Joy

Joy is a dual-channel conversational speech dataset announced October 29, 2026 by Shobhit Banga (co-founder of VoiceArena). It targets training voice models on how people actually talk—when to listen, barge-in, or backchannel—using 100,000 hours of unscripted two-person audio with separate tracks per speaker, 50,000 speakers, and 10 languages.

Launch: @shobhitbanga (Drop 1/15 thread)

Key points

1. Named dataset: introduced simply as Joy (not a generic “VoiceArena corpus”).
2. Scale claim: 100k hours vs prior natural dual-channel sets often <2k hours; speaker-balanced (no speaker >0.02% of hours).
3. Content: real overlaps, pauses, backchannels; sample clips in English and Hindi in the launch thread.
4. Roadmap in-thread: v1 Oct 26 / announced Oct 29 at 100k; v2 Mar 27 → 1M; v3 Mar 28 → 10M hours.
5. Downstream teaser: PersonaPlex fine-tuned on a fraction of Joy for more natural turn-taking.

Why it matters

High-traction proper noun for a voice-AI training resource this window; searchers will look up “Joy dataset” / VoiceArena Joy specifically.

How to follow / obtain

1. Read the launch thread and follow-ups for samples and distribution charts.
2. Watch VoiceArena / @shobhitbanga for download, license, or HF mirrors (public HF path was not confirmed in this scan).
3. Evaluate on turn-taking / backchannel benchmarks rather than clean ASR-only sets.

Treat access as announcement-stage until a stable dataset URL or card is published.