Podcast Transcripts (BOW)

Description:

This dataset provides bag-of-words (BOW) representations of podcast transcripts derived from a large-scale corpus of English-language podcasts. The corpus spans all 19 Apple Podcasts genres and includes over 15,000 episodes from approximately 1,900 shows.

Transcripts were generated using Hugging Face's Distil-Whisper-Large-v3 automatic speech recognition model. Text is lowercased but otherwise unprocessed.

Files:

bow_matrix.npz — Sparse bag-of-words matrix (scipy sparse format, documents × terms)

metadata.csv — Podcast title, episode title, and description for each document (row indices correspond to matrix rows)

vocabulary.csv — Terms in the vocabulary (column indices correspond to matrix

Usage:

from scipy import sparse
import pandas as pd

bow_matrix = sparse.load_npz('bow_matrix.npz')
metadata = pd.read_csv('metadata.csv')
vocabulary = pd.read_csv('vocabulary.csv')['term'].tolist()

Related publication:

For a detailed description of the data collection and transcription process, see:

Verreyen, L. (2025). Podcasts as Data: Building a Dataset for Large-Scale Audio Content Analysis. In T. Arnold, M. Fantoli, & R. Ros (Eds.), Computational Humanities Research 2025 (Vol. 3, pp. 231–248). Anthology of Computers and the Humanities. https://doi.org/10.63744/QgeF94c0fP7D