JP2576703B2 - Musical sound wave generator - Google Patents

Description

DETAILED DESCRIPTION OF THE INVENTION

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a musical tone waveform generator for generating a musical tone waveform which is used as a reference for musical tones produced by electronic musical instruments and the like, and more particularly to a musical tone waveform generator using a neural network.

[0002]

2. Description of the Related Art Conventionally, as a method of generating a waveform amplitude value at each phase angle point of a musical sound waveform as a musical sound waveform generator of an electronic musical instrument or the like, the musical sound waveform data is stored in a memory or the like in advance. Is read out in accordance with phase angle data. Among the waveform memory reading methods, there are a method of storing one cycle of tone waveform data and a method of storing a plurality of cycles.

[0003]

In the method of storing only one cycle, since the waveform from the start to the end of sound generation is fixed, it is difficult to express a temporal change in timbre and the like. Is the degree of deformation due to the filter etc.,
Its expressiveness is limited. On the other hand, in order to express a temporal change of a tone color or the like with high quality, there is a type that stores all waveforms from the start of sound generation to the end (from attack to decay), and reads out the entire waveform in accordance with a key depression. This requires an enormous amount of waveform storage capacity, and it is not easy to create waveform data, requiring a large amount of time, and very expensive in terms of cost.

Therefore, as a device capable of easily creating data while suppressing an increase in the storage capacity, a sound source including a memory in which a required plural-cycle waveform is efficiently stored is currently being commercialized. This sound source stores a plurality of cycles as the waveform of the rising part (attack part) and one cycle as the waveform of the continuation part thereafter, reads out the waveforms of the plurality of cycles in the rising part, and then reads the waveform in the continuation part. It reads the waveform of one cycle repeatedly, and can compress and store the recorded original sound without deteriorating its sound quality, and can faithfully reproduce musical sounds with high reality and expressive sound quality. It is a sound source with excellent features
-1009090).

[0005] However, depending on the type of musical sound produced by the electronic musical instrument, the tone color at the start of sound production is not so large, as in the case of strings, and the tone color mainly changes at the time of aftertouch. When such a musical tone is generated by the above sound source (a waveform of one cycle is repeatedly read out by the sustaining portion), it is necessary to perform a filtering process on the musical tone waveform at the time of after touch to express a temporal change in tone and the like. , Its expressiveness itself was limited.

SUMMARY OF THE INVENTION The present invention has been made in view of the above points, and it is possible to suppress an increase in storage capacity, easily create waveform data, and change the timbre and the like over the entire waveform from the start to the end of sounding. It is an object of the present invention to provide a musical sound waveform generator capable of expressing a sound waveform.

[0007]

According to a first aspect of the present invention, there is provided a musical tone waveform generating apparatus for generating performance information including at least note-on information for instructing generation of a musical tone and pitch information for instructing a pitch of the musical tone. It is constituted by a performance information input means for inputting, and a hierarchical neural network for forming a waveform signal according to the input parameter, and inputs pitch information input from the performance information input means as one of the input parameters. The waveform signal corresponding to the pitch information can be processed in real time in response to the input of the note-on information.
In the waveform forming means for forming, entering the note-on information
In response to the force, it is characterized in further comprising a tone waveform output means for outputting the waveform signal formed by said waveform formation means at pitch corresponding to the pitch information.

[0008] A musical sound waveform generator according to a second invention is characterized in that:
A note-on information indicating the occurrence of the musical tone, the tone Nitsu
Information input means for inputting performance information including at least the above touch information, and a hierarchical neural network for forming a waveform signal according to the input parameter, and the touch information input from the performance information input means. Is input as one of the input parameters, and a waveform signal corresponding to the touch information is input to the note-on.
In real time in response to an input of down information comprises a waveform forming means for forming, in response to an input of the note-on information, the musical tone waveform output means for outputting the waveform signal formed by said waveform formation means It is characterized by the following.

[0009] A musical sound waveform generator according to a third aspect of the present invention comprises:
Performance information input means for inputting performance information including at least note-on information for instructing generation of a musical tone, control information generation means for generating control information that changes in a time series according to the note-on information, A waveform signal according to the control information is constituted by a hierarchical neural network forming a waveform signal according to the parameter, and inputting control information generated from the control information generating means as one of the input parameters , The note-on information
In real time in response to the input of
A waveform forming means for forming the waveform by changing the waveform in time series ,
Wherein in response to the input of note-on information, is characterized in further comprising a tone waveform output means for outputting the waveform signal formed by said waveform formation means.

[0010]

The neural network is an engineering model of a neural network, and is a technology that is being researched and developed as an associative memory in the field of pattern recognition. Then, the hierarchical neural network outputs a signal corresponding to the input parameter and the neural network coefficient (the degree of synapse connection).

