This is the FreeTalk Corpus of conversational speech (video & audio). It contains the set of raw data files with associated time-aligned transcriptions, and scripts that enable fast browsing in a variety of formats

The material was recorded over three days during three 90-minute sessions in early November 2007. One participant was paid as an informant, the rest were volunteers. All have agreed for the recordings to be made public for scientific use. No constraints were made on the content of the discussions and they arose spontaneously as part of the group social interaction. The participants were originally from Belgium, Finland, Japan, Ausrtalia, and the UK, and all conversations were in English, though the recordings were made in Japan and several references to Japanese topics triggered the use of some words in that language.

A small 360-degree lens attached to a Pointgrey Flea2 camera was used to collect rich data and flat video recordings were added as backup and for more detailed views from different angles. Audio was collected by a central Sennheiser MKH30 P48 nucrophone directly to digital memory via a Maranttz PMD 660.

All speech was transcribed manually and annotations were made of topics, topic changes, main speakers, mood of the conversation, participant attention, etc., A particular feature of this corpus is the graphical display of speech activity and the movement traces output from face and body detections algorithms taking input from the 360-degree camera.

The same material can be viewed interactively in a variety of formats at www.speech-data.jp/taba/nov07 (uid nick, pwd campbell)

A do-it-yourself kit containing software and sample code can be downloaded here