Nowadays, Speech Emotion Recognition (SER) system was rapidly growing, it becoming a trend in present era and significant inferences in various fields such as human–computer interactions, and customer services. However, the system struggled on recognizing the subtle and nuanced emotions with the individual speakers. To overcome these drawbacks, in this research proposed an Acoustic Feature Extraction (AFE) for SER Initially, the data is collected from the dataset Ryerson Audio-Visual Database of Emotional Speech and Song (RAVDESS) gives various speech data from numerous samples, then the input signals augmented by adding background noise, then the relevant features are extracted by proposed AFE using Library Reliable Open-Source Audio (LibROSA), it extract the signals from different groups, from the extracted features the selection is performed using ReliefF, and finally, the classification is performed using Hidden Markov Model (HMM) classifies the new or unseen speech data into emotion class. The proposed AFE model archives better results, the obtained values are 0.95% of precision, 0.95% of recall, 0.95% of F1 score and 0.93% of accuracy, when compared to the existing methods respectively.