A Generic Training and Classification System Formirex08 Classification Tasks: Audio Music Mood, Audio Genre, Audio Artist and Audio Tag

This extended abstract details a submission to the Music Information Retrieval Evaluation eXchange (MIREX) 2008 for the training and classification tasks audio music mood, audio genre, audio artist and audio tag. The same system has been submitted for the various tasks without any adaptations to the specific problems. The system named ircamclassification is a generic system which performs batch feature extraction, models training (using various classifiers) and file indexing (or file segmentation) into classes. The features extracted are generic in order to be applicable to many different audio and music indexing problems. The features are not specific to the above mentioned MIREX08 tasks. The goal of this submission is to test the applicability of a generic classification system to those tasks. 1 SYSTEM DESCRIPTION ircamclassification is an extension of a system initially developed for instrument-samples indexing described in [3] using the features described in [4]. Only the subset of features applicable to polyphonic audio signals (music) has been used here. In [5] the system has been extended for speech/music segmentation. It is this system that has been used forMIREX08 tasks. We briefly review it in the following. 2 FEATURE EXTRACTION In the present submission, only three sets of audio features are extracted from the signal. MFCC: The first set aims at describing the shape of the spectrum at each time. Mel Frequency Cepstral Coefficients (40 Mel bands, 13 coefficients including DC component) are extracted every 20ms using a Blackman window of length 40ms. SFM/ SCM: MFCCs only describes the shape of the spectrum whatever the content of the signal is noise or sinusoidal (harmonic) components. In order to describe this noise/ sinusoidal content, we also compute height Spectral Flatness [2] and Spectral Crest Measure coefficients. This is done using the same analysis parameters. Chroma/ PCP: The third set of features gives rough information about the meaning of the harmonic content of the signal. For this, twelve Chroma [6]/ Pitch Class Profiles (PCP) [1] coefficients are computed using a Blackman window of length 100ms synchronized in time with the two other feature sets. Delta and acceleration coefficients of the above mentioned features are also computed. Finally, a simple temporal modelling (mean and standard deviation) of each feature is performed using a sliding window of length 500ms and a hop size of 250ms.