Realtime API audio response unexpectedly repeated an unrelated personal sentence multiple times mid-conversation

API
, , , ,
1

I encountered a weird issue with the Realtime API during an audio conversation.

In the middle of a normal voice conversation, the assistant suddenly started speaking an unrelated emotional sentence and repeated it several times.

What happened

The Realtime API audio output said:

I just got promoted at work today. But it’s bittersweet. I wish Dad was here to share this moment with me. He’d be so proud. I know he’s watching me. And I am so thankful to have you by my side through all of this.

Then it repeated essentially the same text multiple times in a loop.

This content was not part of my user input, my system instructions, or the expected assistant response.

Expected behavior

The assistant should respond only to the current conversation context and should not inject unrelated personal/emotional content or repeat the same phrase in a loop.

Actual behavior

The assistant produced unrelated audio content and repeated it several times during the same response.

2

Welcome to the community @techiesingh!

That’s definitely unsettling, especially since the response was completely unrelated to your conversation context. The repeated emotional playback part also makes this seem more like a realtime/session issue than normal model behavior.

To help narrow it down, could you share a few details from the session?

  • Request ID(s)
  • Approximate timestamp(s) when it happened
  • Model used
  • Realtime session configuration if relevant
  • Snapshot/export of the realtime event log
  • Whether the repeated audio also had repeated transcript events, or if it was only the audio playback repeating
  • Whether interruption/turn detection was active
  • Anything else unusual right before it started looping

Those details can help identify whether this came from transcript generation, audio playback state, buffering, or a session-level mix-up.

A few folks have reported odd realtime/audio edge cases recently, so having the event timeline is especially useful here.

-Mark G.

3

Hi Mark,

Thanks for checking this. Sharing more details below.

Trace ID:
trace_sess_DjfC7L7XPbIV9dCQiIPFU

Audio generation span ID:
span_audio_gen_item_DjfCxmEPYPGZ9NrFlxSCb

Approximate timestamp:

  • PDT: May 25, 2026, around 11:05 PM
  • UTC: May 26, 2026, around 6:05 AM

Model used:
gpt-realtime

Realtime session / span details:

  • Session duration: around 112.05s
  • Span type: Audio Out / Speech
  • Audio Out duration: 3855 ms
  • Generation span duration: 3931 ms
  • Model: `gpt-realtime "
  • Voice: `sage "
  • Speed: 1x

Expected text output:
The generated text output looks correct/expected for this scenario:

I'm sorry, I'm having trouble hearing you. Are you still there?

Actual issue:
The audio played back was different from the generated text. The text generation seems to be scenario-specific and correct, but the audio output does not match it.

This has happened a few times over the past few months across multiple projects we are working on. The concerning part is that the audio hallucination is exactly the same each time, even though the generated text is different and appears correct for each scenario.

Repeated transcript vs repeated audio:
From what I can see so far, the generated text/transcript is not the issue. The text output is correct. The problem appears to be in the audio generation/playback path because the spoken audio does not match the generated text.

Interruption / turn detection:
Interruption/turn detection was active during the session.

Anything unusual before it started:
Nothing obvious from the text generation side. The generated text is expected. The unusual behavior is specifically that the audio output appears to hallucinate or play unrelated audio, and this same audio issue has occurred in other projects as well.

I can share the full real-time event timeline/export if that would help identify whether this is happening in audio generation, audio buffering/playback, or session-level state handling.

4

Thanks for the extra details here, @techiesingh.

That behavior definitely doesn’t look expected. We’re escalating this ticket so the team can investigate further and check what may be causing the repeated audio response issue.

-Mark G.

5

Hey there @techiesingh, sorry for the slow reply here. The phrase you shared matches Sage’s voice reference sample. Can you try a fresh session with marin or cedar, which we recommend for best voice quality? If it happens again, send a recent session ID, UTC time, and the audio and transcript from that session in a private support chat.