GPT Transcribe turns meetings, calls, and streams into clean text. Use gpt-transcribe for finished files and gpt-live-transcribe for low-latency live captions — with scene prompts, keywords, and multilingual hints.
GPT Transcribe
Live transcription preview
Customer called about upgrade plan AC-42. Confirmed billing cycle, noted Spanish code-switch mid-call, and booked a follow-up for Thursday 3pm.
Online generator
Upload a recording for gpt-transcribe, or speak live for streaming captions. Add optional scene context, keywords, and language hints for higher accuracy.
Drop audio here or click to upload
mp3, mp4, m4a, wav, webm · max 25MB · powered by gpt-transcribe
Context is a first-class feature of the new models: describe the scene, list expected terms, and hint mixed languages. Keywords only appear if spoken.
Transcript
Editable result — copy when ready
How to use
Upload or record audio, optionally add a scene prompt plus keywords and languages, run transcription, then copy and lightly edit the text.
Choose file mode for finished recordings, or live mode to capture microphone audio with interim captions.
Describe the scene, list expected terms, and hint languages. Skip any field you do not need.
File jobs use gpt-transcribe. Live sessions stream interim text, then finalize the capture for a cleaner transcript.
Edit the result inline, copy to your clipboard, and re-run with better keywords if a critical term was missed.
Two models, clear jobs
OpenAI split transcription into two recommended entry points. Pick by whether the audio is already finished — or still being spoken.
Optimized for completed audio files and batch workloads: meetings, podcasts, support recordings. Async transcription with optional progressive results.
Built for low-latency live transcription over WebSocket / WebRTC: captions, calls, streaming speech that is still being produced.
Straightforward workflow
Upload or speak, add context the model can actually use, then copy a draft you can lightly edit instead of re-listening the whole recording.
01Drop a finished file for gpt-transcribe, or open the mic for live capture and interim captions.
02Optional scene prompt, keyword list, and language hints — trained so the new models actually improve when you provide them.
03Review the transcript, copy it into notes or docs, and iterate on context if specialized terms need another pass.
What improved
The real leap is not just lower WER. New models learn to use free-form scene descriptions, keyword lists, and multi-language hints. Older models barely moved when given the same context.
+3.6 / +6.1
semantic accuracy points gained with context (file / live)
57+
languages shown in OpenAI demos (verify for your locales)
Tell the model this is a support call, clinical note, or product demo so names, IDs, and domain phrasing land correctly.
Pass product names, drug names, or acronyms. They only appear if spoken — not forced into the transcript.
Supply multiple language codes (e.g. en + es or en + zh-cn). Mid-utterance switches no longer need a manual language flip.
Stronger on accents, short phrases, numbers, specialized terminology, and speech under loud background noise.
Built for real work
From archived recordings to live captions — pick the model that matches whether audio is finished or still streaming.
Batch-transcribe long recordings with gpt-transcribe, then light-edit the draft instead of replaying the whole session.
Feed account IDs and plan names as keywords so order numbers and product terms survive noisy phone audio.
Stream speech with gpt-live-transcribe for calls, classrooms, and events that need words on screen as they are spoken.
Dictation, field notes, and multilingual voice memos with optional language hints for mixed speech.
FAQ
Practical answers about gpt-transcribe, gpt-live-transcribe, and the online generator.
Start with the online generator — file transcription via gpt-transcribe or live capture for streaming captions.