Learning Statistical Regularities of a Simple Musical 'Genre': The Case of Radio Station Jingles

TY - GEN

T1 - Learning Statistical Regularities of a Simple Musical 'Genre': The Case of Radio Station Jingles

AU - Müllensiefen, Daniel

AU - Egermann, Hauke

AU - Burrows, Sean

PY - 2015

Y1 - 2015

N2 - Radio Station Jingles are the fruit flies of systematic music research. They are simple enough to be modelled fully in terms of the features of their musical structure and yet they are complete musical objects that exist in their own right, serve a cognitive function for listeners and are of commercial value to radio stations. This paper aims to identify the statistical regularities of a corpus of radio station jingles by describing the boundaries of this `musical genre' with a probabilistic model that is based on automatically extracted melodic features (Mullensiefen Halpern, 2014). Subsequently we assess the perceptual validity of the probabilistic model with data from two listening experiments that test whether ordinary radio listeners internalise the frequency distributions of melodic features of the radio jingles genre. A corpus of 92 radio station jingles was compiled. Continuous melodic features with a Gaussian distribution (such as mean pitch interval size) were modelled using kernel density estimation; count features (such as jingle length) were modelled with a negative binomial model; and for categorical features (such as tonality) proba- bility estimates were derived from relative class frequencies. Feature probabilities were then used as predictor variables to model the listener response data from two listening experiments showing that ordinary radio listeners are indeed able to distin- guish between jingles that adhere to the stylistic boundaries and those that do not. Backwards model selection was used to arrive at a model that makes use of only 3 features reflecting major v. minor tonality, melodic contour and pitch interval size that are sufficient to describe listener behaviour.

AB - Radio Station Jingles are the fruit flies of systematic music research. They are simple enough to be modelled fully in terms of the features of their musical structure and yet they are complete musical objects that exist in their own right, serve a cognitive function for listeners and are of commercial value to radio stations. This paper aims to identify the statistical regularities of a corpus of radio station jingles by describing the boundaries of this `musical genre' with a probabilistic model that is based on automatically extracted melodic features (Mullensiefen Halpern, 2014). Subsequently we assess the perceptual validity of the probabilistic model with data from two listening experiments that test whether ordinary radio listeners internalise the frequency distributions of melodic features of the radio jingles genre. A corpus of 92 radio station jingles was compiled. Continuous melodic features with a Gaussian distribution (such as mean pitch interval size) were modelled using kernel density estimation; count features (such as jingle length) were modelled with a negative binomial model; and for categorical features (such as tonality) proba- bility estimates were derived from relative class frequencies. Feature probabilities were then used as predictor variables to model the listener response data from two listening experiments showing that ordinary radio listeners are indeed able to distin- guish between jingles that adhere to the stylistic boundaries and those that do not. Backwards model selection was used to arrive at a model that makes use of only 3 features reflecting major v. minor tonality, melodic contour and pitch interval size that are sufficient to describe listener behaviour.

KW - MELODIC FEATURE ANALYSIS,MUSICAL CORPUS ANALYSIS,RADIO STATION JINGLES,STATISTICAL PERCEPTUAL LEARN- ING

M3 - Conference contribution

SP - 1

BT - Proceedings of the European Conference on Data Analysis

A2 - Lausen, Berthold

PB - Gesellschaft für Klassifikation

CY - Colchester

ER -