Skip to main content

openai/gpt-transcribe/realtime-transcription

OpenAI's GPT Transcribe over a transcription WebSocket turns each committed audio segment of up to 60 seconds into text, using context notes, keyword hints and language hints, and carries earlier turns forward as context.

From $0.000015/s80% below the publisher's list price

Workbench

Input

Realtime transcription
0/2048
Explain the recording’s context to GPT Transcribe, such as the discussion topic or specialist vocabulary; do not request an answer or summary. at most 2048 characters
1/1
1 / 1 files
Submit a short compatible WAV to check a single GPT Transcribe segment; its audio is sent through the realtime transcription protocol. at least 1 characters
affects price
Choose GPT Transcribe’s admitted segment length between 1 and 60 seconds; the final reported duration determines settlement. no lower than 1 · no higher than 60
Safety checkerOff

SpicyAPI does not run a safety checker on this model. Your prompt and media go to the model as written, and the model itself decides what comes back.

Checking your account before enabling submission and pricing.

Output

Ready for a new run
Example result
Model response

Missing the model you need?

Two ways to transcribe speech

Key numbers

Segment maximum
60s

What it does

Two ways to transcribe speech

GPT Transcribe accepts an unstructured context note, domain keywords and multiple expected languages. A context note can identify the subject under discussion, while vocabulary hints distinguish an unfamiliar product name from an ordinary word. The model can also use earlier transcribed turns within the continuing session, which is useful when a later phrase refers to a name established earlier. The committed-turn workflow gives an application an explicit point at which to close an audio segment. Transcript deltas can then populate a text preview until the final segment is available. This endpoint meters the final audio duration reported by the model; a caller chooses the maximum admitted segment length before sending samples. GPT Transcribe does not expose Live Transcribe’s adjustable delay profile on this connection. The language hints suit a discussion in which more than one language may appear. For example, a speaker might explain a release in English and then quote its French product name. Give the model the languages that actually occur in that material so that unusual words have plausible linguistic context. Keywords work at a different level from language choice. They can supply an organisation name, a device identifier or a technical term that people pronounce in different ways. A short description of the subject can accompany them. The file result contains the recognized speech as text suitable for searching and correcting.

Where it fits

  • Spoken data entry

    For a spoken data-entry form using GPT Transcribe, let the user finish a field, commit that recording and show the resulting text for correction. Confirm identifiers or amounts in the application before saving them to the record.

  • Interview segment review

    For an interview-review tool, send one bounded utterance at a time with a short description of the topic and names. Keep the association between each submitted segment and its completed text so a reviewer can revisit ambiguous passages.

Model reference

GPT Transcribe API reference

How to call the GPT Transcribe API

  1. Prompt guide & Pro tips

    Choose the file endpoint for a finished recording. Choose realtime transcription to receive text as each spoken turn is committed. Prepare GPT Transcribe context as background about the subject, audience or terminology. Supply the proper spelling of specialist names through keyword hints, and choose likely language codes when the conversation is multilingual. Do not use the context field to ask for a summary or an answer: the task is to recognize speech, so editorial processing should happen after the transcript is finalized. Use manual commits at clear utterance boundaries and keep each segment within the selected one-to-sixty-second limit. The streaming input must match 24 kHz mono 16-bit PCM or 8 kHz G.711; the short-WAV task preserves that same encoding. Review words that matter before triggering a workflow. Finalization occurs per segment, allowing the next segment to be admitted independently. This is the committed realtime-transcription endpoint; the separate file-transcription API has a different request format and is not implied by a WAV test task.

GPT Transcribe API questions

All GPT models18