This dataset contains scripted mock sales conversations performed by former or existing retail, customer service, and sales professionals. The conversations are role-played from scripts, with each one lasting approximately 30 minutes.

The total dataset is around 900 hours of audio. Each conversation is perfectly channel separated, with each person recorded on a separate track. The stereo files provided here combine both speakers into a single file.

Accurate transcripts are available for every conversation, with speaker diarization and timestamps. These transcripts were created directly from the scripts used during recording.

All recordings were captured on voice memos of mobile devices, with formats including m4a, mp4, and mp3 (with some .wav files).

The speakers have mostly western accents, primarily general American, with some British and Scottish accents as well as some Mexican accents.

Dataset Structure

Each sample includes the following fields

audio - the stereo mp3 recording of the full conversation
transcript - the full speaker-diarized transcript of the conversation
speaker 1 gender - the gender of speaker 1
speaker 2 gender - the gender of speaker 2
speaker 1 race - the race of speaker 1
speaker 2 race - the race of speaker 2
speaker 1 accent - the accent of speaker 1
speaker 2 accent - the accent of speaker 2

Downloads last month: 19

Total file size:

22.1 MB