Casual Conversations Speech Recognition Dataset

Casual Conversations is designed to help researchers evaluate the accuracy of their computer vision and audio models across a variety of ages, genders, visible skin tones, and ambient lighting conditions in an effort to eliminate AI bias.

The dataset contains more than 45,000 videos of 3,011 participants, evenly distributed across genders, age groups, and skin colors.

Facebook asked paid participants to submit videos and provide their own age and gender labels to eliminate errors as much as possible. In addition, Facebook also recruited some trained annotators for Casual Conversations. These annotators marked the light level in each video to help measure how the AI model treats people of different skin colors in low-light environments.