JPH03219296A - Musical tone generating device - Google Patents

Abstract

PURPOSE:To easily generate a musical tone as imaged by using a neural network which infers and outputs a waveform determination parameter based on plural image parameters representing the features of timbre. CONSTITUTION:The neural network 14 functions to determine the waveform determination parameter of an FM sound source according to input image parameters. This neural network 14 is a hierarchical neural network and the image parameters are inputted to the input layer from plural timbre image specification dials, so that the waveform determination parameter is outputted to the output layer by inferring based upon the parameter. The waveform determination parameter inferred and outputted by the neural network 14 and algorithm are set in a sound source circuit 17. The FM sound source circuit 17 generates an operator according to plural waveform determination parameters, which are synthesized according to the specific algorithm to generate one musical tone. Consequently, the musical tone can be generated as imaged.

Description

DETAILED DESCRIPTION OF THE INVENTION (Al Industrial Field of Application) The present invention relates to a musical tone generator equipped with a tone generator circuit such as a so-called FM tone generator.

(bl Conventional technology) FM sound sources have been put into practical use.
Independently basic waveform (sine wave, triangle wave, etc.) frequency (pitch)
This is a sound source that generates one musical tone by synthesizing (modulating, adding) a plurality of (4 or 6) operators that can set parameters such as , output level, envelope, etc. based on a specific algorithm. This sound source can generate beautiful musical tones rich in variation with a simple configuration.

(C) Problems to be Solved by the Invention However, it has been said that it is difficult for this FM sound source to generate tones as intended by the author. This is mainly
Since musical tones are generated using multiple parameters and FM modulation, it is difficult to imagine how musical tones will change due to changes in operators or algorithms.

The present invention provides a musical tone generating device that can easily generate musical tones as imagined by using a neural network that infers and outputs waveform determining parameters based on a plurality of image parameters expressing characteristics of timbre. With the goal.

(Means for Solving Problem d1) This invention includes a sound source circuit that generates operators based on a plurality of waveform determination parameters, and synthesizes the generated plurality of operators using a predetermined algorithm to generate one musical tone. A musical tone generator comprising: a means for inputting a plurality of image parameters representing characteristics of a timbre; a neural network for inferring and outputting the waveform determining parameter and the algorithm based on the input plurality of image parameters; and the neural network. means for setting the waveform determining parameters and algorithm inferred and outputted by the sound source circuit into the sound source circuit.

(el) Effects of the Invention The musical tone generator of the present invention generates musical tones by synthesizing a plurality of operators generated based on a plurality of waveform determining parameters. A neural network infers and outputs waveform determining parameters and synthesis algorithms for generating musical tones. The neural network is trained in advance so that it can output the parameters of the musical tone that a typical user imagines. Image parameters expressing characteristics of musical tones are used as input data for inference. As the image parameters, parameters such as the hardness/softness of the sound, the flashiness/subduedness of the sound, the beauty/dirtyness of the sound, and the thickness/thinness of the sound may be used, for example.

By inputting an image, it is possible to produce a musical tone that matches the image.

ffl Embodiment FIG. 1 is a block diagram of a control section of an electronic musical instrument according to an embodiment of the present invention. This electronic musical instrument is controlled by the CPU 10 and generates two musical tones based on the player's operations. CPU10 is connected to each circuit section via a bus 11. As a circuit part, ROM12. RA? l/11
3. Neural network (NN)14. Keyboard 15. Operation panel 16. There is an FM sound source circuit 17. F
A sound system 18 is connected to the M sound source circuit 17 for amplifying the generated musical tones and outputting them from a speaker or the like.

