MIDS Capstone Project Summer 2021

Too Long; Didn't Listen (TL;DL)

There are 2 million podcast shows and 48 million episodes, with average podcast episode lengths of 43 minutes. TL;DL provides concise, accurate, and comprehensive summaries of podcasts to help podcasters and listeners improve podcast content discovery and uptake. We found that in user surveys, podcast listeners were most interested in using TL;DL summaries to discover new podcast content and to catch up on episodes when short on time. Podcast creators would also benefit from TL;DL, as producing accurate transcripts and summaries of podcast episodes also takes a lot of time.

TL;DL uses Google Cloud’s Speech-to-Text API to transcribe podcast audio files, then applies an abstractive summarization model to generate comprehensive synopses. TL;DL leverages the PEGASUS model's architecture, in which a transformer-based encoder-decoder models are pretrained on large text corpora. This pretraining involves masking whole sentences from a document and generating these gap-sentences from the rest of the document.

The TL;DL tool is now available on a web app, from which a user is able to upload an audio file of a podcast and generate a summary in a few minutes.

Last updated: August 5, 2021