[Paper Review] The Spotify Podcasts Dataset

Ann Clifton, Aasish Pappu+5 more|arXiv (Cornell University)|Apr 8, 2020

Radio, Podcasts, and Digital Media5 references4 citations

TL;DR

The paper introduces the Spotify Podcast Dataset, a large-scale collection of approximately 100,000 podcast episodes with raw audio and automatic speech recognition (ASR) transcripts, totaling over 47,000 hours of transcribed audio. This dataset, an order of magnitude larger than prior speech-to-text corpora, enables new research in speech processing, information retrieval, and NLP for diverse podcast formats and genres.

ABSTRACT

Podcasts are a relatively new form of audio media. Episodes appear on a regular cadence, and come in many different formats and levels of formality. They can be formal news journalism or conversational chat; fiction or non-fiction. They are rapidly growing in popularity and yet have been relatively little studied. As an audio format, podcasts are more varied in style and production types than, say, broadcast news, and contain many more genres than typically studied in video research. The medium is therefore a rich domain with many research avenues for the IR and NLP communities. We present the Spotify Podcast Dataset, a set of approximately 100K podcast episodes comprised of raw audio files along with accompanying ASR transcripts. This represents over 47,000 hours of transcribed audio, and is an order of magnitude larger than previous speech-to-text corpora.

Motivation & Objective

• To address the lack of large-scale, diverse audio corpora for podcast-specific research in NLP and information retrieval.
• To provide a rich, heterogeneous dataset capturing the stylistic and formal diversity of modern podcasts, including news, chat, fiction, and non-fiction formats.
• To support research in speech-to-text transcription, audio understanding, and multimodal analysis across a broad range of podcast genres and production qualities.
• To enable large-scale studies of spoken language variation, speaker styles, and content diversity in on-demand audio media.

Proposed method

• Collection of approximately 100,000 podcast episodes from Spotify’s platform, covering a wide range of genres and formats.
• Inclusion of raw audio files and corresponding automatic speech recognition (ASR) transcripts for each episode.
• Aggregation of data across diverse podcast types, including formal journalism, informal conversations, and narrative storytelling.
• Creation of a dataset with over 47,000 hours of transcribed audio, significantly expanding the scale of existing speech-to-text corpora.
• Standardization of metadata and alignment of audio and text for consistent downstream use in NLP and IR tasks.

Experimental results

Research questions

• RQ1 How can large-scale, diverse podcast corpora improve the training and evaluation of automatic speech recognition systems?
• RQ2 What insights can be gained about spoken language variation and stylistic diversity in on-demand audio media?
• RQ3 How does the inclusion of informal, conversational, and narrative podcast formats affect NLP model performance and generalization?
• RQ4 To what extent can this dataset support cross-genre and cross-format research in information retrieval and audio understanding?

Key findings

• The dataset comprises approximately 100,000 podcast episodes, representing over 47,000 hours of transcribed audio, making it an order of magnitude larger than previous speech-to-text corpora.
• The dataset captures a broad spectrum of podcast formats, including formal journalism, casual conversation, fiction, and non-fiction, enabling diverse research applications.
• The inclusion of raw audio and ASR transcripts allows for end-to-end evaluation and development of speech and language processing models.
• The dataset's scale and diversity offer new opportunities for training and benchmarking models in low-resource and multilingual settings.
• The availability of such a large, real-world audio corpus supports research in spoken language understanding, speaker characterization, and content classification.