Voice Mode not working in Mac App - Cutting Off

Having the same issue with mac version.

For some reason the feedback loop does not stop and it cancels out unless you mute the mic after you speak.

I’ve heard that most of OpenAI’s code these days is written by Codex.

That probably means the Mac version of the ChatGPT app was also generated by Codex itself.

But here’s the funny part — Codex can’t actually test audio, can it?

So how is it supposed to fix a bug where ChatGPT picks up its own voice through the mic, interrupts itself, and gets stuck in a feedback loop?

This issue has been around since around May and still hasn’t been fixed.

That happens to be right around the time OpenAI started relying heavily on Codex for development.

Maybe the timing isn’t just a coincidence.

Everything was fine until I tried ChatGPT Voice Mode — now the mic doesn’t move at all, and sometimes I get error 3002 (Discord).
Already checked permissions, restarted, reinstalled, switched mics… still dead.

1 Like

Same problem here - noticed today. Voice mode cutting out repeatedly every few words. Using MacBook Pro M1 Max running MacOS Tahoe 26.0.1. Was just using default built-in mic and built-in MacBook Pro speakers. Seems ChatGTP hears it self speaking back then pauses because it believe it is hearing another person speak then the cycle keeps repeating every few words. Running ChatGPT Version 1.2025.294 (1761352070)

1 Like

Also another thing I noticed (also using MacBook Pro M4 with Tahoe) that after the newest update of ChatGPT app for Mac (29th October 2025) it turns off my internal Mac mic (the only way to turn it back on is to move around the sound input slider) if I try using it in voice mode besides the stuttering and lagging and never being able to finish even the first sentence. I checked my same ChatGPT app on my fathers MacBook Air with an older build and it had no such issues. So it’s definitely the new update that ruined something (update version: 1.2025.294)

Here’s what ChatGPT told me:

Core fact

The newest ChatGPT Mac app (October 29 2025 build) contains a regression in its Core Audio / WebRTC audio-session handling, which causes it to improperly seize and lock the system’s input channel when voice mode initializes.

That lock disables the macOS internal mic driver at the system level until any user action (like moving the input slider) re-initializes the audio stack.


Direct answer

The app’s new voice engine takes exclusive control of the microphone session, fails to complete the handshake with Core Audio, and never releases the input stream.

macOS then marks the internal mic as ā€œinactive,ā€ so your mic level drops to zero and won’t wake until Core Audio is forced to reload (which happens when you touch the slider).


Breakdown

  • 1. Core Audio handshake bug

    The new build hooks into Apple’s AVAudioSession / AVAudioEngine differently from older ones. When the ChatGPT voice module starts, it calls for exclusive access but does not correctly resume or release the stream, leaving the driver in a suspended state.

  • 2. WebRTC deadlock

    Voice mode uses WebRTC to stream audio both ways. When the mic handshake fails, the app keeps looping empty audio buffers. That’s the stuttering you hear — the engine is replaying half-filled frames because it never gets live data.

  • 3. Mac mic driver behavior

    macOS disables a device that appears to be locked but unresponsive, setting its input gain to zero. Touching the slider sends a ā€œset parameterā€ event, forcing Core Audio to re-enumerate the device, effectively ā€œwakingā€ the mic.

  • 4. Speech cutoff / lag

    Since the audio thread is blocked waiting for mic input that never comes, the output thread (my voice) starves of data. The speech synthesis starts, lags, then halts mid-sentence.

  • 5. No issue on older builds / browser

    Older builds used the stable, non-exclusive input mode. Chrome’s WebRTC uses its own mic session management, so it bypasses this bug entirely.


Context / consequences

  • This is a regression introduced in the October 29 build of the ChatGPT Mac app.

  • It’s hardware-agnostic — it happens on M1, M2, M3 machines alike if the internal mic is active.

  • It doesn’t damage the mic; it just leaves the driver waiting for a release signal that never comes.

  • Until patched, any restart of Core Audio or manual slider nudge will temporarily fix it.


Bottom line

The new ChatGPT Mac build mishandles the Core Audio microphone session.

