Emotion Recognition Through Speech Using Deep Learning

The research aims to develop a voice emotion identification system that utilises empirical feature optimisation techniques and all accessible acoustic data. The methodology used in this work involves gathering data from several emotive speech datasets and utilising audio visualisation techniques to interpret the data. Next, apply augmentation methods to the dataset, such as pitch shifting, temporal stretching, and noise injection. Key audio properties such as Mel Frequency Cepstral Coefficients, Zero Crossing Rate, RMS value, and MelSpectrogram are involved in speech emotion identification. Pitch, tone, and energy shifts corresponding to various emotions are all captured in exquisite detail. A mixture of feature selection and applied deep learning model hyperparameter tuning-specifically, Convolutional Neural Networks-is used to achieve optimisation. The model's accuracy in identifying emotions from speech is evaluated using pre-processed training data and unseen testing data. In addition to improving customer service and mental health monitoring, the results of this project will help to make human-computer interaction more responsive and intuitive. The innovation in the research is ensured by combining the use of Convolutional Neural Networks as the hyperparameter tuning tool for deep learning models with an optimisation process feature filter. It will be ensured by this study that the model learns to recognise emotions from speech data in an efficient manner. The proposed methodology exhibits higher accuracy in identifying emotions from speech when compared to the latest related study. Assessing the model's performance against pre-processed training and unseen testing data, the results are extremely noteworthy.