Back home

Open Yap 1K

huggingface(datasets trending) 0.78
Visit source

Intro

Open Yap 1K is a speech dataset released by The Agentic Data Company and trending on Hugging Face in late October 2026. It contains 1,000 hours of dual-channel (two-track) English conversation recorded at 48 kHz, designed specifically for training full-duplex voice models, meaning models that can listen and talk at the same time like a person on a phone call.

A free 8.92-hour sample (16 conversations, CC-BY-4.0) is on the Hugging Face Hub, and the full corpus is free for commercial use on request under the Open Yap 1K Data Use Agreement.

Explain

Key points

  • Natural, self-paired speakers: instead of pairing strangers with assigned topics (as in Fisher or Switchboard), the team built a WhatsApp-like calling app that people used for their normal calls with friends and family. The result is more interruptions, overlaps, laughter, backchannels and short gaps between turns.
  • Wideband audio: 48 kHz recordings, versus roughly 8 kHz telephone-band audio in older corpora like Fisher.
  • Separate tracks: each speaker is captured on their own channel, which is what full-duplex and turn-taking models need.
  • Real rooms: recorded on people's own devices in real environments, with unusable recordings filtered out by an ML screening pipeline.
  • Metadata: per-conversation fields such as topics, relationship, duration, turns per minute, turn-taking gap and speech dominance.

Why it matters

Full-duplex voice agents trained only on polite, turn-by-turn speech fail when real users interrupt. Open Yap 1K positions itself as an alternative to Fisher, Switchboard and CANDOR for exactly that problem.

Usage

How to use it

  1. Browse and listen to the sample in the Hugging Face dataset viewer.
  2. Load the sample in Python:
from datasets import load_dataset
ds = load_dataset("TheAgenticDataCompany/open-yap-1k", "audio_previews", split="Sample")
print(ds[0]["topics"], ds[0]["turns_per_minute"])
  1. For the full 1,000 hours, request access on The Agentic Data Company's page and accept the Data Use Agreement.

Good fits

  • Training or fine-tuning full-duplex speech-to-speech models and voice agents.
  • Turn-taking, interruption and backchannel prediction research.
  • Evaluating ASR or TTS on natural, overlapping conversation.

Notes

The sample is CC-BY-4.0; the full corpus has its own agreement, so check its terms before redistribution.