JPH09212197A - Neural network - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide the neural network which facilitates learning and can obtain a high recognition rate with simple constitution. SOLUTION: A neuron element net 22 consists of self-associative type neural networks ANN1-ANNn. Each ANN is made to correspond to each phoneme and learns only the corresponding phoneme exclusively. Namely, self-associative type learning is performed with a vector string as to each phoneme obtained by the spectrum analysis of an FFT device 21. For speech recognition, on the other hand, the vector string of the spectrum-analyzed speech is inputted to all the ANNs 1-(n). Then the similarity between the inputted vector string and an outputted victor string is calculated by each ANN and the phoneme corresponding to the ANN having the highest similarity is recognized as the phoneme constituting the inputted speech.

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a neural network, for example, a neural network used for shape recognition, voice recognition and the like.

[0002]

2. Description of the Related Art Neural networks have been attracting attention, which are intended to realize information processing by engineeringly realizing the mechanism of the human cranial nerve system. This neural network is composed of a neuron element network composed of a plurality of neuron elements that propagate data and a learning control unit that controls the learning. This neuron element network is generally composed of an input layer to which data is input, an output layer to which data is output in response to input data, and one or a plurality of intermediate layers arranged between these two layers. There is. The neuron elements in each layer of the neuron element network are connected to other neuron elements with a predetermined strength (coupling weight), and the output signal changes according to the difference in the coupling weight values. ing.

In a conventional neural network having such a hierarchical structure, a learning control process is performed by changing the connection weight between the neuron elements by a learning control unit. Learning is performed by analog or binary data (pattern) given corresponding to the number of inputs and outputs of the input layer and the output layer. Now, it is assumed that g1 to g6 are given as data, and when g1 to g3 among them are input as a learning pattern from the input layer, certain output signals p1 to p3 are output from the output layer. When the correct answer of the output signal with respect to this input signal is g4 to g6, these g4 to g6 are generally called teacher signals. Then, the output signals p1 to p3 and the teacher signals g4 to g6
Learning is performed by executing a process of correcting the connection weights of the respective neuron elements so that the error between and becomes the minimum or the same, for a plurality of learning patterns.

As a concrete method for correcting the connection weight between the neuron elements in the neuron element network so that the output signal coincides with the teacher signal, the error back-propagation method (hereinafter referred to as the BP method) has been conventionally used. ) Is often used. The BP method is to correct the connection weight between the neuron elements in all the layers constituting the neural network in order to minimize the error between the output value in the output layer and the teacher signal. That is, it is determined that the error in the output layer is the sum of the individual errors generated in the neuron elements in each intermediate layer, and not only the error from the output layer but also the error in each intermediate layer that causes the error. The connection weight is modified so that the error of the neuron element is also minimized. Therefore, all the errors for each neuron element in the output layer and each intermediate layer are calculated.

In this calculation process, the error value of each neuron element of the output layer is given as an initial condition, and the error value of each neuron element of the nth intermediate layer, the error of the (n-1) th intermediate layer, ......, and so on, the calculation processing is performed in the opposite direction. The error value of each neuron element thus obtained and the connection weight at that time are used to calculate the correction value of the connection weight. The learning is completed by repeating the above-described learning process for all learning patterns until the error with the teacher signal becomes a predetermined value or less, or a predetermined number of times. Using such a neural network, research has been conducted on pattern recognition of characters and figures of various data, analysis and synthesis processing of voice, prediction of generation of time-series pattern of motion, and the like.

[0006]

In such a conventional neural network, one network is used for all cases. That is, in the case of speech recognition, all words and phonemes to be learned are set to 1
I was learning with two neural networks. However, in order to learn each phoneme, it is necessary to carry out learning with a plurality of data. For example, in the case of the phoneme "a", not only the vowel "a" but also "a" included in consonants such as "ma", "sa", "ta", and the same "ma" Even if there is something, such as the one at the beginning of a word like matu, or mayay
It is necessary to learn what is in a word like a and what is at the end like sima. In addition, when applied to unspecified speaker recognition, it is necessary to use each of these phonemes uttered by a plurality of persons as learning data. As described above, since a single neural network has conventionally been used to recognize a large amount of data, in order to increase the recognition rate by performing sufficient learning on all learning data, the number of intermediate layers must be increased. The network size had to be increased by increasing the number of layers and the number of neuron elements in each layer, resulting in a very expensive network. Also, the larger the size, the longer the learning time. On the other hand, when performing voice recognition using a personal computer, etc., the size of the intermediate layer (the number of neuron elements) is used because the device is small.
There is a case that the learning capacity is limited and the learning is not completed enough, and the recognition rate is lowered due to insufficient learning.

