JP5330720B2 - Chord identification method, chord identification device, and learning device - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To provide technology for identifying a chord with high accuracy, without performing complicated operation. <P>SOLUTION: A vector storage section 101 stores a learning vector 100 for expressing a feature of one musical tone which is input by grouping it. A hyper-plane calculation device 102 calculates a hyper-plane for each group by using the learning vector 100 belonging to the group. A parameter for expressing the hyper-plane is stored in one of corresponding storage devices 111 to 114. Evaluation devices 121 to 124 refer to the parameter of the corresponding storage devices 111 to 114, and calculates a determination object vector 150 for expressing the chord and a distance including positive/negative between the vector and the corresponding hyper-plane, as an evaluation value. A determination device 130 determines a group to which a musical tone constituting the chord expressed by the determination object vector 150 belongs, from the evaluation value. <P>COPYRIGHT: (C)2010,JPO&amp;INPIT

Description

 The present invention relates to a technique for identifying a chord composed of a plurality of musical sounds.

  FIG. 6 is a diagram showing a configuration of a conventional chord identification device. With reference to FIG. 6, a conventional chord identification device will be specifically described. As shown in FIG. 6, the identification device includes a selector 601 to which a determination target vector 600 is input, a plurality of filters 621 to 624, a comparison selector 630, a register 632, and a storage device 640. .

  The determination target vector 600 is waveform data obtained by sampling a chord waveform signal, for example. The selector 601 inputs the determination target vector 600 and the vector 634 output from the register 632, and outputs one of them to the filters 621 to 624.

  The chord identification apparatus performs a cyclic operation by switching the vector selected by the selector 601 or updating the vector 634. Through the cyclic operation, the musical tones (frequency) constituting the chord are sequentially determined. At the start of chord identification, that is, in the first cycle, the determination target vector 600 is selected by the selector 601.

  The filters 621 to 624 correspond to specific frequencies F1 to F12, respectively. That is, for example, the filter 621 removes only the signal component corresponding to the frequency F1 and the harmonic frequency thereof (= F1 × k: k is an integer of 2 or more) from the vector input from the selector 621. To work. Therefore, the signal output from the filter 621 becomes a power spectrum obtained by removing the frequency F1 and the signal component of the frequency that is a harmonic thereof from the input vector. FIG. 6 shows only a total of four filters 621 to 624 as filters, but more filters are actually mounted.

  The comparison selector 630 selects the minimum signal (power spectrum) input from each of the filters 621 to 624 and outputs it as a vector 631 to the register 632. Thus, for example, when the minimum signal is input from the filter 621, the vector 631 is selected from the input vector (the determination target vector 600 at the start of chord identification) from which the signal component of the frequency F1 has been removed. -It will be output. Further, the comparison / selection device 630 outputs the filter that has selected the signal, that is, information on the frequency having the largest signal component (frequency information) to the storage device 640 as the signal 635. Thereby, the storage device 640 stores the signal 635. Hereinafter, the signal 635 will be referred to as “frequency information”.

  Saving the frequency information 635 completes the operation of one cycle. In the next second cycle, the selector 601 selects the vector 634 and the same processing is performed.

  The vector 634 selected in the second cycle, that is, the vector 631 stored in the register 632 is the determination target vector 600 from which the maximum frequency component has been removed. Therefore, the comparison selector 630 outputs the determination target vector 600 from which the second largest frequency component is removed in addition to the maximum frequency component to the register 632 as the vector 631. As the frequency information 635, frequency information indicating a frequency having the second largest signal component is output.

In this way, chords are identified by sequentially determining frequencies (musical tones) in descending order of signal components. The determination is performed until the level of the signal selected by the comparison selector 630 is equal to or less than the value set as the threshold value. By making such a determination, the number of musical tones constituting the chord is specified.
JP 2004-341026 A

  In the conventional chord identification apparatus shown in FIG. 6, as described above, the chord is identified in such a manner that the frequency is determined for each musical tone constituting the chord. However, in such a method, the actual situation is that the accuracy of identification is low. The reason is that there are harmonics in a musical tone, and the harmonics of one musical tone may overlap with the fundamental wave of another musical tone. In particular, in the case of chords, there are many cases in which such overlapping occurs because the musical tones that form the chords are often in a relationship of consonant pitches.

