Use realtime transcription when your application needs live speech-to-text without a spoken assistant response. Realtime transcription sessions stream transcript deltas as audio arrives, so users can see text before the full utterance is complete.

For the lowest-latency streaming transcription path, use gpt-realtime-whisper. For offline files or workflows that don’t need streaming deltas, use the standard speech-to-text models in the Audio API.

gpt-realtime-whisper is an alternative for live transcription, not a blanket replacement for every transcription model. Test it against your audio, languages, vocabulary, and latency requirements before switching production traffic.

Realtime transcription uses a session with type: "transcription". You can connect with WebSocket for server-side audio pipelines or WebRTC for browser audio.

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
{
  "type": "session.update",
  "session": {
    "type": "transcription",
    "audio": {
      "input": {
        "format": {
          "type": "audio/pcm",
          "rate": 24000
        },
        "transcription": {
          "model": "gpt-realtime-whisper",
          "language": "en"
        }
      }
    }
  }
}

Session fields

• type: Set to transcription for transcription-only sessions.
• audio.input.format: Input encoding for audio appended to the buffer. Use 24 kHz mono PCM when sending audio/pcm.
• audio.input.transcription.model: Use gpt-realtime-whisper for streaming transcription.
• audio.input.transcription.language: Optional language hint such as en.
• audio.input.transcription.delay: Optional latency/accuracy tradeoff for gpt-realtime-whisper. Supported values are minimal, low, medium, high, and xhigh.
• audio.input.turn_detection: Optional voice activity detection for models that support it. For gpt-realtime-whisper, omit this field or set it to null, then commit audio manually.

Send audio chunks with input_audio_buffer.append:

1
2
3
4
5
6
ws.send(
  JSON.stringify({
    type: "input_audio_buffer.append",
    audio: base64Pcm16,
  })
);

If you disable turn detection, commit the buffer when you want transcription to begin:

1
2
3
4
5
ws.send(
  JSON.stringify({
    type: "input_audio_buffer.commit",
  })
);

For models that support server VAD, the session commits audio automatically when it detects a turn boundary.

Listen for incremental transcript deltas and completion events:

1
2
3
4
5
6
7
8
9
10
11
ws.on("message", (data) => {
  const event = JSON.parse(data);

  if (event.type === "conversation.item.input_audio_transcription.delta") {
    process.stdout.write(event.delta);
  }

  if (event.type === "conversation.item.input_audio_transcription.completed") {
    console.log("\nFinal transcript:", event.transcript);
  }
});

A delta event contains newly available transcript text:

1
2
3
4
5
6
{
  "type": "conversation.item.input_audio_transcription.delta",
  "item_id": "item_003",
  "content_index": 0,
  "delta": "Hello,"
}

A completion event contains the final transcript for the committed item:

1
2
3
4
5
6
{
  "type": "conversation.item.input_audio_transcription.completed",
  "item_id": "item_003",
  "content_index": 0,
  "transcript": "Hello, how are you?"
}

Ordering between completion events from different speech turns isn’t guaranteed. Use item_id to match transcription events to committed input items.

Streaming transcription trades latency for transcript quality. Lower delay settings can produce earlier partial text. Higher delay settings give the model more audio context before emitting text and can improve word error rate.

Start by setting audio.input.transcription.delay and testing against your real audio. Useful starting points are:

• minimal for the most latency-sensitive interactions;
• low for low-latency live captions;
• medium for a balanced latency/accuracy tradeoff;
• high when accuracy matters more than immediate display;
• xhigh when your workflow can tolerate the most delay for additional context.

The exact delay in milliseconds can vary by model configuration, so benchmark with representative audio instead of assuming a fixed timing per level.

Don’t choose a setting from synthetic audio alone. Test with representative microphones, telephony audio, accents, background noise, code-switching, domain vocabulary, and long sessions.

Guide vocabulary and domain terms

If your application depends on exact domain vocabulary, include a language hint and use prompt or keyword steering only when your selected model supports it. For gpt-realtime-whisper in GA Realtime sessions, prompt is not supported.

Where prompt steering is available, use short keyword lists rather than long instructions. The model is already instructed to transcribe, so focus prompts on domain vocabulary, spelling, or style rather than re-stating the transcription task.

Example keyword style:

Keywords: metoprolol, atorvastatin, A1C, systolic, diastolic

For production, treat keyword steering as an aid rather than a guarantee. Continue to evaluate names, numbers, dates, medication names, product names, artist names, and other high-value entities manually.

Only request optional fields that your selected model and endpoint support. If your application needs confidence scoring, timestamps, or diarization, verify support before launch and add fallbacks for fields that aren’t available.

When log probabilities are available, request them with include:

1
2
3
4
5
6
7
8
9
10
11
12
13
14
{
  "type": "session.update",
  "session": {
    "type": "transcription",
    "audio": {
      "input": {
        "transcription": {
          "model": "gpt-realtime-whisper"
        }
      }
    },
    "include": ["item.input_audio_transcription.logprobs"]
  }
}

• Pick a target latency and accuracy threshold before tuning.
• Test against real production audio, not only clean samples.
• Test each target language.
• Include numbers, dates, currency, email addresses, product names, and domain terms in your eval set.
• Track empty, truncated, and delayed transcripts apart from word error rate.
• Decide how your UI should revise partial text when later deltas correct earlier text.
• Use item_id to order and reconcile final transcripts.
• Keep a fallback path for unsupported timestamps, diarization, or confidence fields.