OpenAI: GPT Transcribe

GPT Transcribe is OpenAI’s speech-to-text model for completed audio files, streamed file transcripts and committed audio turns in Realtime sessions. It accepts audio with optional text context and returns recognized text. The producer positions it for accurate transcription of recorded speech in its original language, including domain vocabulary, multilingual recordings and code-switching. Its useful output is a transcript that stays connected to the recording, with detected-language metadata when the model can make a reliable prediction.

Preparing a completed recording

OpenAI’s file-transcription workflow accepts uploads up to 25 MB. The current endpoint reference lists FLAC, MP3, MP4, MPEG, MPGA, M4A, OGG, WAV and WebM input formats and recommends enough file metadata to identify the format, including an appropriate filename extension and content type. For a larger recording, the guide recommends compression or splitting it into files of 25 MB or less. Avoid cutting through a sentence, because removing surrounding speech can make recognition less accurate. Keep the pieces in order and retain their positions in the original recording.

For a meeting, interview or recorded presentation, supply the actual audio and identify the topic and any terminology likely to be spoken. A useful result is the recognized wording rather than an inferred summary of the discussion. Preserve the recording’s identifier with the returned text so that a person reviewing a name, number or uncertain phrase can return to the right source. A transcript can then become material for a separate summary, translation or editorial workflow.

Context, keyword hints and expected languages

The producer separates three kinds of guidance. A prompt gives free-form context about the recording’s topic or setting. Keywords give literal words or phrases expected in the speech. A languages list identifies the expected spoken languages. For a bilingual product discussion, describe the setting in the prompt, list the relevant product names as keywords, and provide the expected languages. These are recognition hints; they are not instructions to insert vocabulary that nobody said.

GPT Transcribe uses the plural languages field instead of the older singular language field; OpenAI says not to send both. Its guide accepts supported ISO 639-1 codes, selected ISO 639-3 codes and regional Chinese locale codes, with unsupported or malformed values rejected. Each keyword must stay on one line and must not contain an angle bracket, carriage return or line feed. An invalid keyword or an overlong prompt can reject the complete request. Keep context relevant and check whether hinted names actually occur in the audio rather than treating the hint list as the transcript.

JSON output and detected-language predictions

The documented file result is JSON containing the transcript and detected languages. Those language predictions are the model’s output, not a copy of the supplied language hints. An empty language list is a valid result when the model cannot make a reliable prediction; it does not establish that the recording was silent or that transcription failed. Keep the recognized text and language metadata separate so downstream processing can use each according to its meaning.

Streaming a file transcript

A completed file can produce transcript text incrementally while it is processed. OpenAI documents transcript.text.delta events for partial text and transcript.text.done for the complete file transcript; the final event also includes detected languages. This streams the output of an already supplied recording. It is useful for displaying progress on a long upload, while the final event establishes the text to retain as the completed result.

Committed turns in Realtime sessions

GPT Transcribe also has a specialized Realtime workflow over WebSocket. Audio is appended to the input buffer, then the application explicitly commits an audio turn. Transcription begins for that committed chunk and may emit deltas before its completion event, including detected-language information. The model automatically uses earlier transcribed turns as context in this producer session flow. This is distinct from continuously emitting live-caption text as microphone audio arrives; OpenAI names GPT Live Transcribe for that continuous workflow.

For several turns, retain the input-item identifier with the corresponding transcript. OpenAI warns that completion events from different turns need not arrive in speech order, so associate each final result with its committed input rather than arranging a conversation solely by completion time. The useful artifact is an ordered transcript assembled from the correct final items, with interim text treated as progress rather than a separate completed utterance.

Specialized outputs and substantive limits

This profile does not supply a universal subtitle, timestamp, translation or speaker-label contract. OpenAI’s migration guide distinguishes those specialized capabilities: native subtitle formats and word timing require a compatible model and format, while speaker labels use GPT-4o Transcribe Diarize. GPT Transcribe’s file result is JSON rather than a drop-in SRT, VTT or verbose-timestamp response. Its exact model card does not publish a numeric context or maximum-output allowance, so a chat-model token window should not be assumed for audio transcription.

Review difficult vocabulary, numbers, accents, background noise and changes of language against the recording. Check that relevant hints improved recognition without introducing unspoken terms and that the final transcript includes the end of the audio. For a multilingual result, preserve the spoken wording before preparing any translated version. These checks make the transcript useful as source material while keeping context hints, detected languages and the recognized speech clearly distinguished.