OCanada-TVNews: A Two-Year Dataset for Gender and Topic Diversity in Canadian TV News Broadcasts

Executive Summary

OCanada-TVNews is a two-year dataset (2023 and 2024) of Canadian TV news broadcasts from four major channels — CBC The National, CTV The National, Global News, and Radio-Canada Info — collected from the channels' YouTube playlists. The dataset comprises 3,982 videos totaling 999 hours of content. Each video is processed through a systematic pipeline involving audio extraction, speaker diarization, gender detection, and topic classification, and includes transcripts, speaker diarization, gender labels (female and male), topic labels (politics, economy, science, and others), and metadata on user engagement.

By merging speaker segments with gender labels and topic categories, the dataset offers a detailed view of who speaks and what is discussed on a daily basis, enabling researchers to track how coverage shifts or remains consistent over time. An exploratory analysis highlights persistent disparities in gender representation across topic areas and provides insights into the topics that dominate coverage on each channel.

The authors also release the full source code used to generate the dataset, including modules for collecting YouTube metadata, performing speaker diarization, gender recognition, topic classification, and data analysis, so that researchers can reuse and adapt the pipeline for similar datasets on other topics or channels. OCanada-TVNews is intended as a foundation for further study of media diversity, content trends, and potential biases in Canadian broadcast journalism.

Key Takeaways

• The dataset contains 3,982 videos totaling 999 hours of Canadian TV news from 2023–2024, collected from CBC The National, CTV The National, Global News, and Radio-Canada Info via YouTube.
• Across all channels, female speakers account for 538.94 hours and male speakers 460.87 hours; excluding presenters, gender representation is near-parity at 349 hours (female) vs. 358 hours (male).
• CBC provides the largest content share at 624 videos and 423.80 hours; Radio-Canada Info contributes the most videos (1,946) but shortest runtime (99.57 hours).
• The automated pipeline uses PyAnnote speaker diarization, wav2vec2 gender recognition (F1 score of 0.9993), WhisperX transcription, and multilingual news topic classification across 17 IPTC categories.
• The dataset and full source code are publicly released, enabling research on gender representation, bilingual coverage, topic diversity, and audience engagement patterns across Canadian broadcast journalism.