Why this matters Natural, unprompted conversations between acquaintances produce short gaps, overlaps, backchannels and interruptions that scripted or stranger-paired corpora rarely capture. Those interaction patterns are precisely what full‑duplex dialogue and voice‑activity models must learn to decide when to start, stop, or yield in real time.
What Sets It Apart
- Channel-separated long-form conversations: both sides are recorded as independent tracks on a single shared timeline, preserving who started when and allowing overlap-aware modeling rather than post-hoc diarisation.
- Scale and accessibility: the full corpus is 1,000 hours across ~1,602 conversations (requestable under the Open Yap 1K Data Use Agreement); the Hugging Face repo provides an 8.9‑hour sample (16 conversations) under CC‑BY‑4.0 so you can inspect schema and audio before requesting the full release.
- Rich per-speaker metadata and time-aligned transcripts: word-level ASR output (Deepgram Nova-3) plus device, headphone/echo state, loudness, effective bandwidth and DNSMOS metrics to support robust training and domain analysis.
- Real-world capture tradeoffs preserved: recordings come from a phone‑call‑like app where people invited existing contacts, so background noise, variable bandwidth (some Bluetooth tracks with limited HF content), and natural conversational artifacts remain intentionally included to improve model robustness.
Who it's for and trade-offs
Great fit if you need channel-separated conversational speech for ASR, overlap-aware turn-taking models, speech separation, conversational TTS, or research into dialogue timing and interruption dynamics. The full corpus is explicitly licensed for commercial and research use under a data‑use agreement.
Look elsewhere if you require perfectly balanced, randomized speaker sampling or fully human‑verified transcripts out of the box: the Hugging Face sample is hand‑picked and the transcripts are machine-generated (expect errors around overlaps and names). Also note some tracks have limited effective bandwidth and audio is delivered un‑normalised, so preprocessing is necessary for some pipelines.
Quick facts
- Sample on Hugging Face: 8.9 hours, 16 conversations, 8 speakers, CC‑BY‑4.0
- Full corpus: 1,000 hours, ~1,602 conversations, ~239 speakers, requestable under the Open Yap 1K Data Use Agreement
- Audio: 48 kHz, 16‑bit per speaker (FLAC/PCM), one file per speaker; per‑word timestamps and extensive per‑speaker metadata
- ASR: machine transcripts generated with Deepgram Nova-3 (word-level)
Where it fits
Use this dataset when you want to train or evaluate systems that must handle realistic turn-taking, overlaps, and backchannels (e.g., full‑duplex dialogue managers, overlap‑aware ASR, speaker‑aware TTS). For tasks that tolerate mixed single‑channel audio only, consider corpora that provide diarised single-channel recordings or human‑verified transcripts.
