RealTime GA text transcription

API
1

Hi OpenAI Team,

I’m writing to request the restoration of the input_audio_transcription feature that was available in the Realtime API beta but removed in the GA release.

What We Had in Beta:

// Session configuration (Beta)
{
  type: "session.update",
  session: {
    modalities: ["text", "audio"],
    turn_detection: { type: "server_vad" },
    input_audio_transcription: { model: "whisper-1" },  // ✅ This existed
    input_audio_format: "g711_ulaw",
    output_audio_format: "g711_ulaw"
  }
}

This would generate conversation.item.input_audio_transcription.completed events containing real-time user speech transcripts within the same session.

What Changed in GA:

The GA release removed input_audio_transcription entirely. Multiple developers in the community have reported this issue:

  • “Input_audio_transcription in realtime-api”
  • “Unable to Access User Audio Transcript in Realtime API”
  • “Problems using session.update with the realtime-api”

Questions:

  • Is input_audio_transcription removal permanent or temporary?
  • What is the recommended approach for real-time user transcription in GA?
  • Are there alternative methods we should be using?

The community would greatly appreciate clarity on this feature’s status.

Thank you for considering this request.

Best regards,
Rohan

2

Rohan, the messages are not gone, you have to request them by providing a transcription model. I get

        session_update_message = {
            "type": "session.update",
            "session": {
                "type": "realtime",
                "model": "gpt-realtime",
                "audio": {
                    "input": {
                        "format": {          
                            "type": "audio/pcm",
                            "rate": NATIVE_OAI_SAMPLE_RATE_HZ
                        },
                        "noise_reduction": {"type":"far_field"},
                        "transcription": {
                            "model": "gpt-4o-mini-transcribe"
                            },
                        "turn_detection": {
                            "create_response": True,
                            "interrupt_response": False,
                            "prefix_padding_ms": 300,
                            "silence_duration_ms": 700,
                            "threshold": 0.5,
                            "type": "server_vad"
                            }
                        },
                    "output": {
                        "format": {
                            "type": "audio/pcm",
                            "rate": 24000
                        },
                        "speed":1,
                        "voice": "marin",
                    }
                },
                "instructions": sp,
                "max_output_tokens": 4096,
                "output_modalities": ["audio"],
                "tool_choice": "auto",
                "tools":[],
                "tracing": None,
                "truncation":"auto"
            }
        }

then I catch these events:

    if event_type == 'response.output_audio_transcript.done':
        # ai transcript

    elif event_type == 'conversation.item.input_audio_transcription.completed':
        # human transcript

hope this helps!!

1 Like
3

Yes! That worked! Thank you so much.