Podcast are a growing spoken medium that is listened to for various reasons in various situations (e.g., for entertainment or educational purposes, on the train or at home) consisting of various types of audio such as (unstructured) speech, music, and other sounds. Traditionally, search and recommendation of spoken content focuses on topical content, derived from text transcriptions, ignoring paralinguistic aspects in spoken language. Instead, we propose to model these paralinguistic aspects, such as speaking style, in podcasts to address both the heterogeneity of type of audio in podcasts and user needs to enable enriched access to this medium. In this paper, we take a first step towards this goal and explore audio-based stylistic variation in podcasts by 1) investigating what facets of stylistic variation are salient and of interest to listeners, and 2) gathering more insights into the kind of stylistic variation that is currently feasible to model with open-source audio tools and that is present in podcasts. We find that much of the stylistic variation mentioned by the users is related to speaking style and music, and we show, using open-source tools, how audio-based stylistic aspects vary across episodes, shows, and genres.
Cite as: Martikainen, K., Karlgren, J., Truong, K. (2022) Exploring audio-based stylistic variation in podcasts. Proc. Interspeech 2022, 2343-2347, doi: 10.21437/Interspeech.2022-10871
@inproceedings{martikainen22_interspeech,
author={Katariina Martikainen and Jussi Karlgren and Khiet Truong},
title={{Exploring audio-based stylistic variation in podcasts}},
year=2022,
booktitle={Proc. Interspeech 2022},
pages={2343--2347},
doi={10.21437/Interspeech.2022-10871}
}