openai / gpt-live-transcribe

Overview

GPT Live Transcribe is OpenAI’s low-latency speech-to-text model for live audio. It processes audio continuously and returns text as speech arrives. It can use added context, expected terms, and language hints to reduce mistakes. It’s useful when you need readable captions or fast transcripts while people speak.

What you can build

• Live captions for meetings, webinars, and classrooms
• Real-time call transcription for customer support and contact centers
• Voice note capture with an editable transcript for summaries and follow-ups
• Moderation and monitoring dashboards that need text from live audio
• Searchable transcripts for long recordings, then downstream indexing

Inputs

• One audio file to transcribe. The file can be up to 60 minutes and 500 MB.
• Optional background context about the recording. Provide a short paragraph that describes the setting and topic. Keep it under 5,000 characters.
• Optional list of literal terms you expect to hear, like product names, acronyms, or ticket IDs. Separate terms with semicolons.
• Optional list of expected spoken languages using ISO codes (for example: en, es, fr, zh-cn). Separate codes with semicolons. Leave it empty to let the model infer.
• Optional latency choice that trades speed for accuracy. Options range from Minimal (fastest) to Extra high (most accurate).

Outputs

The model returns a text transcript of the audio.

• The final result is a single, complete transcript string.
• If your workflow supports incremental updates, you can also receive partial text chunks before the final transcript. Those chunks may be revised before completion.

Recommended settings

• Live subtitles and on-screen captions: choose the Low latency setting for balanced results.
• Interactive typing-style feedback while someone speaks: choose Minimal.
• Post-call transcripts where accuracy matters more than delay: choose High or Extra high.

Limitations

• The model does not provide word-level timestamps or subtitle timing.
• The model does not provide speaker labels.
• The model does not return transcription confidence scores.
• Partial text can change as more audio arrives. Your UI should support revisions.
• Overlapping speakers, heavy background noise, clipped microphones, and strong compression can reduce accuracy.

Safety & compliance

• Get consent before recording or transcribing other people.
• Follow local laws for call recording, workplace monitoring, and notice requirements.
• Only upload audio you have rights to process.
• Follow OpenAI usage rules for sensitive content and personal data.