A Comparative Analysis of SVM, CNN and LSTM Models for Speech Emotion Recognition

Abstract

The Speech Emotion Recognition (SER) project aims to develop an intelligent system capable of recognizing human emotions from speech signals. SER plays a major role in applications such as Human-Computer Interaction (HCI), sentiment analysis and psychological research. In this project, we leverage machine learning techniques and signal processing methods to analyze speech signals and extract features that capture the emotional content, following a structured pipeline that includes data collection, preprocessing, feature extraction, model training and validation. To reduce high frequency noise and retain essential speech characteristics, a low-pass filter is applied and then Mel-Frequency Cepstral Coefficients (MFCCs) is applied to extract meaningful features from audio files, and employing machine learning models like Support Vector Machines (SVM), as well as deep learning models such as Convolutional Neural Networks (CNNs) and Long Short-Term Memory (LSTM), facilitates emotion classification. The system's objective is to accurately distinguish between different emotions like anger, sadness, happiness, and neutral states from speech signals. Moreover, the inclusion of a user-friendly interface enhances accessibility and usability, enabling seamless interaction with the system. Through experimentation and rigorous evaluation, the efficacy of the proposed approach in recognizing emotions from speech is demonstrated. The SER project holds immense potential to contribute to various domains, including HCI, mental health assessment, and affective computing, thereby augmenting our comprehension and interaction with human emotions.