In the first invention, pitch information indicating the pitch of a musical tone output from a keyboard or the like of an electronic musical instrument is input to the waveform forming means as one of the input parameters. Waveform formation means is composed of a hierarchical neural network which forms a waveform signal in response to an input parameter, wherein the performance information input means
Pitch information input from the
The waveform signal corresponding to the pitch information
Issue in real time in response to input of note-on information
Then, form. Therefore, the waveform forming means uses the pitch information as an address, and functions as a waveform memory for forming a waveform signal corresponding to the address.

Since a normal waveform memory stores waveform signals in one-to-one correspondence with addresses, it is necessary to store in advance the number of waveform signals required for sound generation.
Although its storage capacity is enormous, the waveform forming means composed of a hierarchical neural network has an error between a waveform signal formed by inputting input parameters and a tone-specific normalized waveform signal extracted from a tone. By simply storing in advance the neural network coefficients learned so as to minimize, the waveform signal corresponding to the input of any pitch information (address) can be formed in real time . The musical sound waveform output means responds to note-on information instructing generation of a musical tone inputted through the performance information input means.
Then , the waveform signal formed by the waveform forming means is sequentially output with phase data corresponding to the pitch information, and a tone waveform signal having a desired pitch is output. As described above, according to the first invention, it is not necessary to previously store an enormous number of waveform signals required for sound generation as in the related art, and it is possible to suppress an increase in storage capacity, By inputting pitch information as one of the input parameters to the hierarchical neural network, any waveform signal can be easily formed. In addition,
By inputting level information indicating a desired envelope together with the pitch information to the waveform forming means as an input parameter, a temporal change of a timbre or the like is expressed over the entire waveform from the start to the end of sound generation according to the level information. be able to.

In the second invention, touch information output from a keyboard or the like of an electronic musical instrument is input to the waveform forming means as one of input parameters. The touch information is input to the waveform forming means via the performance information input means. Accordingly, the waveform forming means uses the touch information as an address, and functions as a waveform memory for forming a waveform signal corresponding to the address. The waveform signal corresponding to any touch information (address) is input to the note-on information. Realta responded
Im, can be formed. Then, based on the note-on information, the tone waveform output means sequentially outputs the waveform signal formed by the waveform forming means with predetermined phase data, and outputs a desired tone waveform signal. As described above, according to the second aspect, an increase in storage capacity can be suppressed, and all waveform signals can be easily formed by inputting touch information to the hierarchical neural network as one of input parameters. In addition to this, it is possible to express the temporal change of the timbre and the like with high quality over the entire waveform from the start to the end of the sound generation according to the change of the touch information.

In the third invention, note-on information for instructing generation of a musical tone is output from a keyboard or the like of the electronic musical instrument. Therefore, control for generating control information that changes in a time series according to the note-on information is performed. An information generating means is provided, and this control information is input to the waveform forming means as one of the input parameters. Therefore, the waveform forming means uses the control information as an address and functions as a waveform memory for forming a waveform signal corresponding to the address. In response to the control information (address) which changes in time series, the waveform forming means operates in time series. Real-time response of the changed waveform signal to the input of note-on information
, And can be formed in a time-series manner in accordance with the control information . The tone waveform output means sequentially outputs the waveform signal formed by the waveform forming means with predetermined phase data in response to the note-on information, so that the tone waveform signal which changes in time series with the passage of time. Is output. As described above, according to the third aspect, it is possible to suppress an increase in storage capacity, and to input control information that changes in time series as one of the input parameters to the hierarchical neural network, thereby starting sound generation. It is possible to express the temporal change of the timbre or the like with high quality over the entire waveform from the end to the end.

[0015]

DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS Embodiments of the present invention will be described below in detail with reference to the accompanying drawings. FIG. 1A is a diagram showing an overall configuration when a neural network learns a waveform signal based on a tone signal. In FIG. 1A, an analysis unit 1 inputs a voice signal or a musical instrument sound via a microphone or the like,
Sampling it and analog / digital (A / D)
The data is converted into a digital signal by a converter, and based on the digital signal, pitch (pitch) data P, a differential value of pitch ΔP,
The amplitude (level) data L and the differential value ΔL of the level are extracted and output to the input layer neurons I0 to I3 of the hierarchical neural network 2. At this time, the analysis means 1 also extracts a normalized waveform signal (teacher signal) RS for learning and outputs it to the hierarchical neural network 2.

The analysis means 1 includes an autocorrelation method, a zero crossing method, a fast Fourier transform method, and a linear predictive analysis method (Linear predictive analysis method).
ar Predictive Coding: LP
C), Line Spectrum Pair Analysis (Line Spectr)
um Pair: LSP), Composite Sinusoidal modal analysis
el: CSM) or the like using a technique well-known in the field of sound / speech analysis.

