FI84670C - FOERFARANDE FOER BILDANDE AV DIGITALKODSEKVENS, SAERSKILT EN NOTKODSEKVENS. - Google Patents

Description

1 84670

A method for generating a digital code sequence, in particular a note code sequence

The invention generally relates to a method for generating a digital code sequence from a finite number of different code types, in which method new codes are generated one by one after the code sequence on the basis of the codes already in the sequence. In particular, the method is intended for the automatic composition of computer music, each code representing one or more quantized properties of a particular note,

Musical productions that sound pleasant but do not have the form of an independent work of art are often used as background music for films, plays, and other performances. The quantitative need for such music can be considerable. Also in shops and other public spaces, a lot of calm background music is used to entertain customers and create a suitable atmosphere. One way to produce such background music is to use the so-called an electronic machine for producing synthetic music, comprising one or more electronic musical instruments or synthesizers and a device for automatically generating control signals therefor.

One way to produce these control code sequences and signals would be the so-called artificial intelligence programs that exploit rules produced heuristically based on musical expertise. However, the present invention does not relate to such expert methods, but to an apparatus which automatically, on the basis of the example material, generates the necessary rules and automatically generates a new code sequence by means of these rules.

One such known device for generating control signals is based on Markov processes, in which each note (height, length) is treated as one stochastic ti-35 line in a queue formed by spaces. Based on the given example country, 2 84670 material, sheet material, the probability Pr = PriSj | S1_1, St_ 2, ...) is determined for the state in the queue, when the previous states in the queue are Sj _1, St. 2, etc. In music based on Markov processes, 5 satisfactory often the three preceding states to reach a result. New music is generated by means of probability functions stored in memory, starting with a key sequence to which is added the probability function Pr and, for example, the most probable follower mode based on the last three notes or 10 states in the sequence. The sequence thus grown is used as a new key sequence and thus the process of endlessly generating note code material and control signals for electronic musical instruments or synthesizers. In addition, additional operations or rules are needed to produce typical musical structures from parts of the melody.

This known method of generating note codes requires a great deal of teaching material to generate conditional probability density functions. In addition, the synthetic music produced in the above-mentioned manner is usually unsurprising and monotonous, due to the fact that in the stochastic process each note is treated as equivalent, which does not correspond to the characteristics of natural music.

It is an object of the present invention to provide a method for increasing melodic variation and avoiding or alleviating some of the problems of the prior art.

This is achieved by the method according to claim 1.

The method according to the invention for producing a continuous code sequence uses the principle of a dynamically expanding context previously applied in speech recognition and described in [1] "Dynamically expanding context, with application to the cor-35 retention of symbol strings in the recognition of continuous 3 84670 speech ”, Teuvo Kohonen, Proceedings of the Eigth International Conference on Pattern Recognition, October 27-31, 1986, Paris, France (IEEE Computer Society) pp. 1148-1151. The essential difference from a speech recognition application is that when the last said method is used only to correct erroneous codes, so that new stochastic code sequences are continuously created in the presented method.

As in the Markov process, the present roe-10 method determines the code in a code sequence based on the codes immediately preceding it. However, the present invention uses discrete "grammatical" rules in which the length of the conditional content of the rule expressions, i.e., the number of preceding 15 codes required, is a dynamic variable defined by generating rule expressions from example sequences based on the inconsistencies in these examples. when two or more rule clauses are formed, the same condition part 20 but a different consequence part, i.e. the new code, these rule clauses are marked as invalid and their condition part length is increased until unambiguous, ie valid rule clauses are obtained. 25 rule clauses are retained and formed into a tree structure as described below, the dynamically expanding context method being central to this structure Since the above rules are mechanically generated on the basis of the local interdependencies of the symbols present in the sample material, the formation of the rules does not require, for example, a musical theoretical analysis based on the expertise of the sample music material.