FIG. 7 is a diagram for explaining an example of frequency components constituting a musical sound. The horizontal axis represents frequency and the vertical axis represents power. 701 is the signal component of the fundamental wave (fundamental frequency), 702 is the signal component of the second harmonic (frequency twice the fundamental frequency), and 703 is the signal component of the third harmonic (frequency of 3 times the fundamental frequency). Represents. Reference numeral 711 denotes a signal component of a fundamental wave whose fundamental frequency is three times higher than that of 701. Accordingly, if the fundamental wave of 701 is a pitch of “do”, the fundamental wave of 711 corresponds to the pitch of “so”.

  As shown in FIG. 7, the musical tone includes harmonics. In order not to determine each harmonic as one musical tone, each filter 621 to 624 shown in FIG. 6 also removes signal components corresponding to the harmonics. However, when the signal component corresponding to the harmonic is removed in this way, in the example shown in FIG. 7, the signal component 711 of another fundamental wave is removed together with the signal component 701 of the fundamental wave, and another fundamental wave is removed. That is, the signal component 711 of FIG. As a result, it becomes impossible to determine a musical tone to be determined, which leads to a decrease in accuracy of identifying a chord. FIG. 8 is a diagram illustrating a chord identification result obtained by a conventional chord identification apparatus. The lower half of the figure shows the erroneous judgment result, and the upper part shows the musical tone judgment result by the comb filter.

  Chords can be identified in a form that does not individually determine the musical sounds that make up the chords. In other words, it can be identified by learning. However, the number of chords composed of several musical tones is very large considering the combination of pitches, that is, the root pitch. For this reason, complicated work must be performed for learning, and a very long time is required. Considering this, it is considered important to improve chord identification accuracy while avoiding complicated work.

  An object of the present invention is to provide a technique for identifying chords with higher accuracy without performing complicated operations.

The chord identification method of the present invention is a method for identifying a chord composed of a plurality of musical tones by a computer, and for each musical tone that can constitute a chord , a specific musical tone using musical tone characteristic information representing the musical tone. A learning process for calculating a hyperplane that is an index for distinguishing the tone from other musical tones, and distance to the chord using chord feature information representing the power spectrum of the chord to be identified for each hyperplane And determining a musical tone constituting the chord using the calculation result.

  Here, “by tone that can compose a chord” means that two or more tones that make up the same chord do not overlap on the same hyperplane. In other words, the musical tones that make up the same chord correspond to different hyperplanes. Therefore, there may be a plurality of musical sounds corresponding to the hyperplane.

The chord identification device of the present invention is based on the premise that a chord composed of a plurality of musical tones is identified, and for each musical tone that can constitute a chord , a specific musical tone calculated using musical tone characteristic information representing the musical tone, Information storage means for storing hyperplane information, which is a parameter representing the hyperplane that is an index for distinguishing other musical tones, and information input for inputting chord feature information representing the power spectrum of the chord to be identified Using the hyperplane information stored in the information storage means, for each hyperplane, an evaluation means for calculating the distance between the chords represented by the chord feature information, and the distance calculated by the evaluation means for each hyperplane , comprising a tone determining means for determining a tone constituting the chord, the hyperplane information corresponds to each group to divide the tone characteristic information, grouping dividing the musical sound feature information, two or more of the same chords Musical tone to so that it does not belong to the same group.

The learning device according to the present invention is based on the premise that information for identification for identifying a chord composed of a plurality of musical sounds is generated by learning, and input musical feature characteristic information representing the musical sound for each musical sound that can constitute the chord. For identifying a hyperplane that is an index for distinguishing a specific musical sound from other musical sounds by using the musical sound characteristic information for each musical sound represented by the musical sound characteristic information input by the information input means and the musical sound characteristic information input by the information input means Hyperplane calculating means for calculating as information.

  In the present invention, for each tone that can compose a chord, a hyperplane calculated using tone characteristic information representing the tone is used for chord identification. For each hyperplane, the distance from the chord is calculated using the chord feature information representing the chord to be identified, and the musical tone constituting the chord is determined using the calculation result.

