Sound recognition with deep learning methods using spectrogram technique

Voice recognition is the process by which the human voice is detected and recognized by the computer through a microphone. It has started to create an important place for itself in the developing technology process. In this study, deep learning methods were used to recognize previously recorded speech commands. In order to be processed in two-dimensional deep learning structures, audio data were converted to image by using spectogram method. Spectogram method is a sub-sampling technique based on fast fourier transform. Five different convolutional neural network models were created by using the spectogram technique and the trainings were performed with different data numbers. The data set is divided into different number of classes as separate categories. The effect of the number of classes on deep learning structures was investigated. Furthermore, the sensitivity of the network with different training data numbers was evaluated. In order to evaluate the performance of the networks, two classifiers were used as Softmax and Support Vector Machine. Over 95% accuracy was achieved.