Speech emotion recognition using Deep Dropout Autoencoders

This work describes speech emotion recognition in Konkani with Deep Dropout Autoencoder using Multilayer Perceptron trained through backpropagation algorithm (DDA). To learn robust representation and to reduce the chance of co-adaption, hidden units along with their connections are randomly dropped out in Dropout Autoencoders while input layer remain untouched at training time. Dropout Autoencoders are pre-trained to bring the initial weights of the network to some good solution and thereafter can be stacked to form a DDA that then converted to a Deep Classifier by adding a classification layer. A final fine-tune training was applied to the whole classifier. Several configurations have been tested to find a good classifier to predict seven emotion states. To validate the experiment DDA has been compared with other state-of-art systems like Deep Autoencoder using Multilayer Perceptron trained through backpropagation algorithm (DA), Hidden Markov Model (HMM) to evaluate the improvement. It has been found that the overall recognition accuracy of DDA gives better performance than DA and HMM which are 82%, 80% and DDA gives a performance 87% that have been studied by using four fold leave-one-out cross validation.