This page is part of the SALIERI webpages.

[Home] · [Mail]

AVA - experiments

by Wolfgang Chico Töpfer

To give an idea of AVA s capacities I carried out a series of experiments which resulted in an evaluation of 48 compositions generated by AVA and summarized by three tables presented here. The experiments are based on different settings of switches that control AVA s search behaviour when composing ie. searching schemes to be connected and instanced into a playable SALIERI segment. Schemes are organized in a schemes base with a phrasal structure that keeps the original phrase context surrounding a scheme. Thus, a scheme taken from the first variation of KV 265 is surrounded by other schemes taken from the same variation. Such scheme sequences form the base phrases that constitute a schemes base. It is searched according to the following user-set switches:

  1. the functional dependence switch (FCT) with switch values
    1. s : the SPEAC rules [Chi97, Cop96] are followed
    2. e : the musical functions of the theme and their sequence are exactly followed; a selected scheme must have the same function as the theme at the given point
    3. f : only the SPEAC categories of the musical functions of the theme are followed
    4. 0: musically speaking, a type of functional theme independence; only the most similar function of each scheme is selected with respect to the theme. If no similarity is given, an arbitrary function of the scheme in question is chosen. This allows the application of schemes with seeming functional incompatibility in respect of the theme.
  2. the phrasal dependence switch (PHR) with binary switch values: iff on, schemes are selected from as few base phrases as possible (phraith search strategy); otherwise (unphraith search strategy) base phrases are left as soon as possible, thereby ensuring a maximum usage of all the base phrases in the schemes base
  3. the phrasal discrimination switch (DIS) with binary switch values: iff on, only schemes from base phrases of equal type are selectable with respect to the theme phrase to be variated; e.g. if the theme phrase is an A phrase only schemes from A base phrases are allowed.

The generated variations are evaluated through the intuitive listening process with respect to the following three categories:

  1. structural quality (SQ): the most important evaluation category consisting of criteria that judge the musically meaningful placement of schemes and the choice of adequate schemes; one criterion is the correct ending of a phrase by a cadence, for instance. Another is functional correctness which depends on the sequence of functions and its musical logic; e.g. a sequence V-IV or v-IV (as it ocurred in the experiments exp39814 and exp39815) is musically illogical. Furthermore, single scheme pitches are valued according to their fitting the function in use. Correct voice-leading is only checked by the listening process (by myself) which means that some - more difficult to hear - voice-leading faults may have passed unnoticed. Discovering such faults would require a precise visual (ie. non-aural) or algorithmic post examination. However, this would go beyond the original aim of giving a first impression. Later experiments should be based on large scheme bases and on precise and comprehensive examination.
  2. recombination degree (COM): this evaluation aspect identifies the proportion of segments that stem from a single base phrase to those which are combinations made of schemes from several base phrases; a lower recombination degree indicates fewer combinations.
  3. recognizability (COG): this property characterizes how much of a variation is actually recognizable as a derivation of a theme. This means recognizing the melodic or/and harmonic course of the theme

According to these categories each experiment is given marks on a scale from 0 (very low) to 4 points (very high) .

The experimental frame is defined by a small schemes base and the use of a single scheme search method that is based on the event model (thus it is called event search), a specific music structure model that views music as a series of events. Despite AVA s greater capacities the variation type is restricted to melody variation, a limitation which follows quite naturally from the applied schemes base being exclusively made of schemes taken from melody variations. This makes comparing variations with each other easier.

The themes on which the variations are modelled are taken from the theme-and-variations KV 573, 265 and 398. In order to use a regular, less improvisational form that fits AVA s underlying phrase model a fragment of the theme of KV 398 is taken instead of the original.

Experiments on the theme of KV 573

This series contains two variations - exp5736 and exp57310 - that received the highest SQ. Both are identical and follow the theme functions exactly in spite of exp5736 being the only experiment with FCT switched to e . This is caused by always choosing the most similar function of each scheme (see explanation on the FCT switch) with respect to the theme so that 0 may result in e . In addition, exp5736 and exp57310 have the same PHR and DIS switch values. Therefore, both experiments yield identical variations.

Each of the other variations with the same PHR and DIS switch values (but different functional dependence) is marked one point below the winner exp5736/exp57310 due to different functional incompatibilities: functionally speaking, exp5731 is only a weaker version of the winner without any additional, SQ-relevant structures. In contrast, such ones are offered by exp57314; exp57315 is a registerwise improved version of exp57314 that uses a number of new schemes. As far as the scheme transitions are concerned, exp57316 is less coherent than the other functionally independent variations.

Setting the DIS switch to 0 may cause an exhaustion of the schemes base, particularly when it is relatively small. This occurs in the B phrase in exp5733: It remains empty because all those schemes that were once free for B are now already used by A . Therefore no adequate schemes are left; exp5733 is thus only given one point in the COG category. On the other hand, the A phrase offers a much higher recombination that justifies three points in the COM category.

As can be seen from the COG column in table 1, all the variations are average as to their recognizability: The reduced theme melody can only be heard very indirectly. Moreover, the way it has been refined makes the harmonic course thoroughly difficult to recognize (even contradicting it, see above).

The recombination degree goes from average to low as long as phrase dependence is set (compare the columns COM and PHR). This is partly improved by functionally independent variations.

