Collect real two-speaker interaction with separate channels, natural overlap, interruptions, backchannels, and timing metadata for speech-to-speech systems and low-latency voice agents.
Human speakers acknowledge, interrupt, hesitate, restart, and speak at the same time. A model trained only on isolated utterances or neatly alternating dialogue does not learn when to listen, yield, continue, or recover.
Separate synchronized channels preserve those events without forcing diarization to guess who spoke. The annotation scope can stay lightweight or extend into precise turn-taking labels depending on the model job.
Use the free worksheet to define coverage, rights, delivery fields, held-out rules, and acceptance tests before requesting a collection plan.
Build a dataset specificationRequest representative separate-channel conversation samples. The team sends a manually selected slice based on your target language and dialogue setup.
Enter your work email and verify it with a 6-digit code. An admin then reviews the request and sends a relevant sample manually within two business days. Use the optional field to name the language, channel, or condition you need to inspect.
Send the language, channel, volume, annotation, and timing you know. The quote form opens directly in project-brief mode.
