GPT Realtime Whisper overview

Learn about GPT Realtime Whisper, a streaming transcription model for low-latency real-time captioning and monitoring of live audio.

GPT Realtime Whisper is a streaming transcription model that converts live audio to text in real time. You can use it alongside speech-to-speech and translation models to provide continuous input transcription for audio streams.

Key capabilities

• Streaming transcription: Transcribes live audio as it arrives, without waiting for the utterance to complete.
• Low latency: Designed for real-time scenarios where delays aren’t acceptable, such as live captions or quality monitoring.
• Parallel operation: Runs alongside other realtime models (such as GPT Realtime Translate) to provide source-language transcription in parallel with translation.

When to use GPT Realtime Whisper

Use GPT Realtime Whisper when you need:

• Live captions and subtitles for ongoing audio streams.
• Transcription for monitoring, moderation, or analytics workflows.
• Original-language speech captured alongside live translation experiences.
• Text visibility into spoken input while other models process the audio.

Example use cases

• Live event captioning: Provide real-time captions in the speaker’s original language during conferences, webinars, or broadcasts.
• Compliance and quality review: Capture the original conversation as text for regulatory compliance, quality assurance, or analytics.
• Multilingual pipelines: Pair with GPT Realtime Translate to deliver both translated output and a source-language transcript in a single workflow.

Get started

GPT Realtime Whisper is available through the Realtime API. The connection and usage patterns are the same as for other realtime models:

• Use the GPT Realtime API via WebSockets
• Use the GPT Realtime API via WebRTC

Use the same quickstart code patterns, and set your deployment name to the model deployment you created for gpt-realtime-whisper.

Language support guidance

GPT Realtime Whisper is designed for multilingual transcription. Accuracy can vary by language, audio quality, and speaking conditions.

• Provide an ISO-639-1 language hint (for example, en) when available.
• Validate transcription quality with production-like audio before rollout.
• For broader language and locale references, see Language and voice support for the Speech service.

Deployment and availability

GPT Realtime Whisper is available as a Global Standard (pay-as-you-go) deployment in Microsoft Foundry. Deploy the model from the model catalog.

For current region support, see Region availability for Foundry Models sold by Azure.

Pricing

gpt-realtime-whisper is priced hourly. For current rates, see the Audio Models section on the Azure OpenAI pricing page.