Bangla Speech Emotion Recognition Based on Audio Features Using CNN and LSTM

Research on Bangla Speech Emotion Recognition (SER) has been notably limited due to data scarcity.This study utilizes advanced machine learning techniques with the SUBESCO dataset to enhance Bangla SER capabilities.We methodically processed this dataset to extract key features such as Mel-frequency cepstral coefficients (MFCC), zero crossing rate, root mean square energy, etc., owing to the MFCCs' effectiveness in succinctly capturing speech dynamics.A hybrid model combining Convolutional Neural Networks (CNN) with Bidirectional Long Short-Term Memory (LSTM) and dense layers was developed, achieving an accuracy of 88.42% and a loss of 0.3546 on the SUBESCO dataset.The integration of CNN and LSTM enhanced the ability to interpret complex emotional expressions, marking a significant advancement in Bangla SER.These findings not only propel SER forward but also set a foundation for improved human-technology interactions for other underrepresented languages.