Purpose: This study presents a multimodal emotion recognition framework that integrates facial and vocal cues through deep learning. By addressing the complexities of human-computer interaction (HCI), the research develops an end-to-end desktop application for real-time affective state classification. Methodology: The system employs two specialized Convolutional Neural Networks (CNNs). The visual pipeline utilizes 32×32 normalized facial images, while the acoustic pipeline processes a feature fusion of Mel-frequency cepstral coefficients (MFCCs), chroma, and mel-spectrograms. A unified Tkinter-based GUI facilitates the entire lifecycle from dataset preprocessing and model training to performance visualization through automated accuracy/loss plotting. The architecture utilizes a two-stage convolution-pooling backbone with a 256-unit dense layer, optimized for seven facial and eight vocal emotion classes. Findings: Empirical results demonstrate high classification reliability, with both models achieving over 96% training accuracy. The modular design ensures efficient persistence of weights and architectures, allowing for low-latency inference in mental health monitoring and adaptive learning environments. Originality: The integration of dual-modality CNNs into a lightweight, accessible desktop environment bridges the gap between high-complexity research models and practical, user-centric diagnostic tools.

Multimodal Emotion Recognition, CNN, MFCC, Tkinter GUI, Affective Computing, Human-Computer Interaction.