Parameter-Based Categorization for Musical Instrument Retrieval

In the continuing goal of codifying the classification of musical sounds and extracting rules for data mining, we present the following methodology of categorization, based on numerical parameters. The motivation for this paper is based upon the fallibility of Hornbostel and Sachs generic classification scheme, used in Music Information Retrieval for instruments. In eliminating the redundancy and discrepancies of Hornbostel and Sachs’ classification of musical sounds we present a procedure that draws categorization from numerical attributes, describing both time domain and spectrum of sound. Rather than using classification based directly on Hornbostel and Sachs scheme, we rely on the empirical data describing the log attack, sustainability and harmonicity. We propose a categorization system based upon the empirical musical parameters and then incorporating the resultant structure for classification rules. 1 Instrument Classification Information retrieval of musical instruments and their sounds has invoked a need to constructive cataloguing conventions with specialized vocabularies and other encoding schemes. For example the Library of Congress subject headings [1] and the German Schlagwortnormdatei Decimal Classification both use the Dewey classification system [3,11] In 1914 Hornbostel-Sachs devised a classification system, based on the Dewey decimal classification which essentially classified all instruments into strings, wind and percussion. Later it went further and broke instruments into four categories: 1.1 Idiophones, where sound is produced by vibration of the body of the instrument 2.2 Membranophones, where sound produced by the vibration of a membrane 3.3 Chordophones, where sound is produced by the vibration of strings 4.4 Aerophones, where sound is produced by vibrating air. For purposes of music information retrieval, the Hornbostel-Sachs cataloguing convention is problematic, since it contains exceptions, i.e. instruments that could fall into a few categories. This convention is based on what element vibrates to produce sound (air, string, membrane, or elastic solid body), and playing method, shape, relationship of parts of the instrument and so on. Since M. Kryszkiewicz et al. (Eds.): RSEISP 2007, LNAI 4585, pp. 784–792, 2007. c © Springer-Verlag Berlin Heidelberg 2007 Parameter-Based Categorization for Musical Instrument Retrieval 785 this classification follows a humanistic conventions, it makes it incompatible for a knowledge discovery discourse. For example, a piano emits sound when the hammer strikes strings. For many musicians, especially playing jazz, the piano is considered percussive, yet its the string that emits the sound vibrations, so it is classifies as a chordophone, according to Sachs and Hornbostel scheme. Also, the tamborine comprises a membrane and bells making it both an membranophone and an idiophone. Considering this, our paper presents a basis for an empirical music instrument classification system conducive for music information retrieval, specifically for automatic indexing of music instruments. 2 A Three-Level Empirical Tree We focus on three properties of sound waves that can be calculated for any sound and can differentiate. They are: log-attack, harmonicity and sustainability. The first two properties are part of the set of descriptors for audio content description provided in the MPEG-7 standard and have aided us in musical instrument timbre description, audio signature and sound description [16]. The third one is based on observations of sound envelopes for singular sound of various instruments and for various playing method, i.e. articulation. 2.1 LogAttackTime (LAT) The motivation for using the MPEG-7 temporal descriptor, LogAttackTime (LAT ), is because segments containing short LAT periods cut generic percussive (and also sounds of plucked or hammered string) and harmonic (sustained) signals into two separate groups [6,7]. The attack of a sound is the first part of a sound, before a real note develops where the LAT is the logarithm of the time duration between the point where the signal starts to the point it reaches its stable part.[12] The range of the LAT is defined as log10( 1 samplingrate ) and is determined by the length of the signal. Struck instruments, such a most percussive instruments have a short LAT whereas blown or vibrated instruments contain LATs of a longer duration. LAT = log10(T 1− T 0), (1) where T 0 is the time the signal starts; and T 1 is reaches its sustained part (harmonic space) or maximum part (percussive space). 2.2 AudioHarmonicityType (HRM) The motivation for using the MPEG-7 descriptor, AudioHarmonicityType is that it describes the degree of harmonicity of an audio signal.[7] Most ”percussive” instruments contain a latent indefinite pitch that confuses and causes exceptions to parameters set forth in Hornbostel-Sachs. Furthermore, some percussive instruments such as a cuica or guido contain a weak LogAttackTime and therefore fall 786 R. Lewis and A. Wieczorkowska