The hierarchical neural network 2 is composed of a software part mainly composed of a microprocessor unit (MPU) and a hardware part exclusively designed for the neural network in order to increase the processing speed. That is, the hierarchical neural network 2 is an RO that stores a microprocessor unit, a system program, and the like.
M, a microcomputer system comprising a bus connection such as a working RAM and data for storing synapse coupling degrees Vjk, Wij and other data, and a logic circuit for high-speed operation added thereto. I have. In FIG. 1A, the synaptic connectivity Vjk,
In particular, the memory area 3 of the data and working RAM is shown as storing Wij. This synaptic connection degree V
jk and Wij are important neural net coefficients for characterizing the waveform signal. The method of calculating the neural network coefficient will be described later.

FIG. 1 (b) is a diagram showing an entire configuration in the case of creating a musical tone waveform using the hierarchical neural network 2 for which learning of a waveform signal has been completed. Pitch data P for specifying a musical sound waveform to be pronounced, differential value ΔP of pitch,
The level data L and the differential value ΔL of the level are input neurons I0 to I3 of the hierarchical neural network 2 respectively.
And the degree of synapse connection Vjk, W in the memory area 3
The signal corresponding to ij is 1024 output layer neurons Oi
(O0 to O1023). Then, the signals of the output layer neurons Oi corresponding to the phase data i are sequentially read out, and a specific tone waveform signal corresponding to the waveform at the time of learning is output from the hierarchical neural network 2.

FIG. 2 is a diagram showing the concept of a hierarchical neural network 2 composed of a microcomputer system and a logic circuit. In FIG. 2, the hierarchical neural network 2 has four input layer neurons Ik (I0 to I3).
And eight intermediate (hidden (hidden) layers) neurons Hj
(H0 to H7) and 1024 output layer neurons Oi
(O0 to O1023).

The input layer neuron I0 has pitch data P
The input layer neuron I1 has a pitch differential value ΔP,
The level data L is input to the input layer neuron I2, and the level differential value ΔL is input to the input layer neuron I3 from the analysis means 1. The input layer neurons I0 to I3 are connected to the intermediate layer neurons H0 to H7 via synapses of the degree of connection Vjk, respectively. Hereinafter, the term "Vjk" is used to mean the synaptic connection between the k-th input layer neuron Ik and the j-th intermediate layer neuron Hj and the degree of the connection. Therefore, the input layer neuron I0 and the intermediate layer neuron H0 are connected by the synaptic connection V00, and similarly, the input layer neuron I1
Are connected to the intermediate layer neuron H0 via the synaptic connection V01, the input layer neuron I2 is connected to the synaptic connection V02, and the input layer neuron I3 is connected to the intermediate layer neuron H0 via the synaptic connection V03.

The intermediate layer neurons H0 to H7 are respectively connected to the input layer neurons I0 to I3 via the synapse connection Vjk and to the output layer neurons O0 to O1023 via the synapse connection Wij. , And add the sum to the sigmoid function f
(U) and outputs it to the output layer neurons O0 to O1023 as an intermediate layer signal αj. Hereinafter, the expression “Wij” is used to mean the synaptic connection between the j-th intermediate layer neuron Hj and the i-th output layer neuron Oi and the degree of the connection. The sigmoid function f (u) is given by f (u) = 1 / (1 + exp (-u /
T)). For example, the intermediate layer neuron H0 calculates the product (P · V00) of the pitch data P and the synaptic connectivity V00 and the product (ΔP · V) of the pitch differential value ΔP and the synaptic connectivity V01.
01) and the product of the level data L and the synaptic connectivity V02 (L
V02), the level differential value ΔL, and the synaptic connection degree V03
(ΔL · V03), and the sum of these values (P
.V00 + .DELTA.P.V01 + L.V02 + .DELTA.L.V03) is converted by the sigmoid function f (u) and the intermediate layer signal .alpha.0 is output to each output layer neuron O0 to O1023. The intermediate layer neurons H1 to H7 perform the same operation, and output intermediate layer signals α1 to α7 which are the results to the output layer neurons O0 to O1023.

The output-layer neurons O0 to O1023 are connected to the intermediate-layer neurons H0 to H0 through respective synaptic connections Wij.
H7, the products of the intermediate layer signals .alpha.0 to .alpha.7 from the intermediate layer neurons H0 to H7 and the synapses Wij are respectively added, and the added value is again converted by the sigmoid function f (u) to output the output layer signal Fi. Each phase number i of the waveform signal
Is output as the final waveform amplitude value corresponding to. Therefore, the output layer signal F0 of each output layer neuron O0 to O1023 is obtained.
To F1023 in order according to the phase number i, thereby forming a specific musical tone waveform. Here, the number of output layer neurons is 1024, which is the same as the phase number i of the waveform signal. For example, the output layer neuron O0
Is the product of the intermediate layer signal α0 and the synaptic connection W00 (α0 · W0
0) and the product of the intermediate layer signal α1 and the synaptic connection W01 (α1
W01), the product of the intermediate layer signal α2 and the synaptic connection W02 (α2 · W02), and the intermediate layer signal α3 and the synaptic connection W03
(Α3 · W03), the product of the intermediate layer signal α4 and the synaptic connection W04 (α4 · W04), the product of the intermediate layer signal α5 and the synaptic connection W05 (α5 · W05), and the intermediate layer signal α6 and the synaptic connection The product (α6 · W06) of W06 and the product (α7 · W07) of the intermediate layer signal α7 and the synaptic connection W07 are input. Then, the output layer neuron O0 calculates the sum (α0
・ W00 + α1 ・ W01 + α2 ・ W02 + α3 ・ W03 + α4 ・
W04 + α5 · W05 + α6 · W06 + α7 · W07) is converted by a sigmoid function f (u) to output an output layer signal F0. Similarly, the other output layer neurons O1 to O1023 add the products of the respective intermediate layer signals αj and the synaptic connections Wij, convert the values with a sigmoid function f (u), and output the output layer signals F1 to F1023. .

