ChatGPT Voice's update surprised me

I tried ChatGPT Voice very early on.

But for over a year, I basically didn’t use it.

Not because the speech recognition was poor, nor because the voice didn’t sound human, but because I always felt that as soon as I switched to Voice, ChatGPT seemed to get a little weaker.

Casual chat was fine, but as soon as things got a bit more complex, I’d think:

never mind, I’ll type.

Recently OpenAI released a new version of Voice. I used it a few more times, and for the first time I was genuinely a bit impressed.

I still type the vast majority of the time. My ratio of Voice to text may not even be 1:100.

But one change is very clear:

there used to be everyday situations where it would never occur to me to open ChatGPT; now I’ve started using Voice in them.

What actually changed in Voice this time?

My own experience is limited, so afterwards I also looked into what OpenAI actually updated this round.

It really isn’t just “the voice sounds more natural.”

1. Conversation finally doesn’t feel so much like trading voice messages

In July OpenAI launched the new GPT-Live.

One very important change is full-duplex: it can keep listening while you talk, while judging when to answer and when to keep listening.

That sounds very technical, but in actual use it’s easy to understand.

People don’t talk like they’re filling in a form:

you finish a sentence → stop → the other side answers in full → your turn again.

We pause, think for a moment, suddenly cut in, change what we’re saying halfway through.

The new Voice has been optimized for pauses, interruptions, background noise and turn-taking.

So this time what I felt very clearly was:

it finally feels more like chatting than taking turns recording.

2. Voice no longer has to shoulder every complex task by itself

I think this matters more than “does the voice sound human.”

Real-time voice has always had a contradiction:

you want it to answer right away, but the more complex the reasoning, search or task, the more time it usually needs.

The new GPT-Live’s approach is to have different models divide the work.

Voice handles the real-time conversation; when something needs search, complex reasoning or heavier work, it can hand it to a stronger model in the background and bring the result back into the current conversation.

In other words:

the one chatting with you and the one doing the serious work don’t have to be the same model.

That also explains why the “capability gap” between voice and text feels much smaller this time.

3. It can finally connect to more of what you were already doing in ChatGPT

Live Voice can now use web search and memory.

In August, Files and Projects were added too.

For example, if you’ve been discussing something in a Project over time, you can now open Voice directly and keep drawing on that Project’s recent chats, sources and instructions, instead of explaining the background from scratch.

That matters a lot to me.

Because I already keep many things in ChatGPT over the long run: writing, product, job hunting, learning, and some long-running questions in life.

One of Voice’s biggest annoyances used to be:

open your mouth, and you have to explain everything again.

Now that sense of disconnect is much smaller.

4. Voice is also starting to show up in places beyond chat

Recently I’ve also noticed Voice entry points appearing in more and more places in ChatGPT, and Work / Codex on desktop have added voice too.

A distinction is needed here:

ordinary chat Voice currently isn’t “say something and it can call all your connected apps and plugins.”

But in agentic environments like Work / Codex, Voice can use the tools and permissions the current environment already has to start tasks, check progress and adjust what’s being worked on.

That’s starting to make me feel that OpenAI’s positioning of Voice may be changing.

The old question was more like:

“can AI talk with me naturally?”

Now it’s starting to become:

“when my hands and eyes are busy with something else, what can I still have AI do for me?”

But I haven’t typed much less because of it

I think this point matters.

I don’t think Voice will replace text, at least not for me.

Writing articles, revising résumés, reading code, comparing information, checking numbers, doing final review — I prefer text for all of these.

Text is easy to skim, revisit and edit, and more precise.

Where Voice has really started entering my life is a different kind of situation.

For example:

thinking things over while walking.

Before, if a question suddenly popped into my head, I wouldn’t stop by the roadside to type for ten minutes.

Now I can just talk it through as I walk.

Mock interviews.

You can do mock interviews in text, of course, but a real interview doesn’t let you sit there slowly composing a perfect answer.

Voice can ask question by question and keep following up, and the problems it exposes are closer to how you actually speak.

Talking a complicated thing through.

Sometimes I don’t even know yet what the question is.

At times like that, rather than first writing a long, well-structured prompt, I’d rather just talk and let it keep asking.

Practicing a language.

This may be one of Voice’s most natural uses.

Not having AI “fix one English sentence” for me, but actually having to open my mouth, put words together and answer on the spot.

So if I had to sum up how I use them now, I’d say:

text suits “output and review”; Voice suits “process and interaction.”

Not absolutely, of course, but the distinction is very useful to me.

If you want to try Voice again, I’d start with these

Voice doesn’t really need complex prompts.

Rather than specifying the final output format, I think it matters more to tell it up front:

how we’ll talk from here.

For sorting out your thoughts on a walk:

I’m going to walk for 20 minutes and want to think X through. Ask me one question at a time; don’t rush to give answers. Keep your replies short, and at the end help me sum up the few things I’m really torn about.

For a mock interview:

You’re now the interviewer for an X role. Ask one question at a time and follow up based on my answers. Don’t hint at how I should answer during the interview; once it’s over, give feedback on content, structure and delivery.

For working through something that’s been bothering you:

I want to talk something through clearly. Don’t rush to give advice; ask me step by step like an interview, help me separate facts, my interpretations, feelings and what I actually need, and sum up at the end.

Or when you’re cooking or walking and can’t keep looking at the screen:

I can’t keep looking at the screen right now. Search when you need to, tell me the most important conclusion first in a few words, and leave the details in the chat for me to read later.

I even feel now that voice prompts differ a little from text prompts:

in text, we often define “what the final output should be”; in Voice, what’s more worth defining is “how we’ll interact from here.”

How many questions at a time, how long the answers are, when to follow up, when to summarize, when to finally give advice.

Once these few rules are clear, the experience gets much better.

Why now, of all times?

I don’t know how high a strategic priority Voice has inside OpenAI, so I don’t want to draw conclusions for them.

But from the way the product has developed, I think now really is a fairly reasonable moment.

Because several conditions have finally matured somewhat at the same time:

voice interaction itself is smooth enough;

ChatGPT already has long-term context like memory, search, Projects and files;

and complex work can be handed to background models and agents.

So Voice no longer needs to be a standalone “voice version of ChatGPT.”

It’s more like a new door into what ChatGPT can already do.

That’s also why I think the recent spread of Voice entry points everywhere is worth noticing.

Not because everyone will be talking to their computers from now on.

But because many uses of AI used to come with a precondition:

you had to stop what you were doing, pick up your phone or sit down at your computer, and start typing.

Voice is starting to remove part of that precondition.

If you tried ChatGPT Voice a year ago and found it pointless, I think it’s worth opening again now.

No need to force yourself to change your typing habits.

Try it in a situation that was never suited to typing:

walking, cooking, before or after driving, practicing interviews, practicing speaking, or just when there’s something on your mind you want to think through out loud.

For me at least, Voice really started becoming useful this time from these small situations.