Speech Emotion Recognition: Music Generation using Convolutional Neural Network Approach

Music has long been an inseparable companion to humanity, enriching our lives and experiences. Life would feel dull without music, as it plays a vital role in expressing emotions. Considering that music has a powerful effect on human emotions, people frequently use it as a form of therapy to manage and control their anxiety. There is limited study on the recognition of emotional speech among people and the subsequent generation of music based on their present emotional state. Further investigation is required to explore methods for generating music that aligns with the user’s present emotional state. This research applies the Convolutional Neural Network (CNN) to analyze the Mel-frequency cepstral coefficients (MFCC) spectrogram. The MFCC spectrogram is derived from the speech emotion recognition (SER) signal and utilized to generate music by identifying emotional states. A sad or happy emotion is produced by generating music that can improve the person’s mood based on the mood captured. The experiment shows that the CNN classifier model successfully achieved an accuracy of 80.61% during the training phase. Aside from that, utilizing Adam and Stochastic Gradient Descent (SGD) optimizers, the study compares their efficacy in enhancing model accuracy and reliability. The findings reveal that specific configurations of dropout rates and epochs, combined with the choice of optimizer, significantly impact the model’s effectiveness in emotion recognition.