A hierarchical neural network 2 as shown in FIG.
Can be configured by software using only a microcomputer system, but the waveform signal can be processed at high speed.
Meniwa that is, it is better to implement the respective hardware operation between the operation and the intermediate layer neurons Hj and an output layer neuron Oi between the input layer neurons Ik and hidden neurons Hj. Hereinafter, hierarchical neural network 2
3 and 4 for details of the hardware part constituting
This will be described with reference to FIG.

In FIG. 2, a portion for performing an operation between the input layer neuron Ik and the intermediate layer neuron Hj and analyzing the state of the input data is called state analysis means. A portion that performs a calculation between the two and generates a waveform signal is referred to as a waveform signal generating means. FIG. 3 is a diagram showing a circuit configuration corresponding to one intermediate layer neuron Hj of the state analyzing means 2V, and FIG. 4 is a circuit configuration corresponding to one output layer neuron Oi of the waveform signal generating means 2W. FIG. Therefore, the actual state analyzing means 2V is configured by integrating the circuits of FIG. 3 by the same number (eight) as the intermediate layer neurons Hj, and the waveform signal generating means 2W is formed by the same number (1024) as the output layer neurons Oi. It is configured by integrating the circuit of FIG.

The state analysis means 2V shown in FIG. 3 includes four multipliers Mj0, Mj1, Mj2, Mj3 and three adders Aj1, Aj2,
Aj3 and a sigmoid function converter SIG1. Here, j is the number j of the hidden neuron Hj. The multiplier Mj0 receives the pitch data P, the data and the synaptic connection Vj0 stored in the memory area 3V of the working RAM, and multiplies the two (P × Vj0).
Is output to the adder Aj1. The multiplier Mj1 calculates the pitch differential value Δ
P multiplied by the value of synaptic connection Vj1 (ΔP × Vj
1) is output to the adder Aj1. The multiplier Mj2 multiplies the value of the level data L and the value of the synaptic connection Vj2 (L × Vj
2) is output to the adder Aj2. The multiplier Mj3 multiplies the value of the level differential value ΔL and the value of the synaptic connection Vj3 (ΔL ×
Vj3) is output to the adder Aj3.

The adder Aj1 outputs the value (P × Vj0 + ΔP × Vj1) obtained by adding the outputs of the multipliers Mj0 and Mj1 to the next adder Aj2. The adder Aj2 outputs a value (P × Vj0 + ΔP × Vj1 + L × Vj2) obtained by adding the outputs of the adder Aj1 and the multiplier Mj2 to the next adder Aj3. The adder Aj3 adds the output of the adder Aj2 and the output of the multiplier Mj3 (P × Vj0 + ΔP × Vj1 +
L × Vj2 + ΔL × Vj3) is converted to a sigmoid function converter SIG
Output to 1. Therefore, instead of the adders Aj1, Aj2, Aj3, a device that outputs the total value of the multipliers Mj0, Mj1, Mj2, Mj3 may be used. The sigmoid function converter SIG1 outputs the output value (P × Vj0 + ΔP × Vj1 + L × Vj2) of the adder Aj3.
+ ΔL × Vj3) is converted by a sigmoid function f (u) and output as an intermediate layer signal αj to the output layer neurons O0 to O1023 via the synaptic connection Wij.

The waveform generating means 2W of FIG. 4 comprises eight multipliers mi0, mi1, mi2, mi3, mi4, mi5, mi6, mi7,
Seven adders ai1, ai2, ai3, ai4, ai5, ai6, a
i7 and a sigmoid function converter SIG2. Here, i is the number i of the output layer neuron Oi. The multiplier mi0 receives the intermediate layer signal α0 and the synapse connection Wi0 stored in the memory area 3W of the data and working RAM, and multiplies both (α0 × Wi0).
Is output to the adder ai1. The multiplier mi1 outputs the intermediate layer signal α1
Multiplied by the value of the synapse connection Wi1 and (α1 × Wi1)
Is output to the adder ai1. The multiplier mi2 outputs the intermediate layer signal α2
Multiplied by the value of and the synaptic connection Wi2 (α2 × Wi2)
Is output to the adder ai2. Multipliers mi3, mi4, mi5, m
Similarly, the intermediate layer signals α3, α4, α5, α6,
α7 multiplied by the value of synaptic connections Wi3, Wi4, Wi5, Wi6, Wi7 (α3 × Wi3, α4 × Wi4, α5 × Wi
5, α6 × Wi6, α7 × Wi7) are added to respective adders ai
3, ai4, ai5, ai6, ai7.

