This paper proposes a deep learning-based Long Short-Term Memory (LSTM) method to recognize the speech emotions taken from the audio files in a two-stage approach using Mel Frequency Cepstral Coefficients (MFCC) and feature extraction models. Initially audio signals are pre-processed, MFCC features are extracted from the datasets and then the proposed novel method of feature extraction technique is used to build a network which trains the attributes in recognizing emotions such as angry, happy, sad, disgust, fear, and neutral. To achieve better outcomes than other deep learning algorithms, analysis of training and testing datasets are carried out using the Python Colab platform. Using the CREMA-D and SAVEE datasets, accuracy, true positive rates of speech emotions, and spectrograms of human emotions are recognized.

SER, LSTM, MFCC, CREMA-D, SAVEE.