Enhancing audio signal segmentation for musical instruments with refined MFCC-based features

This study investigates the effectiveness of modified Mel-Frequency Cepstral Coefficients (MFCC) features, incorporating Delta and Delta-Delta MFCCs, for classifying musical instruments—specifically drums, guitar, and keyboard. The modified MFCC approach was compared to the standard MFCC in terms of classification accuracy and instrument detection across five cross-validation folds. The results show that the modified MFCC outperforms the standard MFCC, achieving a classification accuracy of 95% compared to 86%, highlighting the importance of temporal features in distinguishing the unique audio patterns of each instrument. Additionally, Fast Fourier Transform (FFT) analysis was employed to identify dominant frequencies, revealing that drums dominate lower frequencies, guitars occupy mid-range frequencies, and keyboards cover a broad spectrum. Clustering of these dominant frequencies further confirmed their distinct spectral characteristics, validating the modified MFCC features as reliable indicators for instrument classification. These findings have significant implications for real-time applications in audio segmentation, music transcription, and AI-based music recognition, with potential use in music production, education, and other related fields.