In recent years, touchless technologies for human-computer interaction have been widely developed. Doppler sonar makes it possible to extract information from hand gestures by emitting/receiving ultrasounds, and gestures recognition is generally achieved using features extracted from a gesture sequence as input to Convolutional Neural Network. This work aims at achieving an accurate and rich acoustical touchless gesture recognition with a low number of transducers and a low complexity real-time classifier. For this purpose, we use a thin plate as an acoustic antenna, excited by a few piezoelectric actuators, and capture the echoes with microphones around the plate. High amplitude emissions on a large bandwidth are achievable with a better-integrated system. Signal features selected to contain meaningful information on rich 3D gestures are computed and used as an input to a small Gated-Recurrent-Unit neural network. We achieve the detection and classification of 11 3D gestures with an accuracy of 93.5% with our system.