Joy introduced as a dual-channel dataset with a claimed 100,000 hours of conversation

@shobhitbanga introduced Joy as a dataset of unscripted two-person conversations, claiming 100,000 hours across 10 languages. The poster intends it to train voice models on when to listen and respond. Other posters describe overlaps and interruptions, but the introductory post does not specify those details. It also gives no download link or training results.