Does the Ear of the Eigenface Hear the Eigensound?
[Case Study] Application of Principal Component Analysis (PCA) for audio feature extraction and music classification.
In this project, we explore the classification of audio files across diverse genres, ranging from classical solo piano (Chopin, Mozart, and Liszt) to pop music (Adele, Billboard Top 30) and jazz. Given the immense feature space of raw audio data (220,500 features per 5-second sample), we utilize ‘eigensounds’—the principal components of our data—for effective preprocessing and dimensionality reduction. With a focus on optimizing model training for environments with low-throughput networks, we determine the lowest audio quality and decomposition levels required to maintain fast and accurate classification using models like SVM, Naive Bayes, and k-NN.
You can access the full paper here (https://sadatian.io/projects/Eigensounds.pdf).
Click here to view the embedded file.