JPH04157499A - Automatic rhythm creation device - Google Patents

Abstract

PURPOSE:To easily generate a rhythm sound strictly in accordance with an image and matched to the tune of a music piece by outputting a rhythm information by inference from a performance data by using a neural network. CONSTITUTION:A performance data of a music piece is input to a neural network means. A neural network is a hierarchical network and treated as a software by CPU 1. Rhythm parts are previously learned about a plural number of music pieces in correspondence with melody parts of famous music pieces in the neural network. The neural network outputs a rhythm information of rhythm accompaniment against the music piece of the performance data by inference on the basis of a plural number of parameters input to neurons I1-I11 of an input layer and a synapse weight (w) in ROM 2. Consequently, it is possible to automatically and easily create a rhythm sound based on the music piece.

Description

[Detailed Description of the Invention] (Industrial Application Field) The present invention relates to an automatic rhythm generation device that provides rhythmic accompaniment to any musical piece.

(Prior Art) Conventionally, methods for a performer to create rhythm data for one song in an automatic rhythm generation device can be roughly divided into the following two methods.
Different methods are known.

The first method is to manually store rhythm data one by one, and the second method is to program a rhythm performance sequence by arbitrarily combining multiple rhythm patterns prepared in advance (Japanese Patent Laid-Open No. 159185/1985). (Problem to be Solved by the Invention) However, with the method of inputting the first rhythm data one by one into the storage means, it takes a lot of effort to input the rhythm data for one song. However, with the method of programming arbitrary combinations of performance patterns, there are limits to the combinations of performance patterns, making it difficult to create a desired rhythm performance sequence. However, it is very difficult to create rhythm accompaniment based on the melody, and therefore it is difficult to generate rhythm data suitable for various songs. This was done in view of the circumstances, and the purpose is to provide an automatic rhythm generation device that can easily generate rhythm sounds based on musical pieces.

(Means for Solving the Problems) A performance pattern information output means that receives performance data of a musical piece and outputs performance pattern information obtained by patterning the performance data, and a performance pattern information outputted from the performance pattern information output means. a neural network means for inferring and outputting rhythm information based on the neural network means; and a rhythm data forming means for forming rhythm data indicating rhythm pronunciation timing based on the rhythm information output from the neural network means, the rhythm data forming means The rhythm sound is generated according to the rhythm data formed by the means.

(Function) In the automatic rhythm generation device of the present invention, performance data of a musical piece is input to the neural network means. The neural network means infers rhythm information by transmitting input to the input layer from the intermediate layer to the output layer through synaptic connections. This neural network means
Rhythm information is learned in advance, and by inferring rhythm information using such a neural network means, rhythm sounds based on music can be automatically and easily generated.

(Embodiment) FIG. 1 is a block diagram of an electronic musical instrument that is an embodiment of the present invention. This electronic musical instrument is controlled by CPU 1
Performance sounds and rhythm sounds of the music are generated based on instructions from the performer. The CPUI is connected to each circuit section via a bus 9, and sends and receives data and the like. The circuit section includes a ROM 2, a RAM 3, a storage means 4, an operation panel 5, and a sound source circuit 6. Mounted on the output side of the sound source circuit 6
A sound system (S, S) 7 is connected to the circuit 6 for amplifying musical tone signals and rhythm tone signals of music generated by the circuit 6 and outputting the amplified signals from a speaker or the like.

The storage means 4 stores performance data (consisting of pitch data and note length data) relating to the melody of an arbitrary musical piece, and is made to be able to read and write the performance data, for example, using a semiconductor memory.

The operation panel 5 is provided with a group of switches for setting a plurality of image parameters expressing the characteristics of the rhythm. Here, image parameters include, for example, tone (cymbals, bass drum, snare drum, etc.), genre (rock, hard rock, ballad, samba, etc.), beat (
4 beats, 8 beats, 16 beats, etc.).

The CPUI is a central processing unit that controls data transmission and reception between circuits and performs various calculations including neural network processing.

