This study presents the design and implementation of a low-cost, automatic voice-based emotion recognition system using a Raspberry Pi 4. The system was developed to classify three basic emotions—happiness, sadness, and anger—from natural, unscripted speech samples recorded in real-world university environments. A significantly expanded and gender-balanced dataset of 259 audio recordings was used, enhancing generalization and ecological validity compared to previous work limited to 51 female-only samples. The preprocessing pipeline included amplitude normalization, conversion to a unified.wav format, and spectral analysis via Fast Fourier Transform (FFT), focusing on the first 1000 Hz of the frequency spectrum. The algorithm, written in Octave, was optimized for batch processing and integrated with the Telegram API, allowing users to send voice messages and receive emotion classifications automatically. Evaluation on 305 audio files yielded an average accuracy of 53.6%, with anger being the most accurately identified emotion (68.0%). The system achieved a 5.3% improvement in accuracy and a 40% reduction in processing time compared to the baseline. Despite its limitations in noisy environments and in differentiating similar emotional tones (e.g., happiness vs. anger), the system demonstrates practical viability for embedded deployment and remote interaction. This work highlights the importance of using diverse and realistic datasets, as well as the potential of embedded systems in democratizing access to affective computing technologies for educational, therapeutic, and assistive applications.