Table 1: AVA s variations of the theme of KV 573, their switch values and evaluation
s 1 1 2 2 2 exp5732
s 0 1 2 3 3 exp5731
s 0 0 1 3 2 exp5733
s 1 0 2 2 2 exp5734
e 1 1 2 2 3 exp5738
e 0 1 2 3 4 exp5736
e 0 0 2 3 3 exp5737
e 1 0 2 2 3 exp5735
f 1 1 2 2 3 exp5739
f 0 1 2 3 4 exp57310
f 0 0 2 3 3 exp57311
f 1 0 2 2 3 exp57312
0 1 1 2 3 2 exp57313
0 0 1 2 3 3 exp57314
0 0 0 2 3 3 exp57315
0 1 0 2 2 2 exp57316

Experiments on the theme of KV 265

This theme proves to be difficult for AVA : Three quarters of the variations range from an average to a low degree of recombination, almost half of them possess a low one (see COG column in table 2). This makes the variations highly recognizable since the original one by Mozart is highly recognizable too. As a rule in most of the variations, functional inconsistencies are a result of applying new schemes. Moreover, the majority of the variations start with a rhythmic irregularity.

Most of the variations with high SQ have earned this mark thanks to the strong similarity to Mozarts own variation. An exception is exp26514 which has - despite functional problems - musically meaningful scheme placement and partly convincing scheme transitions.

As in the experiments on the theme of KV 573, a switched on PHR leads to average and low recombination degrees. This is particularly valid when the theme functions are followed exactly (see exp2658 in table 2).

Table 2: AVA s variations of the theme of KV 265, their switch values and evaluation
s 1 1 3 1 3 exp2652
s 0 1 3 2 3 exp2651
s 0 0 3 3 2 exp2653
s 1 0 2 4 1 exp2654
e 1 1 3 1 3 exp2658
e 0 1 3 1 2 exp2656
e 0 0 3 1 2 exp2657
e 1 0 3 1 2 exp2655
f 1 1 3 1 3 exp2659
f 0 1 3 1 3 exp26510
f 0 0 3 2 2 exp26511
f 1 0 2 2 2 exp26512
0 1 1 3 2 2 exp26513
0 0 1 3 2 3 exp26514
0 0 0 3 3 2 exp26515
0 1 0 2 3 2 exp26516

Experiments on the theme of KV 398

As can be seen in table 3, the experiments on the theme of KV 398 only yield one winner (exp3982) generated by the application of the SPEAC grammar (s ). Its low recombination degree is typical of all the variations that were generated with PHR and DIS switched on. Therefore, the winner differs only in two ornamental figures from the original variation. These figures are placed very effectively at the beginning of the A phrase.

Even less different from the original are the variations exp3988 and exp3989: only a single scheme that was diminuted to an ornamental figure makes each of these variations distinguishable from each other and the original.

A slight exception is exp39813 because of its functionally independent generation, allowing the application of schemes which would otherwise be filtered out. This variation has a functionally highly incompatible scheme that appears at the beginning of the B phrase. As an aside, switching PHR and DIS on also proves to be helpful for a good harmonic course independently of FCT.

A remarkable quarter of the experiments on the theme of KV 398 have no variation of the B part. As already explained, this follows from the small schemes base and the DIS switched off. Adding these variations to those with a low recombination degree shows that half of all the variations are products of low recombination. A high recognizability comes therefore as no surprise.

A comparison between those variations that were generated by SPEAC rules and those without any functional dependence shows very clearly that SPEAC ensures a better functional structure.

All the remaining variations with a high SQ have convincing passages either (a) among less convincing ones or (b) among those that are identical with Mozarts original. Examples for (a) are exp39814 and exp3981; examples for (b) are exp39812 and exp3985.

Table 3: AVA s variations of the theme fragment of KV 398, their switch values and evaluation
s 1 1 3 1 4 exp3982
s 0 1 2 3 3 exp3981
s 0 0 2 3 2 exp3983
s 1 0 3 1 2 exp3984
e 1 1 3 1 3 exp3988
e 0 1 2 3 2 exp3986
e 0 0 2 3 2 exp3987
e 1 0 3 2 3 exp3985
f 1 1 3 1 3 exp3989
f 0 1 2 3 2 exp39810
f 0 0 3 3 2 exp39811
f 1 0 3 2 3 exp39812
0 1 1 3 2 2 exp39813
0 0 1 2 3 3 exp39814
0 0 0 2 2 2 exp39815
0 1 0 3 2 2 exp39816

Experimental results

The applied search method - the event search - proved to generate musically and structurally sensible variations despite the small schemes base. Half of all the variations were given high SQ marks (high structural quality). Only one variation was juged to be of low SQ. Approximately 41 % of the variations possess a high recombination degree. This indicates a well-built schemes base and emphasizes the significance of the experiments. A high recognizability of approximately 48 % mainly results from 75 % of the experiments on the theme of KV 265 having an average or low recombination degree. Only around 31 % of high recognizability is obtained if they are left out of account. This may be understood as an indication that further criteria are needed to ensure recognizability.

The music structure model behind the search method (the event model) is confirmed in its phrase-structural aspects. For instance, there is no variation that simply interrupts a phrase (bear in mind that KV 398 variations are modelled on a fragmented theme, see above). The power of the event model below the phrase-structural level , however, could not be verified clearly. There were both convincing and unconvincing scheme recombinations.

The supposition that functional independence may lead to new, musically meaningful variations (e.g. exp57314) is confirmed. Certainly this stems from the capacity of the applied schemes to fit more than a singular musical function. Another positive consequence is undeniably the increase of scheme recombination. As expected, phrase dependency leads to its decrease.

Obviously, the SPEAC rules (FCT switch s ) offer a richer scheme choice than all the other FCT possibilities. As assumed, the bass model (ie. SPEAC) is verified as to its capacity in building musically sensible function sequences.

In summary, the event search and its underlying event model proved to be a compositionally sensible means to generate music automatically. It ought to be examined thoroughly in further and larger scaled experiments. The reduction model (see description of AVA ), on which all the other music structure models are based, has generally provided sensible diminutions, including most of the generated ornaments.

[Home] · [Mail]
© sic!systems, designed by hh; last update 1997/12/31