Using neural networks to detect and categorize sounds

Artificial neural networks with accurate noise detection can help people with hearing loss be aware of important sounds. Neural networks have already been used for medical purposes and noise detection. However, they typically require large amounts of training data, and background noise can severely decrease the accuracy of the audio detection. The purpose of this project was to examine whether a feed forward neural network (FFNN), recurrent neural network, or convolutional neural network is most effective at audio classification. All three neural networks were trained using the same data of bell sounds, knocking sounds, guitar sounds, and talking. We hypothesized that the convolutional neural network would be the most accurate because it is structured to use more data when it makes predictions and that the FFNN would be the quickest because it requires the least amount of calculations to make predictions. The accuracy of the neural networks was tested with new randomly selected audio of the four categories with no background noise, white noise, environment noise, and busy background noise. Results were compared with the accuracy and times of human participants listening to and categorizing the same sounds. The convolutional neural network was the overall most accurate of the three neural networks, but the feed forward neural network was more accurate when there was little background noise. The recurrent neural network was the least accurate. The feed forward neural network was the fastest among the neural networks and the participants.