Similarly, when 35 new codes are generated after the code sequence utilizing these rule expressions, the last generated code of the code sequence is then compared to the rule expressions of the lookup table, then the two last generated codes, etc. until a match is found with the condition of the last 5 code sequences. the code indicated in the consequence of this rule clause may be added. Said tree structure allows comparisons to be made systematically. This provides an "optimal" code sequence that "stylistically" tends to follow the rules formed from the example sequences. When the method is applied as such to the production of a sheet code sequence, the music produced follows a well-desired style, but may still produce long copies of parts of the sample material.

Variability and surprising changes in the code sequence can be achieved by using, at least occasionally, random selection based on the fact that when a valid rule expression is found, it is randomly replaced with an associated invalid rule expression whose condition is the same as the condition of the found rule expression.

The code automatically generated by the method according to the invention can be used to control electronic musical instruments or synthesizers, either directly or converted into suitable control signals, for example according to the MIDI standard.

The invention will now be described in more detail by means of exemplary embodiments with reference to the accompanying drawing, in which Figure 1 illustrates a code structure that may be used in the present method, Figure 2 illustrates a lookup table used in the method of the invention, and Figure 3 illustrates related rule expressions.

In a preferred embodiment of the method according to the invention, the individual code types represent notes which in this case have been selected as the smallest musical units used, but alternatively the codes may represent other quantities which can be represented by quantized states. In the context of the present invention, a note is described by two or more quantized properties of the note, such as the pitch and length of the note. Figure 1 illustrates a preferred 16-bit code structure in which the seven least significant bits represent the pitch k, which may thus have 128 different values and one of which may indicate a pause, the next seven bits represent the pitch p, which may also have 128 different values, and the two most significant bits represent stage i of the stroke field of the note 15, i.e. the position of the note in the beat or part of the beat, and thus in the quadratic music used in the example, it can have four different states.

The structure of the stored look-up table used in the method according to the invention is illustrated in Figure 2. The look-up table consists of rule expressions each comprising a conditional part X, a consequential part Y and a conflicting point Z. Flag Z state 0 is a valid expression and state 1 is an invalid expression.

The basic principles for generating rule expressions 25 of a look-up table according to the invention are generally set out in the above-mentioned article [1]. The invention applies a special case of the procedure presented in the article, in which only the preceding codes to the left of the code under consideration are taken into account when forming rule expressions.

The procedure can be viewed with a simple example sequence where the letters represent code types: ABCDEFG ... IKFH ... LEFJ ...

35 Now, if we were to try to derive the following code as an example of 6 84670 on the basis of code F alone (which occurs several times), a triple contradiction would arise. Code F could be followed by any of the codes G, H or J. If an attempt is made to increase the accuracy of the code patterns by including the symbol in front of code F in the content of condition 5, there would still be a double contradiction: EF could be followed by G or J. however, an already valid rule expression, which thus results in an unambiguous code H. In cases where F is preceded by E, two symbols in front of the code F may be included in the conditional part, in which case all conflicting situations can be resolved. Thus, two valid rule clauses can be formed in relation to these points, one of which is a conditional part DEF and the consequence part is G and the other 15 conditional part LEF and the consequence part J.

However, invalid rules created during the generation of rule expressions are not destroyed because they are needed both when creating the rule tree described below and when using it to generate new codes for the code sequence. Instead, each generated rule is marked as valid or invalid by the aforementioned conflict flag Z. Similarly, rule expressions are searched separately for each code in the example sequence. This creates a Lookup Table that contains 25 valid and invalid rule expressions with varying lengths of conditionals.

The information structure of the rule expressions stored in the look-up table can be illustrated by a graphical representation similar to Fig. 3, which connects the rules 30 to each other. For each particular code (such as F), there is a tree rooted by a rule expression in which the condition part contains only this code. If there were conflicts in the example sequences, at least two branches branch from this root, leading to nodes with a code other than code F in the conditional part of rule rule 35.