  The hyperplane calculated using the musical tone characteristic information is an index for distinguishing a specific musical tone from other musical tones. In the identification of chords, the reproducible tendency peculiar to the distance to the hyperplane and the direction (positive / negative) was confirmed by experiments, based on the musical tones that make up the chord. For this reason, by using a hyperplane for chord identification, the identification can be performed with higher accuracy. Since it is not necessary to use chords (feature information) for the calculation of the hyperplane, the learning can be performed more easily and quickly than in the case of learning by chords.

 Hereinafter, embodiments of the present invention will be described in detail with reference to the drawings.

  FIG. 1 is a diagram showing a configuration of a chord identification device according to the present embodiment.

  The chord identification device performs learning using the learning vector 100 representing the characteristics of one musical tone, and identifies the chord. As shown in FIG. 1, a vector storage device 101 that stores a learning vector 100, a hyperplane calculation device 102 that calculates a hyperplane for each group into which the learning vector 100 is divided, and a corresponding hyperplane prepared for each group. Storage devices 111 to 114 that store parameters to be expressed, evaluation devices 121 to 124 that are prepared for each group and that calculate the distance between the corresponding hyperplane and the determination target vector 150 as an evaluation value, and the evaluation devices 121 to 124. And a determination device 130 that inputs an evaluation value and identifies a chord.

  As described above, in this embodiment, the learning vector 100 representing the characteristics of one musical tone is used for identifying chords. The learning vector 100 is stored in the vector storage device 101 for each group as shown in FIG. In order to store such information, for example, the learning vector 100 is output to the vector storage device 101 together with group information indicating a group to which the learning vector 100 belongs.

  In the present embodiment, the musical tone constituting the chord is determined by specifying the group to which the musical tone constituting the chord belongs. For this reason, the grouping is performed so that two or more musical tones do not belong to the same group with the same chord. In other words, the musical tones constituting a certain chord are performed so as to belong to different groups. Accordingly, “A1”, “B1”, and “C1” in FIG. 2 represent musical sounds (learning vectors 100) having different pitches (pitches). For the learning vector 100 representing a group, for example, information indicating the type of the group information and the learning vector 100 (musical sound information) such as “<J, A1>”, “<J, A2>”, “<K, B1>” The notation using is adopted.

  The learning vector 100 is the same multidimensional vector as the determination target vector 150. Specifically, for example, a power spectrum indicating the magnitude (power) of a signal component for each frequency. However, for ease of understanding here, each vector 100 and 150 assumes two dimensions. In two dimensions, the hyperplane is a straight line. The generation of the learning vector 100 and the determination target vector 150 is performed using a known technique.

  The number of musical tones belonging to the same group is very small compared to the number of possible chords (number considering the pitch of the root tone in addition to the type). For example, even if attention is paid only to a chord composed of 3 musical tones out of 5 musical tones, the number is 10 (kinds). Therefore, by adopting one musical tone as the learning vector 100, the labor in the learning process is greatly reduced and the required time is also shortened.

  The hyperplane corresponding to group A can be calculated by finding P and Q that maximize DA in equation (1).

  However, Ai is a vector belonging to group A, and Zi is all vectors not belonging to group A. P is a vector of the same dimension as Ai, and Q is a scalar.

  The hyperplane is calculated in the same way for the other groups. Parameters representing the hyperplane corresponding to each group are stored in the storage devices 111 to 114 corresponding to the group. By storing all the parameters in the storage devices 111 to 114, the learning process (step) is completed. By the grouping as described above, the hyperplane is calculated for each tone that can compose a chord. The learning device according to the present embodiment calculates a hyperplane for each musical tone. For this reason, for example, it is realized by the vector storage device 101 and the hyperplane calculation device 102. The chord identification device according to the present embodiment is equipped with the learning device, and the function of the chord identification device itself is realized by a portion excluding the vector storage device 101 and the hyperplane calculation device 102.

