Listen to a sample
The player gives each speaker their own waveform. Switch one off and the conversation between the other two is still there.
- A three-way conversation
Why three people is a different problem
With two people, when one stops, the other is usually next. Add a third and anyone can take the turn, so a system has to work out who is being addressed and whether it should speak at all while two others are talking.
Models trained only on pairs do not see that decision. Separate recordings let you train on it and score it, because you know who spoke at every moment, including during overlap.
Why this matters for humanoids
A humanoid at home or at work will rarely talk to just one person. It will hear people talking to it, near it, and over each other, and it cannot answer everything it hears. Zhu, Miyoshi, and Okafuji list the hard parts for a social robot: overlapping speech, working out the addressee, and reading cues such as gaze. They also name the lack of large datasets of natural multi-party talk as a barrier, and their own set runs to 82 minutes.[1] 1X says NEO’s speech system detects when someone is talking to it and responds only at appropriate moments.[2]
Human conversation will not solve this alone, and the same paper notes that human-only corpora transfer poorly to robots.[1] It is still where a system can learn how three people share the floor and when nobody expects it to speak.
What we measured in the sample
In the 32-second clip above, about 22% of speech time has more than one person talking, and each speaker talks for 34% to 39% of the clip. We estimated this from the signal level on each track in one clip, so it does not describe the whole dataset.
What is included
You get one mono audio file per speaker. Transcripts and turn-taking labels are not part of the set today, and we can add them to your specification.
Data for voice AI
Get the dataset
Tell us what you are building and we will follow up on licensing. If you need larger groups or other languages, we can record to your specification.
References
- He Zhu, Ryo Miyoshi, and Yuki Okafuji. “Whom to Respond To? A Transformer-Based Model for Multi-Party Social Robot Interaction.” 2025. arxiv.org/abs/2507.10960
- The Robot Report. Coverage of the 1X NEO pre-order launch, 2025. www.therobotreport.com/1x-announces-pre-order-launch-neo-humanoid-robot