Turn-taking models are trained on audio where the two speakers are already mixed, and then asked to work out who was talking. We recorded 180 support conversations the other way round: one speaker per channel, overlaps left in. 5.1 hours of it are free on Hugging Face.
These are properties of the recordings, measured per file and published with the set. No claim here is about how a model trained on it performs.
Every conversation is a true stereo pair, each speaker recorded on their own channel rather than split apart afterwards. Per-channel voice activity is ground truth, so turn boundaries do not have to be diarised first. Which channel holds which role varies by file, so the measured speech onset for both channels ships with every row.
Barge-ins, back-channels and interruptions are kept as recorded. Nothing is gated, ducked or edited to tidy the turn boundaries, because those boundaries are the signal.
Resampled to a single rate across the whole release, and loudness-normalised to EBU R128 so files from different sessions can be mixed in one training batch.
Customer-support role-plays across healthcare, billing, airline, insurance and IT support, performed by four recurring speaker pairs (eight performers). Agent and customer roles swap between files.
Tell us who you are and what you are building, and the link comes back straight away — no call, no approval queue. We ask because knowing who trains on this is how we choose what to record next.
Every speaker in this set was engaged and paid by us, and consented in writing to the recording being published and trained on. Nothing here is scraped, and no end customer is recorded: these are performed role-plays, so there is no personal data to redact.
How we handle consentThis set is native US English. We record the same duplex format in other locales to order, including Indic languages, with the channel separation and consent terms unchanged. Say which one on the form and we will tell you what exists and what it would take.
Commission a collection