[WEEK 7] Results

We had trained 3 classifiers(SVM, CNN, KNN) with randomly sampled dataset we had gathered. Even tough KNN is a crude machine learning algorithm we had seen that it shows good results in practice. Unfortunately not in our case. KNN results in this experiment was always under 20 percent accuracy in the training which led to us not testing with test dataset. SVM model have given around between 30–35 percent accuracy. It had maximum 38 percent accuracy. CNN fared better but with much difficulty. Training
CNN was a bothersome, time-consuming processes. Initially we had trained 4 convolutional layers 96 filters for each layers, kernel sizes (1,1) for performance, reasons, 2 dense layers which all layers activation function relu except the last layer which is softmax, with 5 epochs. Training accuracy was around 34 percent and testing was 29 percent. In order to increase accuracy we increased epochs sizes however the training was excruciating slow and only increasing epochs would not do. Hence we drop one convolutional layer and other optimizers and instead made kernel sizes (5,5) and epochs 100. It had reached 80 percent accuracy after 40 epochs. However it started over-fitting. Testing accuracy between 60 and 70 epochs started to decrease which meant over-fitting. We have tested for every 10 epochs therefore the exact point can not be provided. This model took more than hour to train 10 epochs. It can be seen that for every feature we had to train it and test it. In
the end we had skipped some test such as doing it for every 10 epochs or so. Our result might be crude in that regard. The features that have been mentioned are used one by one for every model. In the tests Spectral features have given poor results. Generally lower 10 percent if we make a ratio between other corresponding models and feature accuracies. The best feature to use was MFCC which was the feature used when we hit the 80 percent accuracy result on the training data.