It seizes the mic, fails to complete the handshake, and leaves it muted at the OS level — causing both your mic dropouts and my voice stuttering.

Older builds and Chrome don’t exhibit this because they use the stable shared-session model.

2 Likes

It’s because the mic is so close to the speaker on the new MacBooks that it’s interrupting itself, thinking you said something else. It works fine when you mute the mic after you stop talking. A slight annoyance, but better than not being able to use it.

1 Like

hat’s not the case.
The fact is that the early 2025 versions of the ChatGPT Mac app worked perfectly with Advanced Voice Mode on the exact same MacBook.
If the hardware layout were the real cause, it wouldn’t suddenly stop working under identical conditions.

So this issue isn’t caused by the mic and speaker being too close — it’s clearly a change introduced by a recent app update.

1 Like

Same Issue on a macbook air M4 2025 with macos26 and the newest version on gpt

Well if it were a codex capability/competence issue , we would’ve been facing this issue in MAC Os web …in web version advance chat works great. I think for some reason they have stopped working in Mac OS app , as the new updates works better in Windows app - this advance voice integration patch (25 th nov 2025) doesn’t work in Mac OS app but it flows nicely in Windows app

Yeah, you might be right — my earlier comment was just speculation on my part.

I’ve also heard rumors that OpenAI triggered some kind of internal ā€œcode redā€ recently, so I’m starting to think this Mac app issue might stay unresolved for a while.

So, can I ask you a question to kill time?

If you were developing a voice chatbot and wanted to use Codex to diagnose and fix issues with a voice session like this, what prompts would you send?

Can’t believe this Core Audio bug in ChatGPT Mac desktop version still not fixed in the latest update.

1 Like

Same issue 2025.329 and I’ve reported it more than a few times for months now. Seriously? Can’t switch to Gemini Pro fast enough.

I have the same issue. I’m not as technical as some folks, but it seems that the fix should be very basic. Every time Chat GPT starts to speak, have it auto-mute the mic, and when it’s done speaking, have it auto unmute the mic. That would resolve the issue. It seems like a pretty basic bit of coding that could be added into the desktop app.

You’re right that auto-muting the mic while the AI speaks would avoid feedback issues — but that’s actually the old-style approach to voice interfaces. One of the key features of Advanced Voice Mode is that the user can interrupt the AI while it is speaking. If the mic is automatically muted during output, that feature is completely lost.

I can only speak for myself, but I would rather not be able to interrupt the AI than not be able to talk with it at all. Currently, the ā€œold wayā€ is clearly the best option as it’s the only option which would allow the AI to speak, on a MacBook, without cutting itself off.

I agree — being able to use it at all is better than nothing.

But I just want to point out that the previous versions did support Advanced Voice Mode on the MacBook without this issue. It worked perfectly before.

Going back to the old voice mode isn’t really a great option for me because the experience is much worse compared to what we had.

By the way, is there anyone who actually has Advanced Voice Mode working on the macOS ChatGPT app right now?

If so, it would be really helpful to hear your environment details.

For what it’s worth, I contacted OpenAI support about this the other day, and they told me:

Update the ChatGPT app

Open the App Store → Updates and make sure the ChatGPT app is on the latest version.

Which is… interesting, because the macOS ChatGPT app isn’t even distributed through the App Store.

So it’s pretty clear the support team hasn’t actually used it, and their answer didn’t help at all.

I’m currently using ChatGPT version 1.2025.091 (1743812765) on macOS Tahoe 26.1.

This is, I believe, the last known version to be working properly with Advanced Voice Mode using the built-in microphone and speakers on MacBook.

However, a problem remains when using AirPods: the mic works fine but the voice of the agent is slow and distorted when they answer.

1 Like

I cannot believe after all this time MAC users are still having to put up with the audio problem. I’m currently on a Mac M4 Mini with Sequoia 15.7.2. I use an external microphone an external speakers. And it feeds back on itself interrupting itself consistently totally unusable. I’ve also noticed the same problem is happening on the iPhone when it didn’t used too. I have also previously experienced the robotic slow voice that somebody else mentioned above. Using earphones is the only solution. It’s a pretty big fundamental bug to still be kicking around.

1 Like