Emotions Extraction from Urdu Speech Using Different Linguistic Features

Abstract

Natural Language Processing (NLP) is the study of Machine & Human interaction and to resolve ambiguity where Natural Language is rich in its structure, form and ambiguity. Natural Language is not just a combination of meaningless words but a way to communicate. Out of many features to understand language, Humans process language majorly by the prosodic features like pitch and linguistic features like understanding the word in a speech. Understand emotion using NLP can benefit a number of industrial applications mainly Healthcare, Robotics and education. This research proposes to develop an Emotion Recognition system for recognition of emotion from speech using linguistic features. The proposed system is composed of two modules, developing an Urdu Corpus that contains the list of words with their Part of Speech (POS) tag and Emotion. The second module is the Emotion Recognition system that operates on earlier developed Urdu Dataset by taking Urdu Speech as input, classify it to recognize the emotion and perform the assessment. The system tracks the emotion of the spoken words in Happy, Sad, Angry, Surprise and fear where a normal tag is used for neutral words. Neutral words are words in our speech which does not contain any emotion or words used for completion of the sentence also known as stop words. Features like word stemming, edit distance, stop word removal and synonym list is used to get better results. The Words that the system is unable to classify are stored separately. These words are processed manually by assigning emotion and POS tags following the same process for dataset development and added to the final dataset. This will help in increasing the size of the overall corpus. There is a number of emotions that human expresses in linguistics feature of speech, according to researchers there is more than 135 human expressing emotion. This research works consists of five emotions that are defined as basic emotion by most of the researchers. The proposed system is driven on supervised Machine learning model for Emotion Recognition that is based on the size of the EPOS Model File. Multiple classifiers like SVM, Naïve Bayes and, Deep learning with CNN is used for validation and comparison of results with existing works.