Joy dataset: 100,000 hours of unscripted two-person calls to teach voice models real timing
Scobleizer · x · 2026-10-30
Voicearena released Joy, a dual-channel conversational dataset: 100,000 hours of unscripted two-person calls with each speaker on a separate track, spanning 10 languages and 50,000 speakers. It marks overlaps, backchannels and pauses as they happen—about 350 overlaps per hour. The previous ceiling for such natural dual-channel data was under 2,000 hours. Robert Scoble notes most voice AI still learns to talk from LLM scripts read by TTS, and asks whether timing is what's holding duplex voice agents back.
More from Multimodal
- Armin Ronacher on Qwen3-TTS: custom game narrator voices, cloning limits and workarounds — mitsuhiko · 2026-10-30
- Gemini launches shareable personal avatars for images and videos with friends — GeminiApp · 2026-10-30
- EasyComfee: open-source Higgsfield-style creative studio running free on your own GPU via ComfyUI — sintetizacom · 2026-10-30
- Phonic-1 tops Artificial Analysis speech-to-speech index, first non-big-lab model to do so — graceisford · 2026-10-30
- Claude built the whole thing: 1500+ style wildcards node for Krea 2 in ComfyUI — Dear-Spend-2865 · 2026-10-30
- Stacking 10+ LoRAs on Krea 2 Raw FP8: chasing max photorealism in ComfyUI — DragonfruitNo9822 · 2026-10-30