JPH06266398A - Arithmetic unit using neural network - Google Patents

Abstract

PURPOSE:To enable plural neural networks to be realized by the same program by performing arithmetic for an optional neural network by a neural network arithmetic part according to a selection. CONSTITUTION:A neural network structure storage part 100 sends neural network structure information to the neural network arithmetic part 40. The neural network arithmetic part 40, on receiving the information, determines the structure of the neural network and becomes to be in a state waiting data from a pattern generation part 20. Then, when the pattern generation part 20 generates a speech pattern, the arithmetic of the neural network is performed on the basis of the speech pattern and inter-unit coupling information and a coupling coefficient designated by the neural network structure storage part 100 by a program written in the ROM of the neural network arithmetic part 40.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to an arithmetic unit using a neural network, and particularly to recognition for recognizing a voice pattern or the like by using a characteristic parameter which is a voice analysis result as input information to the neural network. Regarding the device.

[0002]

2. Description of the Related Art In a speech recognition apparatus, it is important for practical use of speech recognition to realize processing on a computer that is similar to the judgment method performed by everyday humans. A neural network realized on a computer as a simple model is widely used.

As a method of calculating the neural network, there are a method using a dedicated LSI and a method using a general-purpose computer. However, since the dedicated LSI leads to an increase in cost, it is possible to program the general-purpose computer to realize the calculation. It is common.

FIG. 3 is a schematic configuration diagram of a voice recognition device using a neural network. In FIG. 3, 11 is a voice analysis unit for analyzing the input voice, and 12 is a feature of the voice analyzed by the voice analysis unit 11. This is a section detection unit that detects a voice section based on a parameter, and the voice analysis section 11 and the section detection section 12 configure a characteristic parameter extraction section 10. Reference numeral 20 is a pattern creation unit that creates a pattern from the feature parameters of the voice section extracted by the feature parameter extraction unit 10, 31 is an inter-unit connection information storage unit in which connection information between units of the neural network is digitized and stored, Reference numeral 32 denotes a coupling coefficient storage unit that stores coupling coefficients between units of the neural network,
This inter-unit coupling information storage unit 31 and coupling coefficient storage unit 3
When 2 is recognized, it is stored in the ROM 30. Reference numeral 40 denotes a neural network for performing a neural network operation from the pattern created by the pattern creating unit 20, the inter-unit coupling information stored in the inter-unit coupling information storage unit 31, and the coupling coefficient stored in the coupling coefficient storage unit 32. The calculation unit 50 is a recognition determination unit that determines recognition based on the calculation result of the neural network calculation unit.

FIG. 4 is a schematic block diagram of a neural network learning device for learning the neural network of the speech recognition device shown in FIG. 3, which has the same function as the speech recognition device using the neural network in FIG. Are denoted by the same reference numerals, and description thereof will be omitted. Reference numeral 60 denotes a learning pattern storage unit that receives characteristic parameters from the pattern creation unit 20 and stores a voice pattern, and 70
Is a teacher signal generation that generates, as a teacher signal, the number of the corresponding category (here, the category is a vocabulary to be recognized) of the learning pattern sent from the learning pattern storage unit 60 to the neural network operation unit 40. Reference numeral 80 denotes an error calculation unit that calculates an error based on the calculation result of the neural network calculation unit 40 and the teacher signal of the teacher signal generation unit 70. Reference numeral 90 denotes the combination of the combination coefficient storage unit 32 based on the calculation result of the error calculation unit 80. It is a learning calculation unit that changes the coefficient.

FIG. 5 is a schematic configuration diagram of a hierarchical neural network. 41a is an input layer, 42a is an intermediate layer, and 4a.
An output layer 3a is composed of I, J, and K units, respectively. As shown in the drawing, the respective units forming the layers adjacent to each other in the vertical direction are connected by the information transmission path based on the coupling information in the inter-unit coupling information storage unit 31. The coupling coefficient of the information transmission path is stored in the coupling coefficient storage unit 32.

Here, the number of units in each layer will be specifically described. For example, the one-digit number voice "Rei", "Ichi",
"Ni", "san", "shi", "go", "roku", "shichi", "hachi", "ku" and their replacements "zero", "maru", "yon" , "Nana", "Kyu"
Take a total of 15 words as an example.

Since the frequency of the input voice is divided into 16 and the time is divided into 8, the number of units in the input layer 41a is I = 16 × 8 = 128.

Since the number of categories to be recognized is 15, the number of units K of the output layer 43a is K = 15, and the number of units J of the intermediate layer 42a is 50.

