Opening Podcast Audio to the Research Community
Podcast listening on Spotify has grown rapidly since 2015, with hundreds of thousands of shows now available. Acquisitions like Gimlet, Parcast, and Anchor have brought a mix of professionally produced content and creator tools that let anyone publish a show. That growth makes one question more pressing than ever: how do listeners find the content that matters to them in an increasingly crowded field?
Spotify is already exploring related questions internally — distinguishing podcasts that interview a figure like Barack Obama from those that merely discuss him, identifying the key moments in a long episode, and defining what makes a podcast informative or engaging. To accelerate that work, the company is opening up its data to external researchers through the Spotify Podcast Dataset and TREC Challenge 2020.
The First Large-Scale Podcast Dataset with Transcripts
The dataset is the first public release of its kind: 100,000 episodes from thousands of shows, each accompanied by its raw audio file, RSS header metadata (title, description, publisher), and an automatically generated transcript. The initial release is English-only, but multilingual versions are planned for the future.
The episodes span a deliberately wide range of sources and formats. Professionally produced shows tend to have clean audio and polished structure, while amateur podcasts introduce significant variability in both recording quality and delivery. Content covers coarse and fine-grained topics — lifestyle, storytelling, sports, news, health, documentary, and commentary — and includes scripted monologues, unscripted conversations, interviews, debates, and segments with non-speech audio. This diversity is intended to reflect the realistic conditions of podcast search and discovery.
Transcripts are word-level and timestamped, as in this sample:
Two Tasks: Segment Search and Summarization
The challenge, run as a shared task at the TREC 2020 Conference hosted by the U.S. National Institute of Standards and Technology, centers on two problems. The first is search: given a natural language query like "current status of legalization of medical marijuana," systems must return a ranked list of podcast segments with specific start indices. This moves beyond episode-level retrieval to pinpoint the exact portion of audio where the topic is discussed, saving users from scanning through minutes of unrelated material.
The second task is summarization: given a full episode's audio and transcript, systems must generate a concise, human-readable summary of the entire episode. Effective summaries would give listeners a quick, reliable basis for deciding whether to press play.
Timeline and Participation
The challenge is designed as a multi-year effort, with tasks expected to grow in complexity over time. For 2020, the schedule runs as follows:
- Dataset release: April 16
- Official task guidelines: May 1
- Experimental work: summer months
- Experiments conclude: September
- Results reported: November
The dataset will be available at the challenge site, where researchers can also join the mailing list. Registration for the challenge and data access is handled through TREC's submission system.