1 84670

The consequence sections of the branches are written next to the branches. The last nodes in the tree that represent the leaves correspond to the final valid rules, while all other nodes correspond to the invalid rules.

5 Next, consider generating a new symbol for the code sequence. Assume that the starting sequence is CDEF. When searching for a new code, a progressively increasing key sequence formed on the basis of one or more recently generated codes of the initial sequence is utilized. 10 The key sequence is always initially formed by the last code of the start sequence, in this case F. When the start sequence is now compared with the look-up table of Figure 2, it is found that a rule expression with such a condition has a conflicting state of 1, so it is invalid. Then, the code E preceding the last code F of the initial sequence is added to the beginning of the key sequence, whereby the key sequence increases to two codes in length and is EF. Comparing the look-up table in Figure 2, this look also results in an invalid rule. The last code of the third to the last sequence of the initial sequence, i.e. D, is then added to the key sequence, whereby the DEF is formed as the key sequence. This search with the key sequence results in a valid rule in memory location 13, in which case the code G indicated by the consequence of this rule is added to the end of the initial sequence. This incremental code sequence is now used as a new increment, the first key sequence containing the code G. rule.

The basic method described above easily results in a 30 code sequence that makes copies of parts of the example material and may even begin to reproduce itself.

Therefore, in the preferred embodiment of the present invention, the code indicated by the valid rule expression found is not always selected as the new code, but a kind of random selection is used in which the intensity of the variations 8,84670 can be adjusted. Such generation of new, partially random sequences can be illustrated by the recent example and Figure 3. Assume again that the initial sequence is CDEF, and search for a valid rule-5 expression according to the previous example. In this case, the search begins at the root F of Figure 3 and follows the branches until a corresponding leaf, in this case DEF, is found. In a recent example, the code G indicated by the found leaf DEF was selected as the new code. However, there are possibly several nodes in the route from root to leaf with different invalid expressions giving alternative new codes. In the present method, the last key sequence that resulted in finding a valid rule expression is randomly truncated by up to a predetermined number of codes, and an invalid rule expression whose conditional part corresponds to this truncated key sequence is randomly selected as the 15 new codes. In other words, we return randomly, at most a limited number of steps back from the rule tree leaf. In this way, randomness is provided in the generated code-20 sequence, which, for example in the case of a note code sequence, leads to variability and surprising changes in the music produced, but still follows a certain regularity.

In the above preferred embodiment 25 of the invention, both the key sequence and the conditional parts of the rule expressions consisted of basic codes only. While the music produced using such a method to produce note code sequences gives a sense of musical continuity, even more harmoniously more regular melodic developments are favored in typical Western music.

This object is achieved, for example, by another embodiment of the invention, in which, for example, only the last two symbols in the key sequence and the conditional part of the rule expression consist of absolute codes and the preceding symbols represent higher level information, each representing a different combination of two or more code groups. Expressed in musical terms, higher level symbols may represent, for example, the chords that best describe melody sequences, 5 formed by notes of previous beats, half beats, quarter beats, etc., and in which the order of the notes in the combination may be arbitrary. Instead of musical chords, other accumulations of note combinations (his-10 tograms) present in the example material can also be used. In this second embodiment, a conditional part of a rule expression is formed from, for example, the last two note codes as described above, but the preceding example sequence is considered in groups of two or more codes compared to a pre-dated higher level symbol library, and when a code combination is identified as a symbol to the conditional part of the rule expression to the left of the first two codes. In the same way, higher-level symbols can be extended from the 20th to the left end of the conditional part. Thus, rule expressions may have one or more such high-level symbols or none at all. Because a melody may contain notes that do not belong to chord or accumulation histograms, the identification of symbols from a short code string must be based on approximate pattern recognition techniques.

The generation of a new code sequence by these rule expressions takes place, for example, by generating the two shortest forms of key sequences as above, and subsequent forms of key sequences by recognizing combinations of two or more codes in the initial sequence and replacing them in the key sequence. with higher level symbols corresponding to the closest combinations.

