Voice input transcription failing in realtime-demo

[vagoel]

Describe the bug

After running the realtime-demo, I am logging all events in the console and noticed an issue with voice input transcription.

https://github.com/openai/openai-agents-js/tree/main/examples/realtime-demo

Expected Behavior

When providing voice input, I should receive:

• Voice output (working ✅)
• Transcribed text for voice output (working ✅)
• Transcribed text for my voice input (failing ❌)

Current Behavior

• Voice output is working correctly
• Transcribed text for voice output is working correctly
• Transcribed text for my voice input is failing

Steps to Reproduce

1. Run the realtime-demo from /examples/realtime-demo
2. Enable console logging for all events
   Add this line to log all events -

session.on("transport_event", (e) => { console.log("Transport event:", e); });

3. Provide voice input
4. Observe that while voice output and its transcription work, the transcription of the user's voice input fails

Environment

• Repository: openai/openai-agents-js
• Demo: examples/realtime-demo
• Browser console logging enabled

Additional Context

This appears to be specifically related to the transcription functionality for user voice input, while other voice-related features are working as expected.

[seratch]

Thanks for flagging this. We'll look into this. If anyone finds a solution, sending a pull request for fixing this would be appreciated.

[agata]

I’ve observed a similar issue in the application I’m developing. In my case, audio input transcription events did not appear unless I explicitly added a transcription property like this:

const session = new RealtimeSession(agent, {
  transport,
  model: "gpt-realtime",
  config: {
    audio: {
      input: {
        // This transcription property seems required to trigger audio input transcripts
        transcription: { model: "gpt-4o-transcribe" },
      },
    },
  },
});

It may be that the missing transcription property is the cause.
(To note: I haven’t tested this with the provided demo.)

[Sanel0101]