The adder ai1 adds the value (α0 × Wi0 + α1 × Wi1) obtained by adding the outputs of the multipliers mi0 and mi1 to the next adder ai2.
Output to The adder ai2 adds the output of the adder ai1 and the output of the multiplier mi2 (α0 × Wi0 + α1 × Wi1 + α2 × Wi
2) is output to the next adder ai3. Similarly, the adders ai3, ai4, ai5, ai6, and ai7 sequentially add the outputs from the multipliers corresponding to the previous-stage adder, and add the final added value (α0
× Wi0 + α1 × Wi1 + α2 × Wi2 + α3 × Wi3 + α4 ×
The adder ai7 outputs Wi4 + α5 × Wi5 + α6 × Wi6 + α7 × Wi7) to the sigmoid function converter SIG2. Therefore, the adders ai1, ai2, ai3, ai4, ai5, ai6, ai7
Instead of the multipliers mi1, mi2, mi3, mi4, mi5, mi
A device that outputs the total value of 6, mi7 may be used. The sigmoid function converter SIG2 outputs the output value of the adder ai7 (α0
× Wi0 + α1 × Wi1 + α2 × Wi2 + α3 × Wi3 + α4 ×
Wi4 + α5 × Wi5 + α6 × Wi6 + α7 × Wi7) is converted by a sigmoid function f (u) and output as an output layer signal Fi of phase number i. Note that the sigmoid function converter SIG2 may be omitted.

Next, the processing of analyzing the neural network coefficients (the synaptic coupling degrees Vjk, Wij) will be described in the order of steps based on the flowchart of FIG. Step 50: Neural network coefficients (synaptic connection V
jk, Wij) are set as initial values so that the output waveform from the output layer neuron Oi becomes substantially a sine wave. Step 51: A musical sound waveform such as an audio signal or a musical instrument sound is input to the analyzing means 1 via a microphone or the like. The tone waveform at this time includes initial touch data, after touch data, and the like.

Steps 52 to 55 are processes performed by the analyzing means 1, and the pitch data P, P
Pitch differential value ΔP, level data L, level differential value ΔL
This is a process for extracting musical tone data relating to musical tones and the like and a normalized waveform signal RS unique to the musical tones.
58 is a process performed by the hierarchical neural network 2,
This is a process of learning and storing neural network coefficients (synapse coupling degrees Vjk, Wij). Step 52: The analysis means 1 samples the input musical sound waveform, converts it into a digital signal by an analog / digital (A / D) converter, extracts an amplitude envelope signal based on the digital signal, and The level data L and the level differential value ΔL are calculated based on the amplitude envelope signal, and output to the hierarchical neural network 2.

Step 53: The level of the amplitude envelope signal of the digital signal extracted in the previous step is equalized so as to be constant over the entire waveform. Step 54: The analyzing means 1 extracts a pitch envelope signal indicating a change in pitch (pitch) based on the amplitude envelope signal uniformized in the previous step, and
Based on the pitch envelope signal, the pitch data P
And its differential value ΔP are calculated and output to the hierarchical neural network 2.

Step 55: The analyzing means 1 equalizes the level of the pitch envelope signal extracted in the previous step so as to be constant over the entire waveform. The pitch envelope signal having a uniform pitch level obtained in this manner is output to the hierarchical neural network 2 as a normalized waveform (teacher waveform) signal RS having a uniform amplitude level and pitch level. The thus-obtained normalized waveform (teacher waveform) signal RS is unique to a musical tone. Step 56: Based on the pitch data P, the pitch differential value ΔP, the level data L, the level differential value ΔL, and the normalized waveform signal RS from the analysis means 1, the hierarchical neural network 2 generates a neural network coefficient (the synaptic coupling degree Vjk, W
ij) is learned by the back propagation method.

That is, in the hierarchical neural network shown in FIG. 2, pitch data P, pitch differential value ΔP, level data L, and level differential value ΔL are input to input layer neurons I0 to I3, respectively. Since initial values are previously set in the neural network coefficients (the degree of synaptic connection Vjk, Wij) in the processing of step 50, the output layer neuron O0 is set.
From 1024 to 1024 according to the neural network coefficient
Output layer signals F0 to F1023 are output. This 102
The four output layer signals Fi are sine wave signals. Therefore, the microcomputer system constituting the hierarchical neural network 2 uses the normalized waveform signal RS
Is converted to 1024 normalized waveform signals RS0 to RS1023 corresponding to the number of output layer signals, and a neural network coefficient (the degree of synaptic coupling Vjk, Wij) that minimizes the shift between the output layer signals F0 to F1023. Is determined by a known back propagation method.

