OpenAI Transcribe

POC for using OpenAI's audio transcribe functionality

OS Requirements

• FFMpeg

Setup

Create an .env file containing your OpenAI API keys.

Usage Notes

If you have MP4 files, you can convert them first to MP3s with FFMpeg:

ffmpeg -i video.mp4 -q:a 0 -map a video_audio.mp3

Then start transcribing it with the tool:

python transcribe.py video_audio.mp3

If you want to download a video you can use:

Fish

yt-dlp --config-location (pwd) URL

Bash

yt-dlp --config-location $(pwd) URL

Transcript → Article

Generate a formatted article from a transcript using a prompt defined in articles.toml.

• Configure article types in articles.toml (each top-level table needs a prompt and can optionally set a model).
• Use {transcript} in the prompt to interpolate the transcript text.

Run:

python transcript_to_article.py <ARTICLE_TYPE> <PATH/TO/transcript.txt>

Output is written to output/<transcript_stem>_<ARTICLE_TYPE>.md by default, or use -o/--output-dir to choose a directory.

Model Selection

You can choose which model generates each article type by setting model in articles.toml. If omitted, the script defaults to gpt-4o-mini.

Recommended options for article generation:

• gpt-4o: Highest quality and formatting fidelity. Great for polished blog posts, nuanced tone, and long transcripts.
• gpt-4.1: Strong reasoning and instruction following. Good for complex outlines and dense technical recaps.
• gpt-4o-mini: Fast and cost‑efficient. Ideal for meeting notes, summaries, and quick drafts.
• gpt-4.1-mini: Balanced quality/cost. Good for structured outlines and straightforward posts.

Tips:

• Use gpt-4o for final, publication-ready blog content.
• Use gpt-4o-mini for iterative drafting, notes, or when cost/latency matters.
• For long or messy transcripts, prefer gpt-4o or gpt-4.1 to preserve nuance.

Example config per type (from articles.toml):

[blog]
model = "gpt-4o"
prompt = "..."

[notes]
model = "gpt-4o-mini"
prompt = "..."

[outline]
model = "gpt-4.1-mini"
prompt = "..."