So yeah, I think I’ve found a solution to the truncation problem with the GPT-4o-transcribe model from OpenAI. Either my solution is the reason transcripts are no longer getting truncated, or OpenAI has made some changes to the model that fixed the issue. However, since I haven’t seen any official statement or announcement from OpenAI about this, I’m assuming the model is still the same as it was when it was released in March.

Anyway, I want to show you the simple solution I discovered. Previously, whenever I used GPT-4o-transcribe, I didn’t supply any prompt in the API call—I just used the default settings. But this time, I tried adding a prompt that specifically told the model not to truncate, omit, summarize, or clean up anything, and to transcribe every spoken word as accurately as possible.

After I did this, I noticed the truncation issue disappeared. The only downside was that the transcripts could be a little disorganized. To improve this, I set the temperature to 0.2. Since then, I’ve been getting really high-quality transcripts—honestly, they beat Whisper 1 almost every time.

I’ve done a lot of tests and experiments. I use this API for client consultations at work, and I’ve run both Whisper 1 and GPT-4o-transcribe in parallel. After these changes with the prompt and temperature, about 99% of the time, the GPT-4o-transcribe transcripts are much better quality than Whisper 1, and I no longer get incomplete or truncated transcripts.

Has anyone else tried this? Or were you aware of this workaround? Maybe give it a try and see if you get the same results. For a long time, I stuck with Whisper 1 because GPT-4o-transcribe had these issues, but now, with this custom prompt and temperature tweak, the problem seems totally solved on my end.

I’ll share the exact prompt I’m using below so you can try it out too. Let me know if it works for you or if you have any feedback!

"formData.append(“model”, “gpt-4o-transcribe”);
formData.append(“temperature”, “0.2”);
formData.append(“prompt”,
“Only transcribe spoken words; exclude all non-verbal and background noises.” +
“Do NOT omit, summarize, or “clean up” anything related to spoken words. " +
“Output every word as spoken. Do NOT truncate or leave out anything in the transcript, that is spoken”
);”