Gpt-4o-transcribe truncates output after ~8-9 minutes even on short segments

API Bugs
1

Hi everyone,

I’ve run into a frustrating issue with the gpt-4o-transcribe (and gpt-4o-mini-transcribe) models. No matter how I prepare my audio, the transcription output always gets truncated after about 8–9 minutes of audio.

Here’s what I’ve tried so far:

  • Converted the source video (.mkv) into clean audio chunks using ffmpeg.

  • Made sure each chunk is mono, 16kHz, normalized with loudnorm, and low/high-pass filtered for clarity.

  • Exported to .m4a (AAC) instead of MP3 to avoid VBR issues.

  • Limited file sizes to well under 25 MB.

  • Limited durations first to 1400s (~23 min), then much shorter 540s (~9 min), even down to 480s (8 min).

  • Sent requests with response_format=json instead of text.

  • Tried both gpt-4o-transcribe and gpt-4o-mini-transcribe.

Despite all of that, the API still only returns text for the first ~8-9 minutes of audio. The rest is cut off completely — no error message, just a truncated transcription.

What’s interesting:

  • The same chunks transcribed with whisper-1 return the full transcript as expected.

  • So the problem seems to be specific to the 4o models.

My questions:

  • Are there recommended best practices to avoid truncation (e.g. maximum safe segment length)?

  • Has anyone found a reliable workaround besides falling back to Whisper?

  • Or is this something the dev team is aware of and working on fixing?

Any insights would be really helpful. Thanks!

2

First of all, the stated “under 25 MB limit“ is a joke. Our tests indicate the limit to be roughly half of that (about 13.30 minutes for .mp3s at mono, 128bps, 22.05 kHz). This is the isssue that OpenAI needs to resolve.

We have not revert to Whisper because we think OpenAI will eventually get arround to increasing the real limit.

3

It could be that you’re hitting the 2,000 output tokens limit of the gpt-4o-transcribe model.

4

Ouch, seems like you are right. Just changed output format to the json and I see “output_tokens”:2048