GPT Realtime Whisper

GPT-Realtime-Whisper is a specialized OpenAI streaming speech-to-text model designed for realtime applications that require low-latency transcript deltas emitted while the speaker is still talking. The model lets developers tune the trade-off between latency and transcription accuracy.

Transcription sessions use a session type of 'transcription' and support both WebSocket transport (24 kHz mono PCM, base64-encoded) and WebRTC. Server VAD (voice activity detection) can be configured with threshold, prefix_padding_ms, and silence_duration_ms parameters, or audio buffers can be committed manually. The model emits conversation.item.input_audio_transcription.delta and .completed events.

Typical use cases include live captions, meeting transcription, lecture capture, telephony transcription, broadcast captioning, and dictation. Pricing is based on audio duration (USD per minute) rather than tokens. Within OpenAI's transcription model family, it is positioned as a natively streaming alternative to GPT-4o Transcribe, GPT-4o mini Transcribe, and Whisper-1.