In recent times, machine learning models have been widely used for emotion recognition from speech. With a large number of users creating and sharing a vast amount of data, including audio on different social media platforms, Speech Emotion Recognition is being explored for purposes like filtering out hate speech, assessing customer satisfaction, and so on. It has been observed that the use of predefined features has not performed very well for classifying various speech emotions. Therefore, it is the need of the hour to develop some automated methods for extracting the features and to build a labeled dataset without human intervention. Hence, in this paper, an autoencoder-based feature extraction method has been proposed that can extract relevant features from an unlabeled dataset. These are then fed into a Convolutional Neural Network (CNN) network to categorize eight different emotions. The obtained results yield an accuracy of 71.45% on the RAVDESS dataset with normal features and a CNN network. Also, use of the proposed autoencoder network increases this accuracy to 75.8%. Another highlight of this paper is to consider the classification of a multitude of emotions totaling eight. Further, it has been demonstrated that the use of an autoencoder yields better values of metrics like the f1 score for individual emotions and prevents the overfitting of a machine learning model.