Speech Emotion Recognition (SER) seeks to identify a speaker’s emotional state from voice signals, enabling more natural human–computer interaction. While recent approaches rely heavily on 2D convolutional neural networks (CNNs) with spectrogram-based features, they often overlook other valuable acoustic cues. This study addresses that gap by proposing a one-dimensional CNN model (CNN_SER) which built around multi acoustic feature fusion concept that integrates three complementary features: Mel-Frequency Cepstral Coefficients (MFCCs), Zero Crossing Rate (ZCR), and Root Mean Square Energy (RMSE). To build a robust dataset, we combined and augmented four widely used speech corpora RAVDESS, SAVEE, TESS, and CREMA-D using pitch shifting, time modification, and signal stretching. The model classifies speech into seven emotions: surprise, fear, disgust, anger, sadness, happiness, and neutral. Performance was evaluated using Accuracy, Precision, Recall, and F1-Score. Results show that fusing MFCC, ZCR, and RMSE delivers a significant boost, with CNN_SER achieving a peak accuracy of 94%. These findings highlight the potential of multi acoustic feature fusion, paired with a streamlined 1D CNN for advancing reliable and efficient speech emotion recognition.