Speech recognition based on spectrograms by using deep learning

Abstract

Speech Recognition is widely being used and it has become part of our day to day. Several massive and popular applications have taken its use to another level. Most of the existing systems use machine learning techniques such as artificial neural networks or fuzzy logic, whereas others may just be based in a comparative analysis of the sound signals with a large lookup tables that contain possible realizations of voice commands. These models base their speech recognition algorithms on the analysis or comparison of the analog acoustic signal itself. The sound has particular characteristics that can not be seen through the representation of its propagation wave in time. This project proposes speech recognition through an innovative model that analyzes the graphic representation of the acustic signal, its spectrogram. Therefore the model does not classify the speech through its acoustic signal but its graphical representation. This leads the research to an approximation of the problem through the use of image classification techniques. Image clasification was considered a task only the humans can do, with the devoloping of machine learning techniques this perception has drastically changed. This project covers several techniques and shows the potential of Deep Learning for objects classification and within this field presents the convolutional neural networks as the most suitable algorithim for the classifcation of spectrograms. As a method to clearly illustrate the efficacy of the proposed model, the used alorithim was trained with two self-obtained datasets. Several experiments were conducted to make a detailed comparison of the system throughput and its levels of accuracy.