Automatic harmonization history
Among musical compositional systems there has been a large number of researchs carried out in the field of automatic harmonization from the early 1950s onwards. When looking at automatic harmonization history the most pioneering work in automatic harmonization is that of Rothgeb who developed a SNOBOL, program to solve the harmonization problem and to identify the voice leadings notes to accompany it. Rothgeb used rule base approach that has a set of ‘if statements’ according ot the musical domain. His main aim was not focusing on automatic harmonzation but to test the computational soundness of two bass harmonization theories from the eighteenth century (Rothgeb, 1969). Afterwards Steals in 1979 proposed a system to use constraints to create passing chords as chords that could be inserted between two given chords. These passing chords must satisfy some musical constraints, such as interval relations between the roots of the first, passing, and last chords. Further, Steals used essentially a frame system, augmented with a bread-first search (Steals, 1979).
The above evidence suggests that the first works for automatic harmonization were carried by using rule base approaches by representing set of musical rules. Furthermore, this indicates that no constraint satisfaction algorithm was used and the main concern was about mastering the combinatorial explosion, by putting more knowledge in the solver.
Automatic Harmonization – AI Techniques
There have been several approaches taken for automatic harmonization in artificial neural networks (ANN). There has been a large number of research carried out in the field of music composition using neural networks available in the literature; Todd (1989, 1991) used a feed-forward ANN with feedback for melody generation, Lewis (1991), Hild et al. (1992) used a neural network to harmonize chorales that divides harmonization in to three parts 1) harmonic skeleton 2) chord skeleton and 3) ornamentation, Mozer (1991, 1994) generated melody using ANN, Stevens and Wiles (1993), Bellgard and Tsang (1994) constructed a Boltzmann machine for harmonization which generates harmonies non-deterministically. Further, Toviainen (1995) trained neural networks for jazz improvisation and later Hörnel and Degenhardt (1997) generated harmonization for four part chorals in baroque style.
The most of the automatic harmonization used neural networks are for four part choral harmonization. Automatic harmonization using neural networks, Schwanauer the developer of MUSE that harmonize chorals, claims that there are five learning techniques learning by rote, learning from failure, learning from examples, learning by analogy and learning from discovery (Schwanauer,1993).
Even though neural network approaches are widely used among musical composition systems Papadopoulos and Wiggins discusses that there are many disadvantages in neural network approach for musical compositions, in his analysis he claims that “The representation of time cannot be dealt efficiently even with ANNs that have feedback” (Papadopoulos & Wiggins, 2007). Further, he states that neural networks cannot reproduce fully trained data set. Even if they are fully trained it is not generalized (Papadopoulos & Wiggins, 2007). Toviainen (1999) claims that neural networks fail to pick up the higher-level features of music such as phrasing or tonal functions. Further, he explains that a neural network often solves toy problems, with many simplifications when compared to knowledge based approaches (Toiviainen, 1999). However, the review reveals that the neural network approaches used for musical compositions suggests that neural networks were used extensively in the past years for musical applications and they have been relatively successful especially in domains such as perception and cognition.
Automatic harmonization – Knowledge Based Systems
The most pioneer technique used for music composition is using knowledge base concepts. Unlike neural network that learns from examples, knowledge based systems work according to the rules. There can be three types of knowledge based systems 1) rule based expert systems, Constraint logic programming and case based reasoning. There has been several approaches on automatic harmonization using rule based approach. Ebcioglu implemented a choral harmonizing system which is a rule based expert system using Backtracking Specification Language (Ebcioglu, 1988). Tsang and Aitken (1991) and Pachet and Roy (1998) used constraint logic programming (CLP) and constraint satisfaction techniques (CSP) respectively for four part choral harmonization. However, Sabater et al also developed four part automatic harmonization for vocals using rules and cased based reasoning but his approach failed due to constructing the harmonization purely on rules. Reason for his failure was because rules don’t make music but music makes rules. The advantage of using a knowledge based approach is because each newly correctly harmonized piece can be memorized and made available as a new example to harmonize other melodies which are also known as learning by experience process (Mantaras & Arcos, 2002). When analyzing the above existing work the main advantages of rule based systems are, their ability to be stand-alone programs, and their ability to explain their choice of action; furthermore, knowledge based systems can introduce explicit structures or rules. Even though knowledge based systems can be a good candidate for musical composition Papadopoulos and Wiggins claims that following disadvantages can be seen in knowledge based systems. 1) Knowledge elicitation is difficult and time consuming, 2) gaps between expert and programmer because representation is not flexible and also knowledge based systems can be too complicated with more exceptions to the rules.
Automatic harmonization – Mathematical Model
Mathematical model also known as Stochastic processes is a machine learning technique which is widely used in many applications and research fields such as; data mining, speech recondition, hand writing recondition and computer vision. One of the most popular mathematical models is Hidden Markov Model; it has been used successfully in genre classification, instrument identification and key estimation in the field of music systems (Levitt, 1993). There are few approaches taken for automatic harmonization using Hidden Markov Model. Allan and William designed a harmonization model for Bach chorales using Hidden Markov Models. In this model, the visible states represent melody notes and the hidden states are chords. This model has to be provided sequences of chords as input, restricting its applicability in more general settings (Allan & Williams, 2005). Hanlon and Ledlie have developed CPU Bach an automatic choral harmonization system which harmonizes melody lines for the style of composer Bach. This system breaks the harmonization process in to two parts 1) generation of a harmonic progression and 2) realization of harmonic progression into four part melody lines. Hanlon and Ledlie use HMM to model the chord progression and derive the realization of vocal lines using constraint satisfaction method. That has been successful for a large number of melodies (Hanlon & Ledlie, 2002). Papadopoulos and Wiggins claims the main draw back of this models are one must needs to find the probablities by analyzing many data therefore, lot of data is needed and also it is difficult to capture higher or more abstract levels of music (Papadopoulos & Wiggins, 2007). However, HMM has been sucessfully used in many applications as disscussed above. Furthermore, HMM is more robust and flexible compared to other models, and also its complexity is low.
Automatic harmonization – Genetic Algorithms
There have been several researches done in automatic composition using Genetic Algorithms (GA). Horner and Ayres have successfully developed a system that generates four-part harmony using genetic algorithms. However, the chords needed to be given and the system produces the four melody lines according to the chords (Horner, 1995). Phon-Amnuaisuk and Wiggins created another harmonizing system using GA in their system, soprano information is input by the users and the GA generates the other three voices with musical domain knowledge encoded in the fitness function. The limitations of using GA are that these systems are subjective and there is no way to simulate human behavior; furthermore, the user must hear all the potential solutions in order to evaluate a population. However, Biles claims that GA has efficient search method which is good for large scale searches and it has the ability to provide multiple solutions. (Bills, 2007).
Musical composition has been the focus of computer science since the 50s, and there exists several ways of representing music using computers. There have been several AI applications for automatic music compositions from early ages. When looking at AI applications, they can be categorized according to their most prominent features; namely, 1) Neural Networks (systems which learn), 2) Knowledge Based systems, 3) Evolutionary methods, and 4) Mathematical Models. Each of these models has their own pros and cons which were discussed in the above sections.