The accumulation of learning patterns in the learning pattern accumulating unit 60 and the learning of the neural network in the above-mentioned configuration will be described. The one-digit number voice “Rei” is input to the voice analysis unit 11. Speech analysis unit 11
Then, after the voice "Rei" is A / D converted by an A / D converter (not explicitly shown), for example, 100 to 600
The frequency band of 0 Hz is divided into 16, and the size in each frequency band, that is, 16 frequency components are extracted every 5 milliseconds, and power information and the like are extracted as necessary. These pieces of information are used as characteristic parameters.

In the section detection unit 12, based on the characteristic parameters extracted by the voice analysis unit 11, the input voice "Rei"
The voice section is determined by determining the start time and end time of. The feature parameter of the voice section determined by the section determining unit 12 is divided into eight by the pattern creating unit 20, and the average of the frequency components extracted every 5 milliseconds in each period is averaged for each frequency component. calculate. That is, a voice pattern consisting of 16 × 8 data is created for one input voice.

The voice pattern created by the pattern creating section 20 is sent to the learning pattern accumulating section 60 and is accumulated as a learning pattern of the category corresponding to "REI".

Thereafter, similar to the above, "ichi", "ni", ...
..., 14 learning patterns of "kyu" are stored in the learning pattern storage unit 60 for each category.

When the learning patterns are accumulated in the learning pattern accumulating unit 60, one learning pattern belonging to the category "Rei" is inputted from the learning pattern accumulating unit 60 into the input layer 41a of the neural network operation unit 40. At the same time, the number of the category c to which the numerical voice “Rei” belongs from the learning pattern storage unit 60 (where 1 ≦ c ≦ K, c is an integer,
In the above example, K = 15. ) Is the teacher signal generator 7
Sent to 0.

In the neural network operation unit 40,
Using the network shape stored in the unit-to-unit coupling information storage unit 31 and the inter-unit coupling coefficient stored in the coupling coefficient storage unit 32 in association with this, the learning pattern input to the input layer 41a is set. On the other hand, a calculation is performed inside the neural network calculation unit 40, and the result is output from the K units of the output layer 43a as output values Ok (k = k).
, 1, ..., K) (hereinafter referred to as Ok) is output.

In the teacher signal generator 70, a teacher signal Tk (k = 1, 2, ..., K) (hereinafter referred to as Tk) corresponding to the category c sent from the learning pattern accumulator 60. )
Is generated and sent to the error calculator 80.

On the other hand, in the error calculator 80, the error between the teacher signal Tk of the teacher signal generator 70 and the output value Ok of the learning neural network calculator 60, that is, Ek = Tk-Ok (k = 1, 2, ····, K) ··· (1) is calculated.

The learning calculation unit 90 refers to the network shape stored in the unit-to-unit coupling information storage unit 31 so that the error Ek shown in the equation (1) is minimized.
The inter-unit coupling coefficient stored in the coupling coefficient storage unit 32 is changed based on the error back propagation method.

By repeating this operation, the inter-unit coupling coefficient stored in the coupling coefficient storage unit 32 is gradually changed to approach the optimum solution.

In practice, it is difficult to find the optimum solution, so that the quasi-optimal solution is sought by sufficiently repeating the learning, and this quasi-optimal solution has no problem in practice.

By the above operation, the coupling coefficient is determined in the coupling coefficient storage unit 32. This coupling coefficient and the inter-unit coupling information in the inter-unit coupling information storage unit 31 are converted to ROM by a ROM writer (not explicitly shown), and R in FIG.
Used as OM30.

Then, the case of performing the voice recognition operation using the neural network which has finished learning will be described.

For example, a case where the one-digit number voice "Rei" is uttered will be described. In the voice analysis unit 11, after the voice "Rei" is A / D converted by an A / D converter (not explicitly shown), for example, the frequency band of 100 to 6000 Hz is divided into 16, and the size in each frequency band, That is, 16 frequency components are extracted every 5 milliseconds, and
Power information and the like are also extracted as needed.

In the section detecting unit 12, the input voice "Rei" is input based on the characteristic parameters extracted by the voice analyzing unit 11.
The voice section is determined by determining the start time and end time of. The feature parameter of the voice section determined by the section determining unit 12 is divided into eight by the pattern creating unit 20, and the average of the frequency components extracted every 5 milliseconds in each period is averaged for each frequency component. calculate. That is, a voice pattern consisting of 16 × 8 data is created for one input voice.

When a voice pattern is created by the pattern creating unit 20, the ROM of the neural network computing unit 40
Based on the program written in (not explicitly shown), calculation is performed using the voice pattern and the inter-unit coupling information and coupling coefficient stored in the ROM 30. As the calculation result, the output value of the output layer 43a corresponding to the calculation result of 15 words in the category to be recognized is transmitted to the recognition determination unit 50. The recognition determination unit 50 determines the category having the largest output value of the output layer 43a sent from the neural network operation unit 40 as the recognition result. In this case, since the output value of the output layer 43a of the category "Rei" is the largest, the category "Rei" is determined as the recognition result.