Also in this embodiment, the random selection mentioned above 10,84670 can be used.

When producing polyphonic music by the above method, a first note code sequence corresponding to the main sound or melody is first formed using the first look-up table as above, and then one or more additional look-up sequences are formed using this first note code sequence as the start sequence. code sequence, each corresponding to one accompanying sound. 10 The additional look-up tables mentioned above contain their own rule expressions for each accompanying voice.

The method according to the invention is in particular intended for producing note code information in digital form for controlling electronic musical instruments 15 or synthesizers or similar devices. The generated note codes can be formed into control signals according to the MIDI standard, which are fed to the above-mentioned devices. MIDI stands for Musical Instrument Digital Interface and means a standard-20 interface through which, for example, synthesizers, rhythm machines, computers, etc. can be connected to each other. Information on the MIDI standard can be found, for example, in the following publications: [2] MIDI 1.0 Specification, Document No. MIDI-1.0, August 1983, International MIDI Association.

The figures and the related description are only intended to illustrate the present invention. The details of the method according to the invention may vary within the scope of the appended claims.

Claims (8)

11 84670 A method for generating a digital code sequence, in particular a note code sequence, from a finite number of different code types, each representing, for example, one or more quantized properties of a particular note, the code sequence generating one code at a time based on the codes already in the sequence. the generation comprises at least the following steps: a) generating a set of progressively increasing key sequences based on the last generated codes of the code sequence such that the first and shortest 15 key sequences are generated based on the last generated code of the code sequence, followed by one or more code steps increasing , (b) when generating key sequences, they are compared with 20 look-up tables containing valid rule expressions, each with a different law and at least two rule clauses, each with at least two rule clauses having the same condition but a different consequence, until at least an approximate match is found between the key-25 sequence and the condition part of a valid rule expression, and (c) the new code is selected as the new code. the code indicated by the consequence of the valid rule expression, or the code indicated by the consequence of the rule expression randomly selected from among the invalid rule clauses comprising a condition at least approximately equal to the condition of the final key sequence by at least a predetermined code number shorter. A method according to claim 1, characterized in that the codes of the i2 84670 code sequence are used directly as the key sequence, starting with the key sequence with the last generated code of the code sequence and adding one code at a time before the code sequence until a match 5 is found. between the part and the key sequence. A method according to claim 1, characterized in that the first and shortest forms of the key sequence are formed on the basis of the last code of the code sequence, possible subsequent key sequence forms based on two or more last codes and all subsequent forms by gradually adding higher level symbols in front of the previous code sequences. is formed by identifying combinations of two or more codes in the code sequence earlier in the code sequence and replacing them in the key sequence with corresponding, proximal combinations corresponding to the closest symbols, and that in at least some of the rule expressions the search portion consists of one or more code terms; degree symbolic term. A method according to claim 3, characterized in that each higher level symbol represents a chord formed by notes occurring in a part of a beat, in one beat or in several beats, or sequences formed by such chords. Method according to claim 3 or 4, characterized in that the identification of the code combinations as certain higher level symbols is performed by a statistical stacking or other statistical approximation pattern recognition method. Method according to one of the preceding claims, characterized in that at least one additional code sequence, e.g. a note code sequence representing a melody accompaniment 35, is generated using a progressively increasing key sequence formed on the basis of the codes of the first code 13464670 and at least one additional look-up table. Method according to one of the preceding claims, characterized in that if a key sequence corresponding to a conditional part of a valid rule expression in a look-up table or a look-up table cannot be generated when generating code sequence 5 or additional code sequence, a random number of pre-generating codes 10 are removed from the first code sequence continue on the basis of the remaining codes. Method according to one of the preceding claims, characterized in that control sequences in accordance with the MIDI standard for controlling electronic musical instruments or synthesizers are formed from the code sequences. 14 84670