What you’ll be able to do

You can talk to an AI the way you’d talk to a person on the phone. No typing, no waiting to read a reply, just you speaking and it speaking back, more or less in real time. ChatGPT, Gemini, Claude and Copilot all have some version of this now. You’ll be able to practise a conversation, think out loud while it responds, get hands-free help while you’re cooking or driving (not while actually driving, obviously, but you know what I mean), or just have something to bounce ideas off without touching your keyboard.

That’s genuinely useful. It’s also worth knowing exactly what it is before you get too comfortable with it, which is what the next bit’s about.

Before you start

Right, the honest bit.

These voice modes are good. Genuinely good. Gemini Live can even use your camera and talk about what it sees. ChatGPT’s voice mode holds a proper back-and-forth, pauses when you pause, picks up again when you start talking. It’s a big jump from typing messages into a box and it does feel more natural, more like a chat with an actual person.

But it isn’t a person. It just isn’t.

These assistants are built to be agreeable. That’s not a flaw exactly, it’s just how they’re designed, to keep the conversation smooth and pleasant and moving forward. A real friend disagrees with you sometimes, tells you you’re wrong, goes quiet because they’re having a bad day. An AI doesn’t do any of that unless you specifically ask it to push back, and even then it’s doing what you asked, not what it actually thinks (because it doesn’t think, not in that sense).

So use it for practising a job interview. Use it for talking through a problem when there’s no one around at 11pm and you just need to hear your own thoughts out loud. Use it to practise a language, or rehearse a difficult conversation before you have the real one. All of that is a sound use of a voice AI.

Don’t use it instead of a person when you’re lonely or in crisis. It can feel like company, and feeling like company isn’t the same as having it. If you’re struggling, please talk to someone, actually talk to someone: Samaritans on 116 123, or NHS 111 if you need help and don’t know where else to turn. An AI will always sound like it’s listening. It won’t ever actually be there.

Right. With that said, let’s get you set up.

Get set up

You’ll need the app, not the browser version, for proper voice conversation. Here’s what’s around at the moment.

ChatGPT: download the ChatGPT app (iOS or Android). Free accounts get some voice access; a paid plan gets you more of it. Open a chat and look for the headphone or waveform icon, usually bottom right of the message box. Tap it and start talking.

Gemini: Gemini Live is built into the Gemini app. Open the app, there’s a Live option you tap into, and it can also switch on your camera so it can talk about what it’s looking at, not just what you’re saying. Handy if you want it to describe something or help you with a task you’re doing with your hands.

Claude: the Claude mobile app has a voice mode too. Open the app, start a conversation, and there’s a microphone or voice icon to switch into spoken mode.

Copilot: Microsoft’s Copilot also has a voice option in its app, works similarly, tap the mic icon and speak.

Give it microphone permission when it asks, obviously, or none of this works. Find somewhere reasonably quiet the first time, not because it can’t cope with noise but because you’ll get a much better feel for how it works if it can hear you clearly. And put your headphones in if you’ve got them. Feedback from your phone’s speaker into its own microphone makes for a very confusing first go.

Try it yourself

Here are three prompts to say out loud once you’re in voice mode. Each does something different.

For practising a real-life conversation:

I've got a difficult conversation coming up with my landlord about a repair that hasn't been done. Can you play the landlord and let me practise how I'd raise it? Be a bit reluctant and give me some pushback so I can practise staying calm.

For thinking out loud when you’re stuck on something:

I'm trying to decide between two job offers and I keep going round in circles. Can you ask me questions one at a time to help me work out what actually matters to me here, rather than just telling me what to do?

For hands-free help while you’re doing something else:

I'm cooking and my hands are covered in flour. Can you talk me through the next steps of a basic bread recipe, one step at a time, and wait for me to tell you when I'm ready for the next one?

What you should get back: a spoken reply, reasonably quick, in a fairly natural tone. It’ll pause for you. It won’t talk over you (usually). With the landlord one, it should actually play a bit of a character rather than just being helpful and nice about it, if you’ve asked it to push back a bit. With the job offers one, it should genuinely ask you questions instead of just handing you a list of pros and cons. If it does the opposite of what you asked, that’s worth noting for the next section.

Check the result

A few things worth actually testing rather than just assuming they work.

Say something and then stop talking mid-sentence, deliberately, and see whether it waits for you or jumps in too early. Good voice mode should give you a beat of silence before it responds.

Ask it to repeat something back to you in different words, to check it actually understood you rather than just responded to the general shape of what you said.

Try interrupting it while it’s talking, mid-sentence, and see if it stops and listens or just keeps going. The better implementations (ChatGPT’s advanced voice, Gemini Live) are meant to handle this reasonably gracefully.

Check the pacing. Does it sound like it’s reading a script at you, or does it sound like it’s actually responding to what you said? A good exchange has a bit of rhythm to it: you talk, small pause, it replies, and the reply actually connects to your specific words rather than being a generic response that could’ve followed anything.

If all of that feels roughly natural, you’re set up properly and the feature’s working as it should on your device.

If it doesn’t work

If voice mode won’t switch on at all, check your app is updated first, that’s the most common culprit. Then check microphone permissions in your phone’s settings, not just the app’s own settings, both need to be right. If you’re on a free tier, check whether voice access is limited or needs a subscription, this varies by app and does change over time, so worth checking the app itself rather than assuming.

If it keeps cutting you off or mishearing you, try a quieter room, or use headphones with a built-in mic rather than relying on your phone’s.

Now the bit that matters more than any of that.

Don’t tell an AI voice assistant anything you wouldn’t want stored somewhere. Voice conversations are recorded and kept in the same way text chats are, they go into your account history, they can be reviewed as part of how these companies improve their models unless you’ve turned that off in your settings. So don’t share passwords, banking details, anyone else’s private information, or anything you’d genuinely mind resurfacing later. Go into the privacy or data settings of whichever app you’re using and see what’s actually being stored and for how long; each one lets you manage this slightly differently.

And don’t use voice mode as a stand-in for a person when things are genuinely hard. If you’re in crisis, an AI is not the right thing to be talking to. Samaritans: 116 123. NHS: 111. Real people, properly there.

Keep exploring

If you liked how the interview-style prompt went, there’s a full guide on how to get AI to help you practise for an interview, worth a look if you’ve actually got one coming up. If you travel or talk to people who don’t speak your language, there’s a guide on how to get AI to translate a live conversation, which uses a lot of the same voice features covered here. And if this whole “AI is agreeable” thing bothers you as much as it should, there’s a guide on how to get AI to be honest with you, which is worth reading before you trust it with anything important.

Sources and review notes

Review date: 15 September 2026 — every source above was opened and checked against the text on that date.