Open Yap 1K Releases 1,000 Hours of Full-Duplex Conversational Speech

What shipped

TheAgenticDataCompany released Open Yap 1K on September 3, 2026 — 1,000 hours of two-speaker English conversation across 1,602 calls, recorded at 48 kHz full bandwidth. The Hugging Face repo holds a CC-BY-4.0 sample (8.9 hours, 16 conversations); the full corpus is free for commercial and research use under the Open Yap 1K Data Use Agreement.

What changed

Open Yap 1K inverts the conversational-speech recipe. Fisher and Switchboard paired strangers on assigned topics over 4 kHz phone lines; newer corpora pair strangers at full bandwidth. Open Yap 1K has speakers invite someone they already know and talk freely — friends and family, real rooms, no assigned topics. The result is what full-duplex models actually have to learn: overlap, backchanneling, laughter, and short gaps between turns. Audio ships as separate per-speaker tracks on one timeline, with transcripts and metadata (turns/minute, turn-taking gaps, speech dominance) in WebDataset shards loadable via load_dataset.

Why a builder cares

Full-duplex voice agents fail on exactly the behaviors Open Yap 1K captures — interruption, backchannel, quick turn-taking — and open corpora have mostly dodged them with clean, stranger-paired studio audio. 1,000 hours of freely usable natural conversation, with a sample you can test under CC-BY-4.0 today, closes that gap.