DESIGN AND EVALUATION OF FPGA-ACCELERATED NEURAL INFERENCE FOR MUSIC EMOTION RECOGNITION

Music emotion recognition (MER) has gained increasing attention due to its applications in affective computing, music recommendation, and intelligent multimedia systems. This thesis investigates machine learning approaches for MER and evaluates FPGA-accelerated neural network inference in embedded environments. Audio features, including Mel-frequency cepstral coefficients (MFCCs), chroma features, spectral contrast, tempo, and pitch information, were extracted from music samples and used to train machine learning models for both emotion classification and valence-arousal predictions. A Multi-Layer Perceptron (MLP) and a Random Forest were evaluated. Experimental results showed that the MLP achieved the highest classification accuracy of approximately 90%, while the Random Forest demonstrated competitive regression performance.

To evaluate hardware acceleration, a complete MER pipeline was implemented on a Raspberry Pi, and the trained MLP model was implemented on an FPGA. FPGA inference achieved a deterministic execution time of approximately 0.21 ms, compared to 3-5 ms for software inference on the Raspberry Pi, providing up to 25x reduction in inference latency while maintaining comparable prediction quality.

Although the evaluation platform required additional communication due to the transfer of feature extraction between the Raspberry Pi and FPGA, the results demonstrate the potential benefits of FPGA-accelerated machine learning for embedded MER systems. The findings suggest that future integrated hardware-software architectures that move additional processing stages onto the FPGA could further reduce end-to-end latency and improve system performance.