Previous works have proven the relevance of using the Deep Learning approach for acoustic source localization, and in particular the efficiency of the approach based on the BeamLearning deep neural network architecture that we have developed. However, the constitution of a dataset remains one of the cornerstones of the supervised learning paradigm. If the constitution of a simulated database is possible, only simple antenna geometries can be used. Indeed, the 3D diffraction phenomena by the antenna body are generally difficult to take into account by simulation. To fully capture the pressure field interactions with the antenna, it is however possible to create a database from real measurements, which can be a time consuming task and requires a priori a large number of different environments. To overcome these drawbacks, the laboratory has a 3D spatializer, which we use to create various data sets for training networks dedicated to source localization tasks. This 5th order ambisonic restitution system can be controlled in real time and offers a perfectly controlled and fully automated synthesis of a large number of source positions. After careful calibration of the spa- tialization system, a large database is built to prove that these data are relevant to train the BeamLearning architecture, and to localize a 3D acoustic source with an angular error surpassing the 3D localization performances of the SH-MUSIC algorithm, while offering a much higher angular estimation rate. We will also show that a training performed on data coming from measurements of a real antenna allows to include in the procedure an implicit calibration of the microphones composing the antenna, which greatly simplifies the problems of calibration of this type of sensors.