The NIST RT evaluations gave researchers an opportunity to build competing systems in tasks such as speaker diarisation, speech-to-text, and more. The meeting dataset used for speaker diarisation in RT07 contains 8 meetings recorded in 4 different rooms with a total of 35 speakers. More details can be found on their website.

We have improved this reference RTTM dataset by manually re-segmenting. It is now accurate to within 0.1 seconds and has speech segments with speaker labels for the complete audio files. We have made it available to download on our website, and would appreciate feedback.