100,000 Podcasts: A Spoken English Document Corpus

Proceedings of the 28th International Conference on Computational Linguistics, pages 5903–5917 Barcelona, Spain (Online), December 8-13, 2020 5903 100,000 Podcasts: A Spoken English Document Corpus Ann Clifton Spotify aclifton@spotify.com Sravana Reddy Spotify sravana@spotify.com Yongze Yu Spotify yongzey@spotify.com Aasish Pappu Spotify aasishp@spotify.com Rezvaneh Rezapour∗ University of Illinois at Urbana-Champaign rezapou2@illinois.edu Hamed Bonab∗ University of Massachusetts Amherst bonab@cs.umass.edu Maria Eskevich CLARIN ERIC maria@clarin.eu Gareth J. F. Jones Dublin City University gareth.jones@dcu.ie Jussi Karlgren Spotify jkarlgren@spotify.com Ben Carterette Spotify benjaminc@spotify.com Rosie Jones Spotify rjones@spotify.com Abstract Podcasts are a large and growing repository of spoken audio. As an audio format, podcasts are more varied in style and production type than broadcast news, contain more genres than typi- cally studied in video data, and are more varied in style and format than previous corpora of conversations. When transcribed with automatic speech recognition they represent a noisy but fascinating collection of documents which can be studied through the lens of natural language processing, information retrieval, and linguistics. Paired with the audio files, they are also a re- source for speech processing and the study of paralinguistic, sociolinguistic, and acoustic aspects of the domain. We introduce the Spotify Podcast Dataset, a new corpus of 100,000 podcasts. We demonstrate the complexity of the domain with a case study of two tasks: (1) passage search and (2) summarization. This is orders of magnitude larger than previous speech corpora used for search and summarization. Our results show that the size and variability of this corpus opens up new avenues for research. 1 Introduction Podcasts come in many formats and levels of formality. Episodes appear on a regular or irregular ca- dence. They can be formal news journalism or conversational chat; fiction or non-fiction. They are sharply growing in popularity (Whitner, 2020) and yet have been relatively little studied. This medium opens up a rich palette of questions and issues for research in speech and language technology, linguis- tics, information access, and media studies. To facilitate research into podcasts, we have produced a corpus of podcast episodes, comprising nearly 60,000 hours of speech. This is orders of magnitude larger than previous transcribed speech datasets, and contains a rich variety of genres, subject matter, speaking styles, and structural formats. Our contribu- tions are four-fold: • The largest corpus of transcribed speech data, from a new and understudied domain, • A set of labeled data for retrieval and summarization on this corpus, • Benchmarking results for retrieval and summarization tasks using standard baselines, • An analysis of the data and benchmarking results, highlighting domain differences from vanilla versions of these tasks to motivate areas of future research. The corpus can be accessed at podcastsdataset.byspotify.com. ∗ Work done while at Spotify This work is licensed under a Creative Commons Attribution 4.0 International License. License details: http://creativecommons.org/licenses/by/4.0/. 5904 2 Related Datasets Earlier speech corpora contained relatively clean audio, often with a single speaker reading from a pre- pared text, such as the TIMIT collection (Garofolo et al., 1990) or broadcast news corpora, which have been …

Content truncated to 500 words