Mittuniversitetet

This project is based upon two previous projects handed to the author by the Norwegian University of Science and Technology in co-operation with Disruptive Technologies.

The report discusses sound sensing and Neural Networks, and their application in IoT. The goal was to determine what type of Neural Networks or classification methods was most suited for audio classification. This was done by applying various classification methods and Neural Networks on a data set consisting of 8732 sound samples. These methods where logistic regression, Feed-Forward Neural Network, Convolutional Neural Network, Gated Recurrent Unit, and Long Short-term Memory network. To compare the Neural Networks the accuracy of the training data set and the validation data set were evaluated. Out of these methods the feed-forward network yielded the highest validation accuracy and is the preferable classification method. However, with more work and refinement the Long Short-term memory may prove to be the better solution.

Future work with a Vesper V1010 piezoelectric microphone and IoT implementation is discussed, as well as the social and ethical difficulties proposed by what is essentially a data gathering system.