Audio Summarization for Podcasts

Aneesh Vartakavi, Amanmeet Garg, Zafar Rafii

Abstract: We propose a novel system to automatically generate audio summaries for podcasts, allowing listeners to quickly preview podcast episodes. The proposed system first transcribes the audio from a podcast using automatic speech recognition (ASR), then summarizes the transcript using extractive text summarization, and finally returns the audio associated with the text summary. Motivated by a lack of relevant datasets for this task, we created our own by transcribing the audio from various podcasts and generating summaries for these transcripts using a manual annotation tool. Using these text summaries, we fine-tuned a recent Transformer-based summarization model to specifically handle podcast summaries. Our system achieves ROUGE-(1/2/L) F-scores of 0.63/0.53/0.63, respectively, showing good performance for podcast summarization. We present some examples of podcast audio summaries here: https://github.com/aneeshvartakavi/podsumm.