Emotion Recognition from Voice Using Multi-Layer Perceptrons

Abstract

Emotion recognition from vocal expressions has become a pivotal task in affective computing, enabling more natural and empathetic human–machine interactions. This manuscript proposes a multi-layer perceptron (MLP)-based framework for classifying discrete emotional states from speech signals. We extract Mel-frequency cepstral coefficients (MFCCs), spectral flux, zero-crossing rate, and chroma features from a balanced corpus of acted and elicited emotional speech. After normalizing features and conducting principal component analysis (PCA) for dimensionality reduction, we train an MLP with two hidden layers of 128 and 64 neurons, respectively, using rectified linear unit (ReLU) activations and dropout regularization. Training is performed with an 80:20 train–test split, employing the Adam optimizer with learning rate scheduling.

The model achieves an overall accuracy of 87.4% on the test set, with balanced precision and recall across five emotions: anger, happiness, sadness, fear, and neutrality. A statistical analysis (ANOVA and pairwise t-tests) confirms that the MLP significantly outperforms a baseline support vector machine (SVM) classifier (p < 0.01). Simulation research explores the network’s sensitivity to hyperparameters and noise levels, demonstrating robustness to up to 20 dB of additive white Gaussian noise. These findings support the feasibility of lightweight MLP architectures for real-time emotion recognition in resource-constrained applications.