Exploring Audio Emotion Classification: A Comparative Analysis of Machine Learning and Deep Learning Models Using the TESS Dataset

The speech recognition is a vital aspect of human-machine interaction, where it is used in virtual assistants as well as voice-controlled systems. This study explores how three different methods of extracting audio features Mel Frequency Cepstral Coefficients (MFCC), Mel Spectrogram, and Chroma features can be integrated to improve the work of a speech recognition system based on machine learning. The paper discusses the contribution of individual features of MFCC, Mel Spectrogram and Chroma in learning the spectral and pitch related features of speech signals. These attributes serve to feed into the different machine learning models such as k-nearest neighbors (KNN), support vectors machines (SVM), Random Forest, multi-layer perceptron (MLP), and Naive Bayes (NB). Comparative analyses are made to identify the performance of each feature set and their combinations on the overall effectiveness and strength of speech recognition. Also, the paper further expands its area of concern to include emotion recognition as part of speech cues. The categories of emotions that are used are; angry, disgusted, fear, happy, neutral, surprise and sad. The most successful machine learning model is taken to identify emotions, and Flask, a web framework, is used to provide a convenient interface of real-time emotion prediction.