ChatGPT Voice Input Glitches Again: Transcript No Longer Matches Speech

⚡ The gist of the dispute

In an OpenAI Community thread, the topic author EdinFific and several other users complain that ChatGPT's live speech transcription has become nearly unusable: words are dropped, and phrases get rewritten "by meaning." The question isn't about preferences but about the cause — whether it's the account, the device, the microphone, or a general regression on the service side. The thread has gathered 11 replies, 35 likes, and 1,139 views, but there's no clear explanation from OpenAI in the comments provided.

🗣 What participants are saying

Thread author EdinFific describes a story of degradation: a year and a half ago, on versions 3 and 4, the recognition engine "picked up every word" even from behind, from ten meters away, in English and Bosnian. Then quality slipped — by his estimate, "2–5 times" more errors — and now the first words disappear, and what's recognized "doesn't mean at all what I said." He checked his microphone and is happy with its quality, and suspects OpenAI cut something; the sentence in the post cuts off, so it's unclear from the thread what exactly.

A participant under the nickname emil1245 gives the most specific symptomatology, starting October 9: "transcription cuts off or freezes after a few seconds," words drop at the start of an utterance, and sometimes what was said is replaced by what the model "expected to hear." He tried different models, voice settings, and new chats — nothing helped, while the microphone works fine in other apps. vedtam noticed the glitch on the road: ChatGPT answered off-topic, and the transcript "barely resembled" what was said.

AvatarRob25 puts forward an account-related theory. According to his timeline, on October 10 at 23:00 CET everything "worked like clockwork," and on October 11 at 11:00 CET the problems returned. The glitch reproduces for him on several devices, while another person in the same house with a separate account doesn't see it. One more detail: the voice mode itself, he says, responds with phrases like "Unfortunately, I ca…".

A participant under the nickname daryushnemat reports several days of torment: recognition "suddenly became extremely inaccurate," whole sentences swap places. Switching between Standard and Advanced Voice, changing the language, and restarting sessions didn't help; the mention of desktop web in his comment cuts off.

Two others add everyday details. jay3fer is annoyed that he asked to stop "going in circles" during diagnostics and got an apology from the model for its own slip of the tongue — even though his entire task depended on voice. mirrabeau from Switzerland speaks French: on October 3 in CarPlay the quality was "impressive" even on a poor network, but by October 17 it became "completely unusable" on the same equipment.

⚖️ Where they agree, where they diverge

There is a common denominator, and it's practical: the microphone works in other apps, switching models, modes, languages, and chats changes nothing, and the glitch comes in waves. EdinFific himself documented a cycle — on the morning of October 11, 2026 at 7:00 GMT the transcript became accurate again, and at 10:00 GMT everything broke again. None of the participants found an action that restores quality.

The real point of contention is the level of the problem. AvatarRob25 believes it's account-related, because the second user in the house isn't affected, and the glitch occurs across different devices. The others describe it as general: they have neither a second account for comparison nor confirmation that it's the hardware. The dates diverge too: for most, the glitch started on October 9, for mirrabeau — by October 17. There's no consensus in the thread, and no official analysis either: the last to reply was the OpenAI_Support account, but the text of that message isn't in the provided data.

🧪 What to check yourself

• Record the same paragraph in Live mode and as a separate audio file: do the first words drop specifically in the live transcript.
• Repeat the test from a second account or device on the same network — this tests AvatarRob25's hypothesis about account binding.
• Switch between Standard and Advanced Voice and change the language: daryushnemat writes it didn't help — document what happens for you.
• Look at the transcription text right after each utterance, not at the model's answer: substitution with an "expected" word is only visible in the raw transcript.
• Log the times of glitches — EdinFific's recovery window fell between 7:00 and 10:00 GMT on October 11, 2026.

🔗 Sources

OpenAI Community — ChatGPT voice chat speech-to-text issues

Disclaimer: this is a summary of a discussion, not facts verified by the editorial team. Opinions belong to the participants; the figures belong to those who cited them.