In recent years, audio sentiment analysis has emerged as one of the most essential tools in customer service, psychological health treatment, and broader human-computer interaction domains. This paper presents a comprehensive methodology for implementing a robust Audio Sentiment Analysis system using the Toronto Emotional Speech Set (TESS) dataset—a widely recognised benchmark that exhaustively covers seven distinct emotional states, making it an ideal foundation for this research. The primary objective is to achieve precise detection and classification of emotions from audio recordings by leveraging cutting-edge machine learning techniques and advanced signal processing methods. The research systematically addresses data preprocessing, including noise reduction, normalisation, and feature extraction. Key acoustic features—Mel-Frequency Cepstral Coefficients (MFCCs), chroma features, spectral contrast, and zero-crossing rate—are extracted to capture subtle variances in emotional expression. These features serve as inputs to four distinct machine learning and deep learning models: a one-dimensional Convolutional Neural Network (CNN-1D), a Long Short-Term Memory (LSTM) network, a Gated Recurrent Unit (GRU) network, and a Support Vector Machine (SVM). Experimental results demonstrate that the CNN-1D model, which automatically learns spatial hierarchies of features, significantly outperforms traditional machine learning approaches, achieving a classification accuracy of 99.58% across all seven emotion classes (anger, disgust, fear, happiness, sadness, surprise, and neutral). The SVM model also performed exceptionally well with 96.00% accuracy, while the GRU and LSTM models achieved 87.00% and 77.86% respectively. These findings validate the efficacy of deep learning architectures for audio-based emotion recognition and establish a strong benchmark for future research. The conclusions extend beyond academic contribution, highlighting practical applications in automated customer service systems, sentiment analysis in social media, and real-time emotional monitoring in therapeutic settings. Future work will focus on integrating multimodal data—combining audio with text and visual cues—to enhance the accuracy and reliability of sentiment analysis systems.
Umair Aziz (Tue,) studied this question.