Voice Arena releases Joy, a 100,000-hour dataset of two-person conversations

Voice Arena has released Joy, a dataset of real, unscripted two-person conversations with each speaker on a separate audio track.

It covers 100,000 hours of audio from 50,000 speakers in 10 languages, and includes overlapping speech, interruptions, pauses and backchannels such as "mm-hmm," according to the release as reported by Santiago. Shobhit Banga introduces the dataset.

Rohan Paul says most current voice AI learns from LLM-written dialogue read aloud by text-to-speech, so it never hears people talk over each other. A chart in his post reports about 350 overlaps and 300 backchannels per average hour of Joy. Elvis Saravia calls it the first dataset built around conversational timing rather than decades-old telephone audio. These figures and comparisons are claims made in those posts and have not been independently verified.