GPT-4o Transcribe — OpenAI's Audio Transcription Model 📝🖼️🔢 → 📝

⭐ Official ▶️ 65.8K runs 📅 May 2025 ⚙️ Cog 0.16.8 ⚖️ License
speech-to-text transcription audio openai

Performance

2.9sTypical run time
66 tok/sThroughput
64.3KTotal runs

Data sampled 2026-07-12

GPT-4o Transcribe is OpenAI's GPT-4o-based speech-to-text model. Run it on Replicate to transcribe audio quickly, without setting up OpenAI infrastructure yourself.

About

GPT-4o Transcribe converts speech to text using OpenAI's multimodal GPT-4o architecture. It handles multiple languages, accents, and background noise, and produces clean, punctuated text — practical for meeting notes, podcast transcripts, subtitles, and voice-to-text workflows.

On Replicate it is one of the fastest transcription options available: the model streams output token by token and typically returns a short clip's transcript in a few seconds.

How it compares to other transcription models on Replicate

Model Strength Word-level timestamps Best for
openai/gpt-4o-transcribe High accuracy, prompt-biased vocabulary No Clean transcripts, context-aware wording
openai/whisper Open, mature, widely used Yes Subtitles / captions needing timestamps
vaibhavs10/incredibly-fast-whisper Batched throughput Yes Long-form audio at scale

Key inputs that matter

  • audio_file (required) — the clip to transcribe.
  • language — set it to skip auto-detection and improve accuracy on known-language audio.
  • prompt — bias the model toward domain vocabulary, names, or acronyms it might otherwise mis-hear.
  • temperature — lower values produce more deterministic transcripts.

When to use this model

  • You want high transcription accuracy with correct punctuation out of the box.
  • Your audio contains domain jargon or proper nouns you can hint via prompt.
  • You need results fast for interactive or near-real-time flows.

When to reach for something else

  • You need word-level timestamps for captions or alignment — use a Whisper variant.
  • You need speaker diarization (who-said-what) — pair with a diarization model; this one does not label speakers.
  • You are transcribing very long files in bulk at the lowest cost — batched Whisper models are cheaper per hour.

Limitations and gotchas

  • No built-in timestamps or speaker labels.
  • Processes one audio file per request — batch by orchestrating multiple calls.
  • Very long recordings may need to be chunked before submission.

Example Output

Output

So we just added GPT-4o transcribe to Replicate and thought you'd want to know. It's basically a speech-to-text model that uses GPT-4o to turn your audio into text. The cool thing is that it's noticeably better than the Whisper models we've been using, fewer errors, better at recognizing different languages, and just more accurate overall. If you've ever been frustrated with transcripts that mess up technical terms or struggle with different accents, you'll probably appreciate this upgrade. It just works better. Some quick tech specs if you're curious. It has a 16,000 token context window, which means it can handle longer audio clips in one go. And it can output up to 2,000 tokens, so you'll get nice complete transcripts. The model's knowledge is current up to June 2024, so it's pretty up-to-date with language and terminology.

Performance Metrics

2.89s Prediction Time
2.90s Total Time
All Input Parameters
{
  "language": "en",
  "audio_file": "https://replicate.delivery/xezq/XoxHeakty0z3KKc46cMLPKC2ct54ekT3EtvcwDQuRIuxfJdpA/tmpsglqtqn5.mp3",
  "temperature": 0
}
Input Parameters
prompt Type: string
An optional text to guide the model's style or continue a previous audio segment. The prompt should match the audio language.
language Type: string
The language of the input audio. Supplying the input language in ISO-639-1 (e.g. en) format will improve accuracy and latency.
audio_file (required) Type: string
The audio file to transcribe. Supported formats: mp3, mp4, mpeg, mpga, m4a, ogg, wav, or webm
temperature Type: numberDefault: 0Range: 0 - 1
Sampling temperature between 0 and 1
Output Schema

Output

Type: arrayItems Type: string

Example Execution Logs
Input audio duration: 54.756 seconds
Input token count: 912
Output token count: 174
Total token count: 1086
TTFT: 1.38s
Version Details
Version ID
cc7638666fc85e9defb010d99e304c0c0e94dcdbd3d31385f28f2730b4cdcc6d
Version Created
November 7, 2025
Run on Replicate →