Neural Network Salt This is a hierarchical bi-ethnic network as shown in FIG. 2, and is processed by the CPU 1 in software. The neural network has previously learned rhythm parts for multiple songs in correspondence with the melody parts of famous songs. This learning is designed to minimize the error between the actual calculation value of the output layer and the rhythm data indicating the timing of the rhythm part of the song when the performance data of the melody part of the song is input. This is performed by changing the synaptic weight W. Note that this learning method can be used, for example, with Backpropagage 1.
Other methods are also possible. This learned synaptic weight W is stored in ROM2
is stored in

The neural network consists of each neuron II in the input layer.
Based on the plurality of parameters inputted to ~III and the synapse weight W in the ROM 2, the rhythm information of the rhythm accompaniment for the music of the performance data is inferred and output. In this case, each neuron from the input layer to the intermediate layer and from the intermediate layer to the output layer is synaptically connected with a predetermined weighting, and the strength of the connection is determined by the synaptic weight W.

The ROM 2 stores a CPUI control program, learned synapse weights W, and the like.

Note that this ROM 2 stores a set of synaptic weights W for each tone color type, for each genre type, and for each beat type.

RAM3 is a readable and writable memory that stores rhythm data indicating the timing of the sound of rhythm sounds obtained from various input parameters, data from each neuron in the input layer, intermediate layer, and output layer, and rhythm information inferred by neural network processing. remember.

The sound source circuit 6 generates sound signals for the musical tones and rhythm sounds of the music based on the performance data of the music stored in the storage means 4 and the rhythm data stored in the RAM 3. This sound signal is sent to the sound source circuit 6. The sound is amplified by the connected sound system 7 and output from a speaker or the like.

Next, assuming that the performance data of the song stored in the storage means 4 is composed of whole notes #half notes, quarter notes, whole rests, half notes, and quarter rests, the second characteristic The rhythm creation operation of the present invention will be explained using FIG.

First, the CPU detects the tone/genre/beat switch states on the operation panel 6, and selects the selected tone/beat.
Each piece of data indicating the genre/beat is stored in the memory portions corresponding to I2, I3, and I4 in RAM3 (
stepL).

The performance data of the melody of the music stored in the storage means 4 is sequentially read out according to instructions from the CPUI. The CPU 1 determines the type of musical rest from the read performance data.One measure is divided into four equal parts and divided into four sections.
The data is converted into data indicating the presence or absence of a note in each of the first to fourth sections divided into sections, and data indicating the connection of each section 1. (step 2L)
The data showing this connection is that musical notes and rests continue between the first sex scene and the second sex scene, the second sex scene and the third sex scene, and the third sex scene and the fourth sex scene. This is data indicating by 110 whether or not. In this case, since one measure is divided into four equal parts, the length of each section corresponds to the length of a quarter note. In the case of a quarter note or a quarter rest, the data indicating the connection will be 0 because it will not continue until the next sekki lesson. If it is a half note or a half rest, it will continue until the next section, so the data indicating the connection will be 1. This data conversion is read based on the quarter note (rest) length. Conversion is performed based on how many times the length of the converted note (rest) is.

Further, the CPU 1 counts the note length data of the performance data sequentially read out from the storage means 4, and when it becomes equal to one measure length, counts the number of measures,
(step 2). Here, the CPU 1 performs each of the above counts by using the register in the RAM 5 as a counter, and uses the counted number of measures and the tone, genre, and beat set on the operation panel 5 in a neural network. Instruct the input layer to input. The input layer of the neural network consists of 12 neurons 11-111,
The number of measures if which represents the number of novels in the song ・Tone 12 (
Cymbals, bass drum, snare drum, etc.) e Genre Mi 3 (rock, hard rock, ballad, samba, etc.)
φbee) I4 (4 beats, 8 beats, 16 beats, etc.)・Presence or absence of notes in each scene i5.17.19, il
Data i8, I8. which represents the connection to the next sexy when it is a half note or a whole note. ilO force ζ is input to each neuron as a parameter.

Here, it is assumed that one measure of performance data is composed of a half note, a quarter note, a quarter note, and a quarter note, as shown in FIG. Since there is a half note in the first sex, the data indicating the presence or absence of a note is 1 and is input to I5.Since the half note continues to be sounded in the second sex, the data indicating the connection is 1. also becomes 1 and is input to I6. Since there is a half note in the second seventh sex, the data indicating the presence or absence of a note becomes 1 and is input to 17. In the third sex, it is 2.
Since no diacritics are produced, the data indicating the connection becomes 0 and is input to I8. 3rd sex, 4 for n
Since there is a minute rest, the data indicating the presence or absence of a note will be 0, which will be input into I9.Also, since this quarter rest does not continue until the fourth sex, the data indicating the connection will also be 0!
10 is input. Since there is a quarter note in the fourth section W, the data indicating the presence or absence of a note becomes 1 and is input to 111. In other words, the input layer of the neural network! 6
~111 contains parameter 1 as performance pattern information.
"1.1.1.0.0.0.1" is input as 5 to ill.

