Evaluation of Computational Cost and Result Accuracy in Design and Efficient Implementation of Log-Mel Spectrogram and MFCC Feature Extraction Using Fixed-Point Arithmetic on FPGA

Mel-frequency cepstral coefficients (MFCCs) are essential features in audio processing, aligning with human auditory perception and are used in applications such as speech recognition, speaker identification, and audio compression. Despite their benefits, the computational complexity of MFCCs poses challenges for real-time applications and embedded systems. Field-programmable gate arrays (FPGAs) offer a solution with their parallel processing capabilities and efficient handling of complex arithmetic operations. This paper presents an optimized implementation of the MFCC algorithm on an FPGA, focusing on computational cost and resulting accuracy. Our key strategies include using fixed-point arithmetic to reduce computational complexity and resource usage. We compare software and hardware implementations of the fast Fourier transform (FFT) to demonstrate the advantages of FPGA-based acceleration. The system is designed for high customization, allowing adaptation to various application requirements. Experimental results explore both the speed of fixed-point arithmetic and the processing results compared to the librosa library in Python, validating the data accuracy under fixed-point operations. These findings provide a foundation for integrating audio feature extraction in system-on-a-chips (SoCs) for various application scenarios.