Noise only and step up to background speech suppression if nearby talk is still a problem. For a receptionist that takes calls from the public, I lean towards Remove Noise + Background Speech, because people ring from cars, shops and kitchens with the telly on. Either way, the help article says noise suppression focuses the audio on the main speaker.

Test tip: make one test call with a radio or a video playing near you, and one with someone talking in the room. Then listen to the recording and check nothing the caller actually said got lost.

Step 6: Save and test with a web call and a real phone call

Click Save, then open the Test panel. HighLevel offers a Web Call in your browser and a Phone Call through a real phone connection. Do both. A web call on a good headset is the easiest version of the problem, and real callers will be on mobiles. HighLevel's testing checklist includes a period of silence and an interruption while the agent is speaking, which are exactly the moments these settings change.

Step 7: Review the call and change one thing at a time

Listen back to the recording and read the transcript in the call history or Voice AI logs. HighLevel's guide says to change one setting at a time and retest, so you know which change made the difference. I keep a simple note of each test: what I changed, what I heard, and whether it got better.

Other settings that help

The three settings above do the heavy lifting, but a few others shape how natural a call feels. These all appear in HighLevel's help docs.

• Idle Reminder Timer (Call Settings). HighLevel's idle reminder article says this is how long the agent waits during caller silence before it prompts them. It can be set from 1 to 20 seconds and the default is 8. If the caller stays silent for more than 15 seconds after the reminder, the call ends. Use a longer timer when callers need to look something up, and a shorter one for quick yes or no questions.
• Response speed and response style (Agent Behavior). These sit next to Interruption Sensitivity. If the agent jumps in before callers finish, or leaves too long a gap before it answers, these are worth a test.
• Voice speed and volume (Voice Settings). A slightly slower voice is often easier to follow on a phone line, especially when the agent reads out times, prices or addresses.
• Pronunciation and keywords (Transcription & Speech). Use these for business names, street names and product names the agent keeps mishearing or saying wrong.
• Speech-to-speech models. HighLevel's speech-to-speech guide says these realtime models skip the separate speech-to-text and text-to-speech steps, which reduces delays. For those agents, the Agent Behavior settings HighLevel describes are Thinking Mode and Eagerness, where Eagerness influences how readily the agent decides the caller has finished speaking. If your agent uses one of these models and you cannot find the settings in this guide, that is likely why.

Prompt habits that make replies sound spoken

Settings fix how the agent listens. The prompt decides how it talks. This is my own advice from building voice agents, not a HighLevel feature:

• Write for the ear. Tell the agent to use short sentences and plain words, and to ask one question at a time.
• No lists or symbols. Ask it never to use bullet points, numbered lists, emojis or symbols like "&" or "/". These either get read out or make the reply sound like a document.
• Say numbers the way people do. Ask it to say times as "half past two" and to read phone numbers in small groups.
• Keep the welcome message short. One line that says who it is and asks how it can help. A long greeting is the first thing callers talk over.
• Pick a voice that fits. Try a few voices on real test calls, not just the preview. A voice that suits the business and the caller's accent does a lot of the work.
• Give it permission to pause. Tell it that if the caller is looking for something, it should wait or say "no rush", rather than repeating the question.

Gotchas to watch for

• Too much backchanneling. HighLevel's own advice is that if you use the Idle Reminder Timer, very frequent backchannels can crowd the conversation, so consider a higher idle threshold. Start at a middle frequency.
• Background speech suppression can go too far. HighLevel's FAQ suggests reviewing recordings and logs for any suppression of voices you meant to keep. If a caller hands the phone to someone else, or two people speak on speakerphone, test that case.
• Web calls hide problems. HighLevel notes that Call Transfer does not run during a Web Call test. Use a real phone call to test transfers, noise and interruptions.
• Changing everything at once. If you change five settings and the call gets worse, you will not know which one did it. One change, one test.
• Settings may differ between accounts. Labs features, the model you pick and HighLevel's updates can all change what you see. If a menu does not match this guide, check the help articles linked below for the latest wording.
• Live agents. Test on a cloned agent or outside business hours. A bad setting on a live receptionist means real callers get the bad version.

Once your agent is taking calls, the calls land in Conversations, and my guide to keyword search and the new filters in GoHighLevel Conversations covers the new Last Call Status filter, which helps you find missed calls and voicemails fast. If you run text-based agents too, see my guide to using Claude or Gemini in Conversation AI, and for the bigger picture, the GoHighLevel AI automation guide.