Next, the CPU performs inference calculations based on the parameters X1 to j11 input to the neural network, and forms and outputs rhythm information of rhythm accompaniment (stops).

This inference calculation is performed by the CPU 1 by appropriately reading out the synaptic weight W stored in the ROM 2. Each neuron It to Ill in the input layer is synaptically connected to all neurons in the intermediate layer, and the strength of the connection is determined by the synaptic weight W. That is, the middle layer consists of m neurons N1 to N■, and the data input to Nl is i I Xwl + i 2 Xw2
+・−・+ i IIXwJI, and the same applies to N2 to Nm. Also, each neuron Nl in the middle layer
-Nm is synaptically connected to all neurons of output l. The output layer consists of four neurons 01 to 04, and the data input from ol to 04 is the middle layer N]
- Same as N-buki. Each neuron 01~ of this output layer
04 outputs rhythm information of four cefsilon rhythm sounds obtained by dividing one bar into four equal parts. This rhythm information is
normalized by an arbitrary threshold (e.g., OI≧
If 0.5, then 0i=1), the rhythm data indicates the presence or absence of sound as 1.0 (step 4). That is, a rhythm pattern for one bar can be created by the output of each neuron Of~04. This rhythm data is RA
It is stored in M3 (step 5). In this way, the performance data of the song is input one measure at a time, one measure of rhythm data is created, and it is sequentially stored in RAM3 (s t
By doing this, it is possible to create rhythm data for rhythm accompaniment based on the music and corresponding to the tone, genre, and beat specified by the operator.

Rhythm data formed in this manner and stored in the RAM 3; tags that are sequentially read out in response to a performance start instruction in the same way as a well-known automatic rhythm performance device; rhythm sounds from the sound source circuit 6 at timings according to the rhythm data; A signal is output and a rhythm sound is produced. Further, by reading out the rhythm data stored in the RAM 3 in synchronization with reading out the performance data of the music piece, the musical tones and rhythm sounds of the music piece can be produced simultaneously.

In this example, ROM2 that has been trained in advance
The force ζ that causes the neural network to perform calculations according to the synaptic weights of ζ may be made to be learned arbitrarily by the user.

In this embodiment, the performance data may be input in real time from the keyboard by using a high-speed calculation means.

In addition, by selecting n tones on the operation panel and providing (equal fraction of the output layer) Xn neurons in the output layer,
It is possible to output multiple types of rhythm rhythms at once.

In addition, the force ζ that has neurons for one measure in the input layer and output layer of the unilateral network can be any length, and 2
It may be a measure or one song.

Furthermore, in this example, one measure is divided into four equal parts and divided into four quarter notes. mark)
- Since it can support sixteenth notes (rests), etc., it is possible to add richer rhythmic accompaniment to more complex melodies.

The neural network may perform processing using hardware instead of the CPU if the same function can be achieved.

In this example, the force ζ pitch data, which does not take into account the pitch of the music (performance data), is input into a predetermined input layer of the neural network by digitizing it into a predetermined value, and the pitch change pattern and the rhythm are By learning the 221 Ral network for songs that have a certain relationship, rhythm data can be created based on both note length data and pitch data, and rhythm sounds more suitable for the song can be obtained.

(Effects of the Invention) As described above, the automatic rhythm generation device of this invention is effective. !
By inferring and outputting rhythm information from performance data using a neural network, it is possible to easily generate rhythmic sounds that are exactly as imagined and match the idea of the song.

[Brief explanation of drawings]

Fig. 1 is a block diagram of an automatic rhythm generation device according to an embodiment of the present invention. Fig. 2 shows a schematic configuration of a neural network used in the automatic rhythm generation device. Fig. 3 is a flowchart showing the operation of the CPU. . -CPU -ROM -RAM 4-Storage means 5-Operation panel 6-Sound source circuit 7-5. S (sound system) 8-bass

Claims (1)

[Scope of Claims] Performance pattern information output means that receives performance data of a piece of music and outputs performance pattern information that is patterned from the performance data; Neural network means for inferring and outputting information; and rhythm data forming means for forming rhythm data indicating rhythm pronunciation timing based on the rhythm information output from the neural network means, the rhythm data forming means forming rhythm data. An automatic rhythm generation device characterized in that it generates rhythm sounds according to rhythm data that has been generated.