Step 57: It is determined whether the processing of steps 51 to 56 has been performed N times or more. If the processing has been performed N times or more (YES), the process proceeds to the next step 58, and if not, steps 51 to 56 have been performed. Is repeatedly executed. Step 58: As a result of executing the processing of steps 51 to 56 N times, it is determined whether or not the neural network coefficient gradually converges as the number of executions increases. Are learned and stored, the learning process for this musical tone waveform is terminated, and if not converged, the processes of steps 51 to 56 are executed a predetermined number of times again. The number of times at this time may be the same as N in step 57, may be more or less than this, and may be repeatedly executed until it is determined in step 58 that convergence is achieved. Is also good.

Next, a description will be given of an embodiment of an electronic musical instrument using the hierarchical neural network 2 for which the learning process has been completed as musical tone waveform generating means. FIG. 6 uses the hierarchical neural network 2 of FIG. 1 as musical tone waveform generating means. In the embodiment of FIG. 6, the keyboard 10 is provided with a plurality of keys for selecting the pitch of a musical tone to be produced, and has a key switch corresponding to each key. And a touch detecting means such as a pressing force detecting device.

The key press detecting means 11 detects the state of the keyboard 10 (whether or not there is a key pressed), that is, the on / off of each key switch in the keyboard 10, and for example, turns each key switch in order. And a circuit that encodes the scanning result,
It outputs a key code KC indicating a depressed key, a note-on signal Non indicating that the key has been pressed, and a note-off signal Nof (not shown) indicating that the key has been released. Also,
The key press detection means 11 determines the key press operation speed at the time of depression based on the output from the key switch, outputs the initial touch data IT, and relates to each key of the keyboard, the pressing force when the key press is continued. The pressing force is detected from the output of the detecting device and the after touch data AT is output.

The key press detecting means 11 outputs the note-on signal No.
n, the initial touch data IT and the after touch data AT are output to the pitch envelope signal generating means 12 and the amplitude envelope signal generating means 14, and the key code K
C is output to the adder 13. The pitch envelope signal generating means 12 outputs to the adder 13 a pitch envelope signal PE indicating the amount of deviation of the key code KC from the reference pitch based on the note-on signal Non, the initial touch data IT and the after touch data AT.

The adder 13 outputs pitch data P obtained by changing the key code KC by the pitch envelope signal PE to the musical tone waveform generator 2, the differentiator 15, and the pitch number generator 17. Amplitude envelope signal generating means 14
Are the note-on signal Non and the initial touch data I
Based on T and the after touch data AT, an amplitude envelope signal (level data) L from the start of tone generation to the end of tone generation is output to the multiplier 19, the differentiator 16 and the musical tone waveform generating means 2, and the initial touch data IT
And / or variably control the amplitude envelope signal according to the after touch data AT, or generate a different amplitude envelope signal depending on the tone color of a musical tone.

The differentiator 15 receives the pitch data P from the adder 13 and outputs a pitch differential value ΔP obtained by differentiating the pitch data P to the musical tone waveform generating means 2. The differentiator 16 receives the amplitude envelope signal (level data) L from the amplitude envelope signal generator 14 and outputs a level differential value ΔL obtained by differentiating the signal to the musical sound waveform generator 2. The pitch number generating means 17 receives the pitch data P from the adder 13 and outputs a corresponding pitch number Pn to the accumulator 18.

The accumulator 18 accumulates the pitch number Pn and generates phase data i for setting the pitch of the musical tone waveform signal Fi output from the musical tone waveform generating means 2. Therefore, when the value of the pitch number Pn is small, the speed of accumulation becomes slow, the amount of increase of the phase data i becomes small, and the pitch of the musical tone waveform signal Fi also becomes low. Conversely, when the pitch number Pn is large, the speed of accumulation increases, the amount of increase in the phase data i increases, and the pitch of the musical tone waveform signal Fi also increases. When the accumulated value overflows, it returns to the initial value again and repeats the accumulation.

The tone waveform generating means 2 includes pitch data P, pitch differential value ΔP, level data L, and level differential value ΔL.
Into each of the input layer neurons I0 to I3,
A tone waveform signal corresponding to a neural network coefficient (synaptic coupling degree Vjk, Wij) is output to each output layer neuron O0 to O1023.
Output from At this time, from the musical tone waveform generating means 2,
A tone waveform signal Fi corresponding to the phase data i of the accumulator 18 is finally output.

The multiplier 19 multiplies the tone waveform signal Fi output from the tone waveform generating means 2 by the amplitude envelope signal L from the amplitude envelope signal generating means 14, and uses the result of the multiplication as a tone signal as a D / A converter. 20. The digital tone signal output from the multiplier 19 is converted into an analog tone signal by a digital / analog (D / A) converter 20 and output to the sound system 21.

