Sounder Transcript Highlight Prediction

Overview

Utilizes an implementation of TextRank to extract a highlight from a given transcript along with the associated start and end timestamps from the audio.

Pros

• Smaller, simpler, and faster than many other methods.
• Only takes between a few seconds to minutes to return a highlight.

Cons

• TextRank is a somewhat basic means of extracting important information from text.
• This implementation of TextRank only extracts single sentences, though it should be relatively easy to rewrite the TextRank implementation accept a window size for how much it should consider and return (e.g. 3 consecutive sentences instead of 1).
• This only looks at the text context of the podcast, so you lose a lot of additional information like tone and volume from the audio that could also indicate some portion of it should be highlight.

Requirements

A system with Docker installed

Build

Navigate to the highlight directory (the one this file is in) and run:

docker-compose build

Start

After building, from the highlight directory:

docker-compose up

Use

• GET 127.0.0.1:8000/: Returns status for healthchecks
• POST 127.0.0.1:8000/predict: POSTing a transcript JSON to this endpoint returns the predicted highlight text as well as the start and end timestamps