Year: 2026
Type: article
Abstract: The classification of music genres is a fundamental task in music information field. This project investigated how the characteristics of audio, the model architecture, and the selection of input segment length jointly affect the classification accuracy on the FMA small dataset. Lightweight self-designed convolution network CompactResNet trains on 3 second fragments by using these three major feature representations: Mel spectrogram, stft spectrogram, mfcc. Then it is compared with svm baselines with rbf kernels in those same features. This paper has done more research about the segments of length between 3 and 30: The CNN using MFE gets the best result, which is 90.5%; and the SVM with STFT gets 89.7%, so these two kinds of models like features differently: The shorter segment was better than the longer one: Accuracy decreased monotonically from 90.5% at three seconds and 53.8% at 30s, so data enrichment via segmentation is better than adding more time context.
Author: Yuxing Jiang
Language: English
Topic: Music and Audio Processing
Subfield: Signal Processing
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7161775356