PCA summarization for audio song identification using Gaussian Mixture models

Abstract

In an audio fingerprinting system, the song identification task should be performed within a few seconds. To address the need for fast and robust song identification system, we design fingerprints based on Gaussian Mixture Modeling (GMM) of delta Mel-frequency cepstrum coefficients (ΔMFCC) or delta chroma features (Δchroma). In order to summarize the extracted features over time, a novel implementation of Principal Component Analysis (PCA) is introduced. Experimental evaluations performed on a database of 10000 songs confirm that the proposed PCA summarization technique provides a significant increase in speed in the system's query time. Furthermore, the fingerprints prove to be quite robust against various common distortions, while by using non-distorted test song segments of 10 seconds, the system achieves high identification rates.