Joy releases a 100,000-hour dual-track real conversation dataset covering 10 languages
Shobhit Banga releases Joy, a dual-track conversation dataset containing 100,000 hours of real, unscripted two-person conversations, with each person on a separate audio track, covering 50,000 speakers and 10 languages. Rohan Paul points out that current speech AI is mainly trained on fake conversations written by LLMs and read aloud by TTS, which cannot capture people talking over each other, and he is curious about how these conversations are sustained for 10 to 20 minutes.
Why read this
AI score 58
Speech models have long relied on synthetic conversation training; this dataset fills the gap in scenarios such as overlapping speech with real two-person conversations.
Topics
View original postDrop 1/15: Introducing Joy, a dual-channel conversational dataset. Real, unscripted conversations between two people, each on their own track. Built to teach voice models how people actually talk: when to listen, when to jump in, when to say “hmm”. 100,000 hours. 50,000 speakers. 10 languages.
View original
Drop 1/15: Introducing Joy, a dual-channel conversational dataset. Real, unscripted conversations between two people, each on their own track. Built to teach voice models how people actually talk: when to listen, when to jump in, when to say “hmm”. 100,000 hours. 50,000 speakers. 10 languages.
Body
Unscripted calls are hard to collect at scale. Voice AI today mostly learns from fake dialogue written by LLMs and read aloud by TTS, so it never hears people talking over each other;
Really curious how they kept the conversations going for 10 to 20 minutes.
They released Joy, a 100,000 hours of unscripted, two-person conversation across 10 languages
https://t.co/KFRib2aMI6