The above method is used when the vocabulary to be recognized is constant. However, the recognition target vocabulary in speech recognition is not always constant, and the type of target vocabulary and the number of target words generally change. For example, when the number of vocabulary to be recognized changes, it becomes impossible to execute the operation with the neural network having the same structure, and the number of units in the output layer, the number of layers in the intermediate layer, the number of units in the intermediate layer, and the input It was necessary to change the number of units in a layer.

Conventionally, since the neural network operation unit 40 has a plurality of neural networks in order to support a plurality of recognition targets, a plurality of programs are used and a program is selected according to the target vocabulary. Therefore, as the number of programs increases with the increase in the types of recognition target vocabulary, the program capacity increases.

[0028]

Therefore, the present invention has been made in view of the above problems, and it is an object of the present invention to perform a neural network operation with a single program in response to changes in the vocabulary to be recognized. .

[0029]

An arithmetic unit using a neural network according to the present invention comprises a neural network selection unit for selecting a neural network, and a neural network structure storage unit for storing at least one network structure Network structure information (number of intermediate layers, number of units in input layer, number of units in each intermediate layer, number of units in output layer, address information of ROM storing inter-unit coupling information or coupling coefficient) Address information of existing ROM) is made into a table, and according to the selection of the neural network selection unit, the neural network operation unit
It is possible to operate on any neural network.

[0030]

With the above-described structure, the neural network structure information is tabulated, the neural network selection unit selects the neural network structure information in the neural network structure storage unit when the neural network calculation unit calculates, and the neural network calculation unit Calculation is performed based on the neural network structure information.

[0031]

1 is a schematic block diagram of an embodiment of a neural network of the present invention.

FIG. 2 is a schematic diagram of a hierarchical neural network having a structure different from that of FIG.

In FIG. 1, those having the same functions as those in FIG. 3 showing the conventional example are designated by the same reference numerals, and the description thereof will be omitted.

The present invention is different from the conventional example in that a neural network structure storage unit 100 and a neural network selection unit 110 are provided. The neural network structure storage unit 100 includes the number of intermediate layers and the number of input layer units. , The number of units in each intermediate layer, the number of units in the output layer, and the like are stored as one set, while the neural network selection unit 110 is stored.
Has a function of selecting neural network structure information used in the calculation in the neural network calculation unit 40.

In FIG. 2, 41b is an input layer, 42b is a first intermediate layer coupled to the input layer, 42c is a second intermediate layer connected to the first intermediate layer, and 43b is a second intermediate layer 42.
The output layer is connected to b, and is composed of L, M, N, and O units, respectively. As shown in the figure, the respective units constituting the vertically adjacent layers are connected by an information transmission path, and the coupling coefficient of the information transmission path is stored in the ROM 30.

Here, the number of units in each layer in FIG. 2 will be specifically described. The input voice is patterned by dividing the frequency band into 16 parts and the time into 8 parts. 41a number of units L = 16 × 8 =
128.

The categories to be recognized are, for example, 8 words of place names (“Tokyo”, “Osaka”, “Kyoto”, “Kobe”,
“Yokohama”, “Nagoya”, “Sendai”, “Sapporo”), the number of units O in the output layer 43b is 8, and the number of units M, N in the first middle layer 42b and the second middle layer 42c.
Are 30 respectively.

In the following description, the case where a one-digit number is a recognition target is Stage 1 and the case where a place name is a recognition target is Stage 2.

Table 1 shows the neural network structure information of stage 1 and stage 2.

[0040]

[Table 1]

For example, the one-digit numeral voice "Rei" that has been uttered
The case of recognizing will be described. The learning of the stage 1 and stage 2 neural networks is the same as in the conventional example, and will not be described here.

First, the user specifies the stage 1 by the neural network selection unit 110. When stage 1 is designated, the neural network structure storage unit 100
Neural network structure information shown in stage 1 of Table 1 (the middle layer is one layer, the input layer is 128 units, the middle layer is 50 units, the output layer is 15 units, and the address of the ROM 30 in which the inter-unit coupling information is stored is 10
00 and the address of the ROM 30 in which the coupling coefficient is stored are 2000) are transmitted to the neural network operation unit 40.

The neural network calculation unit 40 receives the above information, determines the structure of the neural network, and enters a data standby state from the pattern generation unit 20.

When the user utters "Rei" into the microphone (not explicitly shown), the voice analysis unit 11 A / D-converts the voice "Rei" and then, for example, 100
The frequency band of up to 6000 Hz is divided into 16 parts, and the size in each frequency band, that is, 16 frequency components are extracted every 5 milliseconds, and power information and the like are extracted as necessary.

