OpenAI’s GPT-4o model emulates the user’s voice in a noisy background because it gets confused, but the issue has been mitigated at a “system level”