The FBKDICIT-FrontBack Dataset has been collected to experiment on the capability of discriminating between front and rear speakers when using a linear microphone array

The dataset is partitioned into the training and testing subset, both were recorded at 96 kHz.

The training set involves only 1 speaker, with a total number of 16 recordings that last around 2.15 hours and include 340 front/back segments.

The test set records for the other nine speakers, with a total number of 23 recordings that last for around 3.16 hours and include 415 front/back segments.

The time periods when speakers are in front or behind back the array were manually labelled as the ground truth.

To access the dataset please contact: Brutti Alessio