The section detecting unit 12 determines the start section and the end time of the input voice "REI" based on the frequency component and the power information extracted by the voice analyzing unit 11 to determine the voice section. The feature parameter of the voice section determined by the section determining unit 12 is divided into eight by the pattern creating unit 20, and the average of the frequency components extracted every 5 milliseconds in each period is averaged for each frequency component. calculate. That is, a voice pattern consisting of 16 × 8 data is created for one input voice.

When a voice pattern is created by the pattern creating unit 20, the ROM of the neural network computing unit 40
Based on the program written (not explicitly shown), the neural network is calculated based on the voice pattern, the inter-unit coupling information and the coupling coefficient designated in the neural network structure storage unit 100. As the calculation result, the output value of the output layer 43b corresponding to the category of 15 words to be recognized is transmitted to the recognition determination unit 50 as the calculation result. The recognition determination unit 50 determines the category with the largest output value of the output layers 43b of the recognition neural network operation unit 40 as the recognition result. In this case, since the output value of the category "Rei" is the largest, the category "Rei" is determined as the recognition result.

Then, the stage 2 is designated by the neural network selection unit 110 in order to recognize the place name. Stage 2 in the neural network selection unit 110
Is specified, the neural network structure storage unit 10
0 is the neural network structure information shown in stage 2 of Table 1 (two layers for the intermediate layer, 128 units for the input layer, 30 units for the first intermediate layer, and 30 units for the second intermediate layer).
(Units, output layer is 8 units, address of ROM1 storing coupling information between units is 3000 and address of ROM1 storing coupling coefficient is 4000)
Is transmitted to the neural network calculation unit 40. Since the processing from the voice input unit 11 to the neural network calculation unit 40 is the same, it is omitted here. As an operation result, the output value of the output layer 43b corresponding to the category of 8 words to be recognized is transmitted to the recognition determination unit 50. The recognition determination unit 50 determines that the output value of the output layer 43b sent from the neural network operation unit 40 that has the largest value is the recognition result. In this case, since the output value of the category "Osaka" is the largest, the category "Osaka" is determined as the recognition result.

By the above-mentioned operation, even when the category to be recognized is changed, the neural network structure information is given to the neural network calculation unit 40, so that a plurality of neural networks can be calculated easily.

In the present embodiment, the number of intermediate layers,
Although the number of units in each intermediate layer, the number of units in the input layer, and the number of units in the output layer are limited and described, the present invention is not limited to these numbers. Also, the pattern creation unit 2
As the input to 0, the voice analysis result for use in voice recognition is used, but it is also possible to use a characteristic parameter for character recognition and the like, and the input to the pattern creating unit 20 is not limited. .

[0050]

According to the present invention, it is not necessary to change the program for performing the operation of the neural network according to the change of the category of the voice recognition target, so that a plurality of neural networks are realized by the same program. Therefore, it is possible to realize the operation of the neural network with a small program capacity.

[Brief description of drawings]

FIG. 1 is a schematic configuration diagram of an embodiment when an embodiment of the neural network of the present invention is applied to speech recognition.

FIG. 2 is a schematic configuration diagram of a hierarchical neural network.

FIG. 3 is a schematic configuration diagram of an embodiment in which a conventional neural network is applied to voice recognition.

FIG. 4 is a schematic configuration diagram of an embodiment of a learning unit of a neural network.

FIG. 5 is a schematic configuration diagram of another hierarchical neural network.

[Explanation of symbols]

 11 Voice Analysis Section 12 Section Detection Section 20 Pattern Creation Section 31 Inter-Unit Coupling Information Storage Section 32 Coupling Coefficient Storage Section 40 Neural Network Operation Section 50 Recognition Determination Section 100 Neural Network Structure Storage Section 110 Neural Network Designation Section

Claims (2)

[Claims] 1. A feature parameter extracting section for extracting a feature parameter of an input signal, a pattern creating section for creating a pattern based on the feature parameter extracted by the feature parameter extracting section, and a pattern creating section A neural network operation unit that executes an operation of the neural network using the generated pattern as input information; a neural network structure storage unit that stores the structure information of the neural network necessary for the operation in the neural network operation unit; And a neural network selection unit that selects one set of neural network structure information from a plurality of sets of neural network structure information stored in the network structure storage unit, and the neural network operation unit is created by the pattern creation unit. Patta An input layer for inputting an input layer, an intermediate layer connected to the input layer, and an output layer connected to the intermediate layer. The neural network structure storage unit includes the number of layers of the intermediate layer and a unit of the input layer. Number, the number of units in the intermediate layer, or
A plurality of sets of neural network structure information having at least one of the number of units of the output layer as a constituent element is stored, and the neural network operation unit is based on the neural network structure information selected by the neural network selection unit. An arithmetic unit using a neural network, which is characterized by performing arithmetic operations. 2. The arithmetic unit using a neural network according to claim 1, wherein the input signal to the characteristic parameter extraction unit is a voice signal or character information.