Recognition of Voice Commands using CNN for Romanian Language

In this paper, we studied the capability of the DLCNN (Deep Learning Convolutional Neural Network) classifier to be included into an intelligent human-machine interface. This interface will interact with a small number of people (generally one person, the owner of the device: autonomous car, smart home, intelligent wheelchairs, etc.). The main goal was to use a classifier capable of being trained quickly (maximum 10 training epochs), with a minimal set of training data. We need an algorithm of reasonable complexity, that can be integrated on small single-board computers (SBCs) like Raspberry Pi for example. We used 2 sets of voice commands, one in English and one in Romanian, which will train two different DL-CNNs. MFSC (mel-frequency spectral coefficients) coefficients did not provide the expected results, but with MFCC (mel-frequency cepstral coefficients) coefficients we obtained a 92% accuracy rate in the voice commands recognition.