Realtime voice (GPT-Live) transcription becomes extremely slow after ~3 minutes of conversation

Open 💬 0 comments Opened Jul 26, 2026 by thewh1teagle

Description

The realtime voice conversation (GPT-Live) in Codex is super fun to talk with, but there's an issue: after roughly ~3 minutes of talking, transcription becomes very slow and barely transcribes what I say, and I have to open a new session to fix it.

I can literally see the transcription very slowly adding more words — even 5 seconds after I speak, the words still arrive slowly, one by one.

Notes

  • This happens only in Codex — the live conversation in the ChatGPT app does not have this issue.
  • Workaround: opening a new session restores normal transcription speed, but only for another ~3 minutes.

Steps to reproduce

  1. Start a realtime voice conversation in Codex
  2. Keep talking for ~3 minutes
  3. Transcription slows down dramatically and barely keeps up with speech

Version

Codex Version 26.721.41059 (5848)

System info

  • macOS 26.5.2 (build 25F84)
  • MacBook Pro, Apple M4 Pro (arm64)
  • 24 GB RAM

View original on GitHub ↗