Genre classification of music by tonal harmony

We present a genre classification framework for audio music based on a symbolic classification system. Audio signals are transformed to a symbolic representation of harmony using a chord transcription algorithm, by computing Harmonic Pitch Class Profiles. Then, language models built from a groundtruth of chord progressions for each genre are used to perform classification. We show that chord progressions are a suitable feature to represent musical genre, as they capture the harmonic rules relevant in each musical period or style. 1. Genre classification Organization of large music repositories is a tedious and time-intensive task for which music genre is an important meta-data. Automatic genre and style classification have become popular topics in Music Information Retrieval (MIR) research because musical genres are categorical labels created by humans to characterize pieces of music and this nature provides the genre meta-data with a high semantic and cultural information to the music items in the collection. Traditionally, the research domain of genre classification has been divided into the audio and symbolic music analysis and retrieval domains. Nevertheless, some authors have paid attention recently on making use of the best of both worlds. The work by Lidy et al. (Lidy et al., 2007) deals with audio to MIDI transcription in order to extract features from both signals and then combine the decisions of the different classifiers. On the other hand, Cataltepe and coworkers’ approach (Cataltepe et al., 2007) is just the opposite: to synthesize audio from MIDI and then analyze both signals to integrate the classifications. Our proposal is to use tonal harmonic information to distinguish between musical genres. The underlying hypothesis is that each musical genre makes use of different rules that allow or forbid specific chord progressions. As it can be found in (Piston, 1987), some rules that were almost forbidden in a period have been accepted afterwards. Also, it is well known that pop-rock tunes mainly follow the classical tonicsubdominant-dominant chord sequence, whereas jazz harmony books propose different series of chord progressions as a standard. The goal of this work is to classify digital audio music using a language modelling system trained from a groundtruth of chord progressions, bridging the gap between audio and symbolic by means of a chord transcription algorithm. 2. Chord transcription The Pitch Class Profile (PCP) measure has been used in automatic chord recognition or key extraction since its introduction by Fujishima (Fujishima, 1999). The perception of musical pitch has two main attributes: height and chroma. Pitch height moves vertically in octaves telling which octave a note belongs to, while chroma tells its position in relation to others within an octave. A chromagram or a pitch class profile is a 12dimensional vector representation of a chroma, which represents the relative intensity in each of twelve semitones in a chromatic scale. Since a chord is composed of a set of tones, and its label is only determined by the position of those tones in a chroma, regardless of their heights, chromagram seems to be an ideal feature to represent a musical chord. Genre Classification of Music by Tonal Harmony Table 1. Classification accuracy percentages using different n-gram lengths. Data set 2-grams 3-grams 4-grams 3 classes 54.4 61.1 61.1 popular vs. jazz 70.9 83.4 83.4 academic vs. jazz 75.0 75.0 75.0 academic vs. popular 56.7 66.7 66.7 In this paper we have obtained the Harmonic Pitch Class Profile (HCPC) by applying the algorithm in (Gómez & Herrera, 2004), which deviates from Fujishima’s PCP measure by distributing spectral peak contributions to several adjacent HCPC bins and taking a peaks harmonics into account. The feature vectors are then processed to obtain a symbolic representation of chords in the form of triads. Thus, each song is represented as a string of chord progressions, with an alphabet of 24 symbols (major and minor triads).