llm_podgen

Rough Overview

1. Create prompt engineering pipeline that takes user input and generates a podcast script

2. Feed podcast script into text to audio pipeline

   • Elevenlabs API is the best API for now, but there are still artefacts and randomness with each generation

Getting Started:

1. Run the following setup

python3 -m venv .venv
source .venv/bin/activate
# pip3 install pandas requests configparser pydub static-ffmpeg
pip install -r requirements.txt

2. Then run this python code in .venv
   • Its important you add this in your virtual environment

static_ffmpeg.add_paths()

3. Rename config_template.ini to config.ini and save your api keys
4. Create a podcast script file in data/podcast_scripts as <podcast_name>.txt
   • Interviewer paragraphs should start with **host:**
   • Guest paragraphs should start with **guest:**
   • Will update to better tags later, that are not markdown lol
   • example script
5. Define podcast script file name in the pod_script_fn variable
6. Alternative speaker ids can be found in speakers_eleven.json

Test Case

Health system-scale language models are all-purpose prediction engines

• jupyter
• Used perplexity to generate podcast script from this publication
• perplexity prompt
• stitched together final audio