EC447: Pattern Recognition and Machine Learning Course Project
Vowel-only versus consonant sound classification using EEG data corresponding to speech prompts
The data for this project was taken from Univeristy of Toronto - KaraOne database. The main objective of the project was to recognise imagined speech by finding correlations between EEG and facial features with speech features. Instead of using multi-modal data, we use only EEG data and try to accomplish one of the sub tasks - given a processed EEG feature vector, classify whether the corresponsing speech prompt had a vowel sound or not.
We tried out different methods - Gaussian Mixture Model, Logistic Regression and Neural Networks. We implemented the GMM and Logistic regression from scratch using native Python and tested for stability too. Neural network models were built with Keras with TensorFlow backend.
The results for the different classifiers, different parameter sweeps and corresponding inferences can be found in the project reports. The best result we obtained was for a Dense sequential network with 1 hidden layer of just 5 neurons, no regularization, batch size of 128, SGD optimizer, ReLu non linearities, and 0.01 learning rate - 61% test accuracy.