What's the best solution for speeding up proofreading?

What we do

Proofread transcribed text(Japanese) to improve its readability by using GPT4.
Meetings we transcribe would be 30min - 1h30min long, and it takes about 25 mins for proofreading.

Want

Speeding up the process of proofreading with properly improved result.
Preferrably faster than 20 mins.

  • Completing proofreading within seconds for transcripts that costs 1,000 tokens.
  • Completing proofreading within a minute for transcripts that costs 2,000 tokens.

Known issue

More tokens will take more time.
1,000 tokens for transcript takes 1 minute for proofreading, and ChatGPT times out if it takes over 6 minutes for generating texts.

prompt we use

system_prompt

# Directive

# Meeting Transcription

The meeting transcription is a written record of the contents of a meeting conducted in Japanese. The content of the meeting involves a certain company explaining its financial results to investors. Please answer #User Commands considering the following points:

# Notes

* The transcription is performed by a machine learning model, and its accuracy is not 100%.
* The transcription results may include descriptions related to the progress of the meeting and background noise, in addition to the participants' remarks. Please ignore any irrelevant noise.
* The transcription results include fillers. Fillers are words or sounds used in conversation to fill gaps while choosing the next words, such as "um," "uh," and "but..."
* The first half of the meeting consists of explanations from the company's directors and officers, while the second half is a question and answer session.

# Meeting Transcription

''' {transcript} ‘''

user_prompt

# User Directive

If background music is played before or after the start of the meeting, it may lead to incorrect transcriptions like the examples in #Silent Transcription Examples, even though there are no spoken words. Please delete such incorrect transcriptions from #Meeting Transcription. If no incorrect transcriptions are found, please return the #Meeting Transcription as it is.

# Examples of Silent Transcription

---

* Yes, hmm.
* Oh, hmm.
* Wow!

---

# Notes

---

* Do not alter the original meaning of the text.
* Do not omit or change numbers.
* Do not omit or change the subject.

That would be amazing if anyone could advise any solutions to improve the speed. Feel free to ask me if you need further information.

Thanks!