The sound system 21 is composed of a speaker, an amplifier and the like, and generates a tone according to an analog tone signal from the D / A converter 20. Musical sound wave generating means 2
Has described the case where a waveform for one cycle is output as musical sound waveform data, but the present invention is not limited to this, and a waveform for a half cycle or a quarter cycle may be output. The number of output layer neurons is also 1024, but is not limited to this.

Next, another embodiment of the present invention will be described with reference to FIG.
This will be described with reference to FIG. 7, the same components as those in FIG. 6 are denoted by the same reference numerals, and the description thereof will be omitted. The embodiment of FIG. 7 is different from that of FIG. 6 in that initial touch data IT, after touch data AT,
Differential value ΔAT of after touch data and frame number F
N is input and a tone waveform signal Fi is generated based on these input data.

The musical sound waveform generating means 4 inputs the initial touch data IT, the after touch data AT, the differential value .DELTA.AT of the after touch data and the frame number FN to the input layer neurons I0 to I3 during learning, and responds to these input data. Neural network coefficients (synaptic coupling degree Vjk,
Wij) is stored in the memory area of the data and working RAM. Therefore, the musical tone waveform generating means 4 differs from the musical tone waveform generating means 2 of FIG. 6 only in the value of the neural network coefficient (the degree of synaptic connection Vjk, Wij), and there is no difference in hardware.

Therefore, in the embodiment of FIG. 7, the circuit configuration for providing input data (tone data) to the tone waveform generating means 4 is as follows. First, the musical sound waveform generating means 4
Time counter 22 to supply the frame number FN to the
And a frame designating circuit 23. The supply of the initial touch data IT and the after touch data AT is performed directly from the key press detection means 11, and the supply of the differential value ΔAT of the after touch data is performed via the differentiator 24.

The time counter 22 outputs a note-on signal Non
The counting process is started in response to the input of the, and the increment amount to be counted is changed in accordance with the initial touch data IT and the after touch data AT, thereby variably controlling the counting process. The frame designating circuit 23 receives the count value from the time counter 22 and outputs a frame number FN corresponding to the count value to the musical tone waveform generating means 4.

Here, the frame number FN is learned by dividing all the waveforms of the musical tone from the start of sound generation to the end (from attack to decay) by dividing them into groups at intervals of a plurality of periods, and assigning a number to the group in order from the start of sound generation. That is. For example, if the tone waveform signals F0 to F1023 for one cycle output first from the tone waveform generating means 4 immediately after the start of sound generation are set as the first waveform signal FW1, the tone waveform generating means 4 sets the first frame number FN = 1. To change the waveform signal FW1 to the cycle number 0.
In the next frame number FN = 2, the second waveform signal FW2 having a slightly different waveform is provided for six cycles of the cycle numbers 4 to 9, and in the frame number FN = 3, the third waveform signal FW3 is provided for the cycle number 10 in the next frame number FN = 2. Frame number FN = 11 periods of ~ 20
At 4, the fourth waveform signal FW4 is set to 20 of the cycle numbers 21 to 40.
In the frame number FN = 5, the fifth waveform signal FW5
For 30 cycles of cycle numbers 41 to 70 and frame number FN
When = 6, the sixth waveform signal FW6 is learned and stored in association with the frame number FN and the cycle number, such as 50 cycles of the cycle numbers 71 to 120.

When a tone waveform is generated, the time counter 22 outputs a count value incremented according to the values of the initial touch data IT and the after touch data AT to the frame designating circuit 23. The frame numbers FN are no longer output in numerical order, and the tone color of the tone waveform output from the tone waveform generating means 4 effectively changes according to the values of the initial touch data IT and the after touch data AT. That is, when the increment amount increases, the second waveform signal FW2 is output as the fourth waveform of the cycle number without outputting the first waveform signal FW1. Conversely, when the increment amount is small, the second waveform signal FW2 is not output as the fifth waveform of the cycle number and the first waveform is not output.
The waveform signal FW1 is output.

In the embodiment of FIG. 6, the amplitude envelope signal L of the amplitude envelope signal generating means 14 is differentiator 1
6 and the musical sound waveform generating means 2, FIG.
In this embodiment, the amplitude envelope signal L is output only to the multiplier 19. Thus, according to the embodiment of FIG.
The tone color of the musical tone waveform output from the musical tone waveform generating means 4 can be effectively changed according to the values of the initial touch data IT and the after touch data AT.

In the embodiments shown in FIGS. 6 and 7, the device is constituted by a dedicated hardware circuit, but the same function may be realized by software processing. In the above embodiment, a case where a single sound is generated has been described. However, a sound generation process may be performed on a plurality of channels by time division processing, parallel processing, or the like. In the above embodiment, the electronic musical instrument for designating a desired musical tone by a keyboard has been described. However, the present invention is not limited to this, and the performance operating means for specifying the pitch may be other than the keyboard. In addition, the present invention can be similarly applied to a sound source module unit without a keyboard, and the like.
In such a case, it is only necessary to be able to input performance information such as pitch information, note-on information, and touch information. In the present invention, the key is not limited to a key on a keyboard, but also includes other musical sound designation operation means. In order to control the characteristic of the temporal change of the waveform, for example, a process such as passing a waveform output through a comb filter using a 1024-stage delay circuit may be performed.

