Statistical Learning of Harmonic Movement

Predictive information processing in music cognition. A critical review

“…N-gram and multiple-viewpoint methods have proven powerful for melodic prediction (Conklin and Witten, 1995;Pearce andWiggins, 2004, 2006), key region prediction (Rohrmeier, 2007b), melodic and vertical voice-leading patterns (Conklin and Anagnostopoulou, 2001;Conklin, 2002Conklin, , 2010 as well as in initial results for harmonic prediction (Ponsford et al, 1999;Whorley et al, 2010). As cognitive model, IDyOM has further been evaluated with a series of experimental results.…”

“…N-gram models have been successfully applied for modelling the prediction of melody (Pearce and Wiggins, 2006) as well as harmony (Ponsford et al, 1999;Whorley et al, 2010). N-gram models entail a generalisation that hand-crafted models such as pitch profiles (Krumhansl, 1990), for instance, are revealed to be a special case of unigram models (with a zero-length context, i.e.…”

AI Methods in Algorithmic Composition: A Comprehensive Survey

“…For example, Ponsford et al (1999) used a corpus of sarabande pieces (relatively simple dance music) to generate new compositions using Markov models 20 , but with a pre-processing stage to automatically annotate the compositions of the corpus with symbols to make explicit their structure, and a post-processing stage using a template to constrain the structure of the synthesized composition, in order to generate minimally acceptable results. Another way is the hybridization of Markov chains with other methods.…”

Principles of structure building in music, language and animal song

“…This is demonstrated by a study of harmonic structure using a corpus of seventeenth century dance music [73]: an n-gram model taking into account three-beat metrical structure (and representing each beat by one symbol) seems to favour 4-grams. This is probably because the first beat of a bar is more musically salient than the other two, in harmonic terms, and a 4-gram is able to directly capture at least some of this importance, in this representation, where a 3-gram necessarily cannot.…”

“…[55,56]). In particular, general n-th order Markov models have been proposed for melody and harmony [39][40][41][42][73][74][75] and employed for segmentation and boundary entropy detection [76]. In practical, ecological contexts, it is a common finding that n-gram models with large values of the context length n result in suboptimal models owing to sparsity issues or overfitting.…”