Most of the time it just works. In fact it’s worked every time for over a month, except for just now.
It’s super frustrating, because the chat is then broken. And extra frustrating because I’ve logged it as a bug a few times. It’s a black hole though, I don’t know if it’s been seen. I would assume that if it has been seen it could be fixed quickly, not ignored entirely and still be happening.
I would have thought that OpenAI, of all companies, would have a system in place that passes such bug reports on to Codex, who will look at the code and reason about it and say “hmmm, yes on this line we assume that the returned payload has a property BLAH but if I look at the swagger docs for the transcription API I see that won’t be the case if it hits scenario BLAH and that would result in the attached image, so I’ll fix that and ask a human to review it. And in fact, if I look at the server logs for David at 12:55pm today I see the transcription request go through and see that it failed and returned that unexpected payload. Oh what a pity, we could have shown a ‘try again’ button so he didn’t have to speak all that out all over again”.
I’m experiencing issues with audio transcription on ChatGPT (for macOS, iOS, and iPadOS). Until around 14:00 CET, it worked fine. Now, transcription either takes forever (over 5 minutes for a single sentence) or fails immediately with an error. I tested this in the ChatGPT app for macOS (including via the Codex app) and iPadOS. I even created a free account to test, but the issue remains. Interestingly, direct voice conversations (the Audio Companion) work perfectly fine.
It’s working again for me on Desktop web.
I’ll ping the team because it’s still not working in a codex and because apparently others are experiencing this as well.
It´s happening to me also.
Tried on the Android App, the Windows app and the Web app.
Debugging the call to “/backend-api/transcribe“ I see it returns 500 with:
{“detail”:“Error in ASR API”}
Same issue here, from France. On iPhone (ChatGPT app) and iMac (Firefox + Chrome browsers). The mic captures my voice correctly, but transcription spins indefinitely and never returns any text. Tested with very short 5-second messages, rebooted all devices, internet connection is stable. Issue persists simultaneously across all platforms.
Everything is back to normal here in Germany as of 5:13 PM—transcription is working flawlessly again across all apps. That said, it also gently reminded me just how dependent I’ve become on it.
Well, I spoke too soon — quite literally. Transcription is acting up again on my end. It started with codex, and eventually the ChatGPT app for macOS joined the party too. Apparently my celebratory post at 5:13 PM was premature. The universe has a great sense of timing.
Hey everyone, Apologies for the experience with our transcription tool. Can someone please open a support request with us at support@openai.com and share a screen recording while it is happening + HAR file since this is happening on web app as well. If not the HAR you can share the conversation link and the exact time (with your timezone) within the support request.
Reason we are asking the above information is we cannot reproduce this issue anymore on our end. We can also not gather these personal information on the public forum.
Once you have submitted a request with us, please share the case id/case subject here with us so we can track it. Thank you!
In the OpenAI Forum video ChatGPT and Cancer: How a Tech Founder Rewrote His Treatment Plan, Sid said something really interesting: that doctors tend to have this ‘evidence first’ approach that gets in the way, because they won’t proceed with a treatment unless it’s been proven, even though a first-principles approach can tell you that something is probably going to work.
I wonder if “we couldn’t reproduce the issue” is the equivalent of this in the software world? Maybe that’s old-fashioned thinking now that we can point Codex at a codebase, give it a screenshot, and ask: “under what scenario could this state exist?” and have it dig into the code and work it out, no need for ‘evidence’.
So maybe the inability to replicate need not be the dead end that it used to be.