Further, in this embodiment, pitch data P, pitch differential value ΔP, level data L, level differential value ΔL,
Initial touch data IT, after touch data A
T, the after-touch data differential value ΔAT, and the frame number FN have been described as examples. However, the present invention is not limited thereto, and the position data of sound image localization, the brilliance parameter data,
Alternatively, various pedal data or the like may be used. Needless to say, each input layer neuron I and intermediate layer neuron of the hierarchical neural network can be realized. Further, a new tone waveform signal can be created by appropriately variably controlling the neural network coefficient in accordance with the operation amount of the manipulator, tone data, and the like.

[0053]

As described above, according to the musical tone waveform generator of the present invention, an increase in storage capacity can be suppressed, and even a tone waveform signal that did not exist at the time of storage can be easily created. This has an excellent effect that it is possible to express the temporal change of the timbre and the like over the entire waveform from the start to the end of the sound generation.

[Brief description of the drawings]

FIG. 1A shows an entire configuration in a case where a neural network learns a waveform signal based on a musical tone signal;
FIG. 3B is a diagram showing an entire configuration in the case of creating a musical tone waveform using a hierarchical neural network for which learning of a waveform signal has been completed.

FIG. 2 is a diagram showing the concept of the hierarchical neural network of FIG.

FIG. 3 is a diagram showing a circuit configuration corresponding to one hidden neuron of the state analysis means of FIG. 2;

FIG. 4 is a diagram showing a circuit configuration corresponding to one output layer neuron of the waveform signal generating means.

FIG. 5 shows a neural network coefficient (synaptic coupling degree Vj).
FIG. 9 is a flowchart for explaining the analysis process of (k, Wij).

FIG. 6 is a diagram showing an embodiment of an electronic musical instrument configured using the musical tone waveform generator of the present invention.

FIG. 7 is a diagram showing another embodiment of an electronic musical instrument configured using the musical tone waveform generating device of the present invention.

[Explanation of symbols]

1 ... analysis means, 2, 4 ... hierarchical neural network, 3 ...
Memory area, 10 ... keyboard, 11 ... key press detection means, 12 ...
Pitch envelope signal generating means, 13 ... adder, 14
... amplitude envelope signal generating means, 15, 16, 24 ...
Differentiator, 17 pitch number generating means, 18 accumulator, 19 multiplier, 20 D / A converter, 21 sound system, 22 time counter, 23 frame designation circuit, I input neuron, H ... Intermediate layer neuron, O ... Output layer neuron, Wij, Vjk ... Synaptic coupling degree, Aj1-Aj3, ai1-ai7 ... Adder, Mj0-Aj3, m
i0-ai7 ... Multiplier

Claims (4)

(57) [Claims] 1. Performance information input means for inputting performance information including at least note-on information instructing generation of a musical tone and pitch information instructing a pitch of the musical tone, and a waveform corresponding to an input parameter. A pitch signal input from the performance information input means as one of the input parameters, thereby forming a waveform signal corresponding to the pitch information into the note-on signal. Real tie responding to input of information
In arm, a waveform forming means for forming, the note-on information in response to the input of tone waveform output means for outputting the waveform signal formed by said waveform formation means at pitch corresponding to the pitch data tone waveform generation apparatus comprising and. 2. Performance information input means for inputting performance information including at least note-on information for instructing generation of a musical tone and touch information on the musical tone , and a hierarchy for forming a waveform signal corresponding to the input parameter. A touch-sensitive information input from the performance information input means as one of the input parameters, thereby generating a waveform signal corresponding to the touch information in response to the input of the note-on information.
In time, the waveform forming means for forming, in response to an input of the note-on information, the musical sound waveform generator and a musical tone waveform output means for outputting the waveform signal formed by said waveform formation means. 3. Performance information input means for inputting performance information including at least note-on information for instructing generation of a musical tone, and control information for generating control information which changes in a time series according to the note-on information. Generating means, and a hierarchical neural network for forming a waveform signal corresponding to the input parameter, wherein the control information generated by the control information generating means is input as one of the input parameters, thereby responding to the control information. Real-time response to the input of the note-on information.
And in a time-series manner according to the control information ,
A waveform formation means for forming, in response to an input of the note-on information, the musical sound waveform generator and a musical tone waveform output means for outputting the waveform signal formed by said waveform formation means. 4. The neural network according to claim 1, wherein said waveform forming means learns a neural signal so as to minimize an error between a waveform signal formed by inputting said input parameter and a tone-specific normalized waveform signal extracted from said tone. 4. The musical tone waveform generator according to claim 1, wherein a net coefficient is stored.