error: cannot update a conversation's voice if assistant audio is present bug

[shehan360]

Describe the bug

Started receiving the following error message through the realtime session recently,

Error(message="Cannot update a conversation's voice if assistant audio is present.", type='invalid_request_error', code='cannot_update_voice', event_id=None, param=None)

This only seems to happens if you have a model_config passed to runner.run() method.

Debug information

• Agents SDK version: 0.2.10
• Python version: Python 3.10

Repro steps

1. Setup the examples/realtime/app/server.py
2. Change the runner.run() line (

   openai-agents-python/examples/realtime/app/server.py

   Line 46 in 6df6fe1

   session_context = await runner.run()

   ) to the following

        session_context = await runner.run(
            model_config=RealtimeModelConfig(
                initial_model_settings=RealtimeSessionModelSettings(
                    model_name="gpt-4o-realtime-preview-2025-06-03",
                    input_audio_format="pcm16",
                    output_audio_format="pcm16",
                    input_audio_transcription=RealtimeInputAudioTranscriptionConfig(
                        model="gpt-4o-transcribe",
                        language="en",
                    ),
                    voice="alloy",
                    speed=1.2,
                ),
            )
        )

3. Talk to the agent and get past the handover.
4. Right after the handover you should see

Error(message="Cannot update a conversation's voice if assistant audio is present.", type='invalid_request_error', code='cannot_update_voice', event_id=None, param=None)

Expected behavior

A clear and concise description of what you expected to happen.

[shehan360]

This issue doesnt occur if you provide the model settings when creating the runner!

runner = RealtimeRunner(
            starting_agent=self.realtime_agent,
            config=RealtimeRunConfig(
                model_settings=RealtimeSessionModelSettings(
                    model_name="gpt-4o-realtime-preview-2025-06-03",
                    input_audio_format="pcm16",
                    output_audio_format="pcm16",
                    input_audio_transcription=RealtimeInputAudioTranscriptionConfig(
                        model="gpt-4o-transcribe",
                        language="en",
                    ),
                    voice="alloy",
                    speed=1.2,
                ),
            ),
        )
self.session_context = await runner.run()

[seratch]

Thanks for reporting this issue. We're working on gpt-realtime (GA) migration. I will check this issue as well.