A Comparative Analysis on Speech Emotion Recognition

This is my final year thesis project. In this project, I tried to show a comparative analysis on speech emotion recognition.


Here, We worked on IEMOCAP dataset which was realized by the University of Southern California.  The dataset has three types of data: text, audio, and video.  We worked on the only text and audio dataset.  The abstract part is given below.

Abstract— Recently, speech recognition has been one of the most necessary domains in machine learning and deep learning. People tend to order a machine by speech more comfortably and so this field has emerged to fulfill this necessity. In this paper, we proceed towards a feature engineering-based approach for detecting emotion in speech domain. Handcrafted features from multiple audio files are inclined to feed into learning models Features extracted from text domain are also included for resolving ambiguity. We follow both machine learning and deep learning-based approach where extracted features are fed into six machine learning classifiers namely, Random Forest, Multinomial Naïve Bayes, Support Vector Machine, Logistic Regression and Gradient Boosting whereas deep learning consists of feed-forward neural network and LSTM based classifiers for feature extraction. For both domains, eight hand-crafted features are extracted. Finally, we compare machine learning model to deep learning model using accuracy, precision, recall and F-measure where we observe that shallow machine learning models achieve higher performance than renowned deep learning models for recognizing emotion. 

We have divided our experiment into three parts:
1. Audio only
2. Text only
3. Combined(Audio+Text)

We have used five machine learning and two deep learning models.

Machine Learning
1. Random Forest
2. Xtreme Gradient Boosting
3. Multinomial Naive Bayes 
4. Support Vector Machine 
5. Logistic Regression 

Deep Learning
1. Multi-layer Perceptron
2. Long Short Term Memory (Audio only)


Steps:
1. First, we extract the emotion label from the datasets.
2. audio vectors 
3. extract features from audio wav files
4. pre-process audio and text features dataset. 
5. audio classification models
6. text classification models
7. combined (audio+text) classification models

Github: https://github.com/Imtiajul/A-comparative-analysis-on-speeech-emotion-classification-models