openai / gpt-realtime-whisper

Overview

GPT Realtime Whisper by OpenAI is a streaming speech-to-text model built for live transcription. It emits transcript deltas as speech arrives, then produces a final transcript for each committed chunk. You can tune a delay setting to trade faster text updates for stronger accuracy.

This model fits apps that must show words quickly, like captions and call monitoring. It reduces the need to wait for full-file transcription before users see useful text.

What you can build

• Live captions for meetings, webinars, and events
• Real-time call transcripts for support and sales QA
• Voice note transcription for field workers and clinicians
• Accessibility overlays for live audio streams
• Moderation and escalation triggers from spoken conversations
• Dual-track workflows that transcribe now and clean up later with a text model

Inputs

• An audio recording to transcribe, uploaded as a single file (maximum 60 minutes and 500 MB).
• Optional context about the recording, like the scenario or domain terms. Some Realtime GA sessions don’t support a transcription prompt for this model.
• Optional expected spoken language as a short code (example: en). Leave it blank to let the system auto-detect.
• Optional latency versus accuracy preference using a delay level. Higher delay waits longer before emitting text.

Outputs

• A speech-to-text transcript returned as text.
• For streaming workflows, the transcript can arrive as incremental text deltas that may revise earlier words.
• A final transcript is produced for each committed portion of audio.

Recommended settings

• Use the lowest delay level when your UI must react instantly to speech.
• Use a low delay level for live captioning that still needs readable phrasing.
• Use a medium delay level for general meeting notes and mixed speakers.
• Use high or extra-high delay when accuracy matters more than immediacy.
• Provide a language hint when you know it. It often improves accuracy and stability.

Limitations

• Audio input only. It doesn’t generate spoken audio output.
• In GA Realtime sessions, this model doesn’t support a transcription prompt. Your context text may be ignored or rejected.
• Very long speech can exceed single-message text limits. Expect long audio to be split across updates.
• Streaming transcripts can change as more audio arrives. Don’t treat early deltas as final.
• Low-quality inputs hurt results. This includes noisy rooms, crosstalk, strong reverb, clipped audio, and low-bitrate telephony.
• Mixed languages, heavy accents, or dense jargon can increase errors. Test with your real microphones and speakers.

Safety & compliance

• Get consent before recording or transcribing people.
• Don’t use transcripts as the only source of truth for high-stakes decisions.
• Review outputs when you handle names, numbers, dates, or medical and legal content.
• Follow OpenAI’s usage policies for disallowed content and sensitive data handling.