Emotional Speech Generation: An Approach Using Convolutional Neural Networks (CNN) Based Generative Adversarial Network

The objective of emotional speech generation is to create synthetic speech that convincingly conveys specific emotions, enhancing the emotional quality of human-computer interactions. However, existing techniques often fall short of capturing the subtle emotional nuances, leading to speech that feels inauthentic. Additionally, many models lack the robustness needed to perform well across various emotional contexts, which limits their adaptability. Some methods may also generate overly exaggerated or artificial emotional responses, diminishing their effectiveness in real-world scenarios. This research explores using Generative Adversarial Networks (GAN) combined with Convolutional Neural Networks (CNN) for emotional speech generation. The process begins with audio preprocessing using Mel spectrograms for noise reduction and min-max normalization. A CNN-based GAN is then applied for feature extraction. The combination of CNN and GAN is used to classify emotions such as fear, anger, sadness, and happiness from the extracted features. The performance of the proposed method was evaluated using two datasets: RAVDESS and IEMOCAP. Results show that this approach can effectively detect speech emotions, achieving average accuracies of 99% on both datasets.