The development and validation of advanced algorithms requires the pre-processing of multiple sources of information (contained in audio, video and images captured). Consortium members have developed a very large database of pre-processed meeting recordings. The actions, words and all data (slides, white board drawings and hand written notes) associated with a set of scripted meetings were captured using highly instrumented meeting rooms [more information, movie (52Mo)].

The AMI Consortium also develops tools for annotating meeting archives and promotes these worldwide as recommended structure for meeting metadata. Our goal is that all future meeting capture systems store their data (and metadata, i.e., information extracted from the meetings) in a common AMI-like file format.

Once the meetings in the AMI meeting database are fully (manually and automatically) annotated, a very significant investment on the part of the Consortium, the scientists develop algorithms and models. The models are trained on the ground truth in the annotated meeting database. Then AMI models are compared for their ability to detect salient human behavior and information in meetings and to automatically create metadata on meetings.

We know that in the future meeting archives produced in businesses will not be as “fine grained” as those the AMI Consortium uses to train the models. In addition to using the highly instrumented meeting rooms, studies also compare the models’ abilities to use microphone arrays and lower quality audio and video inputs as sources for meeting recordings.