Sample bundles

Hear the data before you license it.

Each bundle is 10 to 30 minutes of a dataset, cut the way the full set is delivered. Choose one, and we email you a download link that works for 24 hours.

What is in a bundle

A README with the dataset card excerpt for that dataset.
The audio, in the same layout and format as the full dataset.
Transcripts and metadata in the same files and fields you would receive with the full set.
LICENSE-SAMPLE.txt: the bundle is for evaluation only.

How to read it

Numbers that identify a person appear as [pii] in the transcripts and as a tone in the audio. Other tags in transcripts are in square brackets, like [filler] and [overlap].

Delivery layouts for the full sets are on formats; licences are quoted per use, as described on licensing.

Whole conversations

Complete recordings with speaker-labelled transcripts.

Full-duplex speech to speech

Each speaker on a separate channel, timing preserved.

Hindi and Tamil full-duplex call-centre conversations

Speech recognition

Audio with time-aligned transcripts in native script.

Diarisation

Single-channel recordings with speaker turns marked.