FIG. 3 is a diagram for explaining a hyperplane calculated using the learning vector 100. The hyperplane 300 shown in FIG. 3 is calculated when the learning vector 100 is two-dimensionally expressed in the xy plane. In that case, the hyperplane is a straight line represented by the following equation. Note that 311 to 313 and 321 to 323 in FIG. 5 represent three learning vectors 100 belonging to groups J and K, respectively.
ax + by + c = 0 (2)

  In the hyperplane that can be expressed as in Expression (2), a, b, and c in Expression (2) are stored in the respective storage devices 111 to 114 as parameters.

  Each of the evaluation devices 121 to 124 calculates the distance between the hyperplane represented by the parameter and the determination target vector 150 using the parameter stored in the corresponding storage device among the storage devices 111 to 114, and The calculation result is output as an evaluation value. In the example shown in FIG. 3, the evaluation value is positive for the learning vectors 311 to 313 belonging to the group J and negative for the learning vectors 321 to 323 belonging to the group K. Thereby, the positive / negative of the evaluation value has an important meaning for specifying the group.

  The determination device 130 inputs evaluation values (here, a total of n (n is an integer of 5 or more) evaluation values) from each of the evaluation devices 111 to 114, and the chord represented by the determination target vector 150 based on the magnitude and positive / negative. The group to which the musical sound belongs is determined for each musical sound that constitutes. The chord is identified by determining the musical sound belonging to the group for each determined group. Musical sounds belonging to a group can be accurately identified or estimated from the frequency at which the signal component is maximum, that is, the fundamental frequency of the root tone.

  FIG. 4 is a diagram for explaining a hyperplane calculated in the learning process and a chord identification method using the hyperplane. Here, with reference to FIG. 4, the chord identification method performed using the hyperplane calculated in the learning process will be described in detail. In FIG. 4, a straight line 401 is a hyperplane calculated by the group to which the learning vectors 411 to 413 belong (here, from FIG. 2 to group J), that is, the group J is one group, and the other group is another group. The hyperplane calculated as a group, the straight line 402 is the hyperplane calculated in the group to which the learning vectors 421 to 423 belong (also referred to as group K from FIG. 2), that is, the group K is one group, and the other groups The hyperplane calculated as another group, the straight line 403 is the hyperplane calculated in the group to which the learning vectors 431 to 433 belong (also referred to as group L from FIG. 2), that is, the group L is one group, and the others The hyperplane calculated with one group as another group is shown.

  Since these three groups of hyperplanes are shown, it is assumed in FIG. 4 that there are only three musical sounds A, B, and C, and a chord composed of two musical sounds is identified. In that case, there are three types of chords: A + B, B + C, and C + A. 4 in FIG. 4 represents a vector input as the determination target vector 150 of the chord A + B. e1 to e3 indicate distances between the straight lines 601 to 603 and the determination target vector 600, respectively. The distance e1 is negative and has a small absolute value, the distance e2 is positive and has a small absolute value, and the distance e3 is negative and has a large absolute value. The magnitude relationship in absolute value is e3> e2> e1.

  The correspondence between the evaluation device and the group is one-to-one. Therefore, the group to which the musical sound constituting the chord belongs is determined by extracting the distance corresponding to the number of constituent sounds of the chord from the input distance having the larger absolute value. Thereby, in the example shown in FIG. 4, distances e3 and e2 are extracted. Since the distances e3 and e2 correspond to the straight lines 603 and 602, respectively, the two musical sounds constituting the chord represented by the determination target vector 600 are determined to belong to the groups L and K, respectively. .

  FIG. 5 is a diagram showing a chord identification result of the chord identification device according to the present embodiment. As in FIG. 8, an erroneous determination result is shown at the bottom of the figure, and a musical sound determination result by the comb filter is shown at the top. Compared with FIG. 8, it can be seen that the chords are identified with higher accuracy. This has been confirmed from other experiments.

  In this embodiment, the learning device is mounted on the chord identification device. However, the chord identification device and the learning device may be mounted on different devices or realized as different devices. In addition, any of these devices can be realized by causing a computer to execute a program that realizes the functions described above. For this reason, such a program may be recorded and distributed in a portable flash memory or the like, or may be received via a communication network.