The ROM 12 stores synapse weights for programs and presets, and the RAM 13 stores registers for storing various data generated during performance, as well as synapse weights that are the result of arbitrary learning by the user. There is also a neural network 14 in which the weights are stored.
has a function of determining waveform determining parameters for the FM sound source based on input image parameters. This neural network 14 is a hierarchical neural network as shown schematically in FIG.
(Image parameters are input from f, and waveform determination parameters as shown in Figure 2 are output to the output layer through inference based on these parameters. This neural network is capable of the hierarchical inference shown in the figure. The keyboard 15 is for performance and has a range of about five octaves.The operation panel 16 also has the tone image designation dials 16a-16d. In addition to the above, a learning mode/normal mode selection switch, a synapse weight selection switch, etc. are provided.Learning mode means inputting an image of the currently set waveform determining parameters etc. using the tone image designation dial. This mode is used to learn the tones.Normal mode is the normal performance mode.The synapse weight selection switch is
The FM sound source circuit 17, which is a switch for selecting whether to use the synapse weights stored in advance in the ROM 12 or the synapse weights learned in the learning mode and stored in the RAM 13, is configured to select whether to use the synapse weights stored in advance in the ROM 12 or the synapse weights learned in the learning mode and stored in the RAM 13. This circuit determines the content by setting a plurality of parameters, and generates musical tones by synthesizing these operators using a specified algorithm (-how to combine operators, how to modulate). By appropriately adjusting each parameter and algorithm, complex tonal changes and high-order overtones can be obtained. Each parameter and algorithm is cented by CPL;10 before performance.
Based on the key-on signal and key code sent from CPUl0 during performance, a musical tone of the pitch is generated.

FIG. 2 shows a schematic diagram of the tone color image designation dials 168 to 16d and a schematic diagram of the configuration of the neural network 14. The tone image designation dials 162 to 16d can set the degree of four types of tone images, 16a is the hardness of the sound (hard/soft), 16b is the thickness of the sound (thick/thin), 16c is is a dial that specifies the beauty of the sound (beautiful/dirty), and 16d is a dial that specifies the level of flashiness (flashy/quiet) of the sound. The values designated by these tone color image designation dials 162 to 16d are input to the neural network 14 as image parameters. These four image parameters are input to the input layer of the neural network 14.

Each neuron in the combination layer of the input layer is synaptically connected with a predetermined weight, and each neuron in the combination layer and the output layer is also synaptically connected with a predetermined weight. The output of each neuron in the output layer corresponds to the waveform determining parameters and algorithm of each operator. The waveform determining parameters for each operator are shown in this figure as ON/OFF to simplify the explanation.

Although only the frequency ratio (the pitch to the frequency of the musical tone to be sounded) and the envelope rate are shown, the actual operator is determined by many more parameters. This output is sent to the FM sound source circuit 17 via the CPU10, FM
The sound source circuit 17 generates musical tones based on these parameters.

This neural network 14 is trained in advance using a plurality of pieces of teacher data, and based on this learning, it is possible to determine output parameters commensurate with whatever parameters are input. The training data uses statistical data from a large number of unspecified authors, and outputs waveform determining parameters, etc. such that the FM sound source circuit 17 generates a musical tone according to the image specified by the tone image specification dial 16. is learned to do. This greatly reduces the burden of sound creation.

In the above embodiment, the learned synaptic weight stored in the RAM 13 may be a plurality of cents. Further, the musical tone data to be learned may not be the waveform determining parameters set in this electronic musical instrument, but may be data input from another device.

(g) Effects of the invention As described above, the musical tone generating device of this invention can generate musical tones using a neural network, so it can easily generate musical tones that match the image without manipulating complicated parameters. can do. This makes it possible, for example, to easily edit the timbre of the FM sound source circuit.

[Brief explanation of drawings]

FIG. 1 is a block diagram of a control section of an electronic musical instrument according to an embodiment of the present invention, and FIG. 2 is a diagram showing the configuration of a neural network used in the control section. 14-neural network, 16a-16d-tone image setting dial.

Claims (1)

[Claims] (1) In a musical tone generator equipped with a tone generator circuit that generates operators based on multiple waveform determining parameters and synthesizes the generated operators using a predetermined algorithm to generate one musical tone, the characteristics of the timbre are means for inputting a plurality of image parameters to be expressed; a neural network that infers and outputs the waveform determination parameter and the algorithm based on the input plurality of image parameters; and a neural network that infers and outputs the waveform determination parameter and algorithm that the neural network infers and outputs. A musical tone generating device comprising: means for setting in the tone generator circuit.