Open Yap 1K brings 1,000 hours of natural speech to voice AI builders

The Agentic Data Company has released Open Yap 1K, a speech dataset containing 1,000 hours of English conversation recorded in real-world environments. The Hugging Face Blog announcement says the corpus is free for commercial and research use, making it a potentially useful resource for teams building voice agents.

Unlike many established speech datasets, Open Yap 1K records friends and family members speaking naturally through a phone-like app. The two speakers are captured on separate tracks at 48 kHz, preserving interruptions, laughter, backchannels, overlapping speech and short pauses rather than forcing clean, alternating turns.

Why it matters

Full-duplex systems need to listen and respond while another person is still speaking. Training primarily on polite, stranger-to-stranger conversations can leave models poorly prepared for real calls, where people interrupt, react mid-sentence and talk over one another. Open Yap 1K intentionally retains those behaviors, along with moderate background noise from kitchens, homes and other everyday spaces.

The corpus contains 1,602 conversations, averaging 37.5 minutes, with a median length of 30 minutes. At the midpoint, 8.3% of voiced time overlaps and conversations contain 11.9 turns per minute; the dataset also includes examples reaching 20.9% overlap and 20 turns per minute. Friends account for 70.9% of recorded relationships, followed by colleagues at 10.8%, romantic partners at 9.8% and family at 8.5%.

An 8.9-hour sample covering 16 conversations is available on Hugging Face under CC BY 4.0. The full 1,000-hour release is available on request under the Open Yap 1K Data Use Agreement, free for commercial use. For developers, that offers a larger and more behaviorally realistic training option for conversational speech research, evaluation and production voice applications.

Source: Hugging Face Blog