It is a figure which shows the structure of the chord identification apparatus by this embodiment. It is a figure explaining the learning vector memorize | stored in a vector memory | storage device. It is a figure explaining the hyperplane calculated using the learning vector. It is a figure explaining the identification method of the chord using the hyperplane calculated in a learning process, and the hyperplane. It is a figure which shows the identification result of the chord of the chord identification apparatus by this embodiment. It is a figure which shows the structure of the conventional chord identification apparatus. It is a figure explaining the example of the frequency component which comprises a musical tone. It is a figure which shows the identification result of the chord by the conventional chord identification apparatus.

Explanation of symbols

100, 311 to 313, 321 to 322, 411 to 413, 421 to 423, 431 to 433 Learning vector 101 Vector storage device 102 Hyperplane calculation device 111 to 114 Storage device 121 to 124 Evaluation device 130 Determination device 150 Determination target vector

Claims (6)

A method for identifying a chord composed of a plurality of musical sounds by a computer,
A learning step of calculating a hyperplane that is an index for distinguishing a specific musical tone from other musical tones using musical tone characteristic information representing the musical tone for each musical tone that can compose the chord;
A determination step of calculating a distance to the chord using chord feature information representing a power spectrum of a chord to be identified for each hyperplane, and determining a musical tone constituting the chord using the calculation result; ,
A chord identification method characterized by comprising: In a chord identification device that identifies a chord composed of multiple musical tones,
Hyperplane information that is a parameter representing a hyperplane that is an index for distinguishing a specific musical tone from other musical tones calculated using musical tone characteristic information representing the musical tone for each musical tone that can compose the chord. Information storage means for storing
Information input means for inputting chord feature information representing the power spectrum of the chord to be identified;
Using the hyperplane information stored in the information storage means, for each hyperplane, evaluation means for calculating the distance from the chord represented by the chord feature information;
From the distance calculated by the evaluation means for each hyperplane, a musical sound determination means for determining a musical sound constituting the chord;
The hyperplane information corresponds to each group that divides the musical tone characteristic information, and the grouping that divides the musical tone characteristic information is such that two or more musical sounds with the same chord do not belong to the same group. A chord identification device characterized by: In a learning device that generates identification information for identifying a chord composed of a plurality of musical sounds,
Information input means for inputting musical tone characteristic information representing the musical tone for each musical tone that can compose the chord,
For each musical tone represented by the musical tone characteristic information input by the information input means , a hyperplane, which is an index for distinguishing a specific musical tone from other musical tones, is calculated as the identification information using the musical tone characteristic information. Hyperplane calculation means;
A learning apparatus comprising: A program for causing a computer to perform processing for identifying a chord composed of a plurality of musical sounds by a computer,
To the computer,
A learning process for calculating a hyperplane that is an index for distinguishing a specific musical tone from other musical tones using musical tone characteristic information representing the musical tone for each musical tone that can constitute the chord;
A determination process for calculating a distance from the chord using chord feature information representing a power spectrum of the chord to be identified for each hyperplane, and determining a musical tone constituting the chord using the calculation result; ,
The line wa cell program. Used in a chord identification device that identifies chords composed of multiple musical tones,
In the chord identification device,
Hyperplane information that is a parameter representing a hyperplane that is an index for distinguishing a specific musical tone from other musical tones calculated using musical tone characteristic information representing the musical tone for each musical tone that can compose the chord. Information storage processing that stores
An information input process for inputting chord feature information representing the power spectrum of the chord to be identified;
Using the hyperplane information stored by the information storage process, for each hyperplane, an evaluation process for calculating a distance from the chord represented by the chord feature information;
A musical tone determination process for determining a musical tone constituting the chord from the distance calculated for each hyperplane by the evaluation process ;
Rows of Align, the hyperplane information corresponds to each group to divide the tone characteristic information, grouping dividing the musical sound feature information, as more than one tone in the same chord is not to belong to the same group Program to make . Used in learning devices that generate identification information for learning to identify chords composed of multiple musical sounds,
In the learning device,
For each tone that can compose the chord, an information input process for inputting tone characteristic information representing the tone;
For each musical tone represented by the musical tone characteristic information input by the information input process , a hyperplane that is an index for distinguishing a specific musical tone from other musical tones is calculated as the identification information using the musical tone characteristic information. Hyperplane calculation processing,
The line wa cell program.