Use speech to text models and LLMs to process Youtube videos and podcasts 10x as fast.
This is a personal tool I created which makes it easy to find interesting moments within longer podcasts and youtube videos. Video & audio content is often harder to parse/skim through than written content, but running tool makes this easier to manage.
This app is also a great demonstration of AssemblyAI's product suite which includes automatic speech recognition, automatic chapter generation, and LeMUR - a framework for pairing transcripts with LLMs.
I recommend that you start by running this project locally. If you want to deploy it, you can do so using the Streamlit community cloud or your method of choice.
This is a streamlit app, with all logic in qa.py
Follow these steps to run the app:
This app uses Supabase for persistent transcript storage.
Create a new project, then add the Supabase URL and Supabase Key as local environment variables like this:
export SUPABASE_URL=YOUR_SUPABASE_URLexport SUPABASE_KEY=YOUR_SUPABASE_KEY
You'll need to create a new table in supabase called transcripts and add 2 text columns: content_name and transcript_id
You can rename the table and the columns if you'd like, but if you do so be sure to make the modifications necessary in `qa.py``
Head to assemblyai.com and sign up for an API Key, then add this to your local env variables:
export ASSEMBLYAI_API_KEY=YOUR_AAI_KEY
Once you have Supabase and AssemblyAI ready to go, you can install deps and run the app.
pip install -r requirements.txt
streamlit run qa.py