[Production Incident] Severe Static / Crackling Noise in GPT-Realtime Assistant Audio (WebRTC)

API Bugs
1

We are experiencing a production-impacting issue with GPT-Realtime over WebRTC.

Starting today (June 12), we began receiving reports from users that assistant-generated audio intermittently contains extremely loud static/crackling noise.

Environment:

  • GPT-Realtime 1.0

  • WebRTC

  • Production educational voice application

Symptoms:

  • The issue occurs in assistant-generated audio output.

  • It sounds similar to severe microphone clipping, electrical interference, or corrupted audio frames.

  • The noise is loud enough to disrupt lessons and cause significant user discomfort.

  • Once it starts, it continues until the current assistant utterance is interrupted.

Impact:

  • Approximately 50% of live lessons are affected.

  • Affected lessons experience the issue roughly 2 times during a 20-minute session.

  • Prior to today, we had not observed this behavior.

  • No deployment or audio pipeline changes were made on our side before the issue began.

Additional Notes:

  • The problem appears in model-generated speech, not user microphone input.

  • We are currently evaluating GPT-Realtime 1.5 as a temporary mitigation, but GPT-Realtime 1.0 is our production configuration.

Questions:

  1. Are there any known issues affecting GPT-Realtime audio generation today?

  2. Were there any recent changes to the Realtime audio stack or voice generation pipeline?

  3. Has anyone else observed similar corruption in assistant audio output over WebRTC?

  4. Is there any diagnostic information we should collect to help the Realtime team investigate?

**Prior to today, this issue was effectively 0%. It appeared suddenly and is now affecting roughly half of active lessons.

This issue is actively affecting production users and we would appreciate any guidance from the Realtime engineering team.

Thank you.

2 Likes