Audio fingerprinting system

Abstract

Over the past few years, the online multimedia collection has grown at a fast past pace. Several companies showed interest to study the different ways to organise the amount of audio information without the need of human intervention to generate metadata. In the past few years, many applications have emerged on the market which are capable of identifying a piece of music in a short time. Different audio effects and degradation make it much harder to identify the unknown piece. In this dissertation, an audio fingerprinting system which identifies an unknown piece of audio is presented. It is compared with the previously computed fingerprints stored in the database. A fingerprint must be compact, discriminant and invariant to distorted audio clips. A set of 155 audio files is inputted to the system and preprocessing techniques are performed on the audio files. The feature extraction methods employed are the Mel Spectrum Coefficients and the MPEG-7 basic descriptors. The Mel Spectrum Coefficients are extracted from the Mel Frequency Cepstrum Coefficients (MFCC) process. The MPEG-7 basic descriptors are composed of four descriptors, Audio Spectrum Envelope (ASE), Audio Spectrum Centroid (ASC), Audio Spectrum Spread (ASS) and Audio Spectrum Flatness (ASF). Wavelets were also implemented as a comparison to the other feature extraction methods. The features are then modelled using a binning technique and Gaussian Mixture Models (GMMs). The unknown feature set is identified by taking the minimum Arithmetic Distance between the tested audio piece and each audio file registered. The results show that the best method overall is a combination of MFCCs, ASE and ASF as feature extraction methods and these are modelled using the binning technique.