Section: New Results

The Kinovis Multiple-Speaker Tracking Datasets

The Kinovis multiple speaker tracking (Kinovis-MST) datasets contain live acoustic recordings of multiple moving speakers in a reverberant environment. The data were recorded in the Kinovis multiple-camera laboratory at Inria Grenoble Rhône-Alpes. The room size is $10.2\times 9.9\times 5.6$ meters with T60 = 0.53 seconds. The data were recorded with four microphones embedded into the head of a NAO robot. Because there is a fan located inside the robot head nearby the microphones, there is a fair amount of stationary and spatially correlated microphone noise. The signal-to-noise ratio of the microphone signals is of approximately 2.7 dB. The recordings contain between one and three moving participants that speak naturally, hence the number of active speech sources varies over time. The robot-to-speaker distance ranges between 1.5 and 3.5 meters. Ground-truth trajectories and speech activity information were obtained in the following way. Participants were wearing optical markers placed on their heads such that the Kinovis motion capture system provides accurate 3D trajectories for each participant. Moreover, an infrared marker is placed on the participants' foreheads. This enables the identification of each participant over time. Whenever time a participant is silent, he/she hides his/her infrared marker, thus allowing speaking/silent annotations of the recordings.

Website: https://team.inria.fr/perception/the-kinovis-mst-dataset/.