Speech Emotion Recognition using AdaMax and Weighted Adam Optimizers

Machine learning and speech emotion detection are two technologies that are being used to create more natural and engaging ways for humans to interact with computers. These technologies are still in their early stages of development, but they have the potential to revolutionize the way we interact with technology. This work proposes the use of weighted Adam (AdamW) and AdaMax (Adamax) Optimizers to improve results on the current schemes. Weighted Adam is an extension of Adam that assigns different learning rates to different parameters. It can help to improve convergence and avoid overfitting. Adamax is a variant of Adam that uses a more robust learning rate schedule. It can be more stable than Adam, especially in high-dimensional problems. We combine three well known speech emotion datasets - RAVDESS, SAVEE and TESS. SVM Classifiers and Convolutional Neural Networks were used for the prediction process and a wide range of features were extracted from the wav file in the preprocessing phase. We achieved a testing accuracy of 84.79%, 82.73% and 88.53% with Adam, AdamW and AdaMax respectively.