Further, it is necessary to previously adapt the input spectrum to the first position of the neural network for speech recognition. Therefore, the existing method cannot handle continuous speech recognition in which the start timing of the phoneme changes freely. Furthermore, in the conventional speech recognition for a neural network, each phoneme spectrum is given independently. However, since the state of each phoneme during continuous speech recognition is affected by the state of the phoneme that appears before each, the existing neural network that performs individual recognition for each phoneme recognizes the previously presented phoneme. Since the information was not available, it was not suitable for continuous speech recognition. These problems existed not only in voice recognition, but also in figure recognition, character recognition, prediction of the occurrence of a time-series pattern of movement, and the like.

Therefore, it is a first object of the present invention to provide a neural network having a simple structure and capable of obtaining a high recognition rate and the like. A second object of the present invention is to provide a neural network capable of performing learning easily in a short time.

[0009]

According to a first aspect of the present invention, an input layer having a plurality of neuron elements, an intermediate layer having a smaller number of neuron elements than the input layer, and the same number of neuron elements as the input layer are provided. A plurality of bottleneck neuron element networks each having a different specific meaning associated with each other, and input means for inputting a vector sequence to each data input layer of the neuron element network at the bottleneck, Similarity calculation means for calculating the similarity between the output vector sequence of each bottleneck neuron element network and the input vector sequence by the input of the vector sequence by the input means, and the bottle having the highest similarity degree calculated by this similarity calculation means Output means for outputting a specific meaning corresponding to the neck neuron element network as the meaning of the vector sequence input to the input means, It is provided in-menu neural network to achieve the first purpose.

According to a second aspect of the present invention, in the neural network according to the first aspect, the plurality of bottleneck neuron element networks are self-associative with corresponding vector strings of specific meaning as input data and teacher signals. Use what you have learned. According to a third aspect of the present invention, in the neural network according to the first aspect, self-associative learning is performed for each of the bottleneck neuron element networks using a corresponding vector string of a specific meaning as input data and a teacher signal. A learning means is provided to achieve the second object. According to the invention described in claim 4, in the neural network according to claim 2 or 3, self-associative learning is performed by the back propagation rule. According to a fifth aspect of the invention, in the neural network according to any one of the first to fourth aspects, the specific meaning is a phoneme that constitutes a voice and is input to the input layer. The vector sequence
A vector sequence representing the feature amount of the phonemes analyzed in time series is used. According to a sixth aspect of the invention, in the neural network according to the fifth aspect, at least one of the spectrum data of the voice and the cepstrum data is used.

[0011]

BEST MODE FOR CARRYING OUT THE INVENTION An embodiment of the neural network of the present invention will be described in detail below with reference to FIGS. FIG. 1 shows a system configuration of a voice recognition device using a neural network. This speech recognition device is designed to input a vector sequence for learning to a neuron element network and input a teacher signal (vector sequence) to an output layer, change a connection weight between each neuron element by learning, and A CPU 11 is provided for performing various processes such as voice recognition and control based on the output signal. This CPU11
Is connected to the ROM 1 via a bus line 12 such as a data bus.
3, RAM 14, communication control device 15, printer 16, display device 17, keyboard 18, FFT (fast Fourier transform) device 21, neuron element network 22 having n self-associative NN (neural networks), and graphic reading device 24 is connected.

The ROM 13 is a read-only memory that stores various programs and data for the CPU 11 to perform processing and control such as voice recognition and learning of a neuron element network. In the ROM 13, for example, a program for performing learning based on a backpropagation rule as learning of a neuron element network, a program for sequentially recognizing phonemes from an input signal, a voice for recognizing phonemes from the recognized phonemes, and a recognized voice. It also stores the Japanese conversion system program that converts the text into text.

In the RAM 14, a predetermined program stored in the ROM 13 is downloaded and stored, and at the same time,
It is a random access memory used as a working memory of the CPU 11. FFT in RAM14
Voice data analyzed by the device 21 or the communication control device 1
For the voice data received from No. 5, a vector string storage area for temporarily storing the power at each time and each frequency is secured. The power value at each frequency becomes a vector sequence input to the input layer I of each self-associative NN of the neuron element network 22. RAM 14
When recognizing characters, figures, etc. by a neural network, the image data read by the figure reading device 24 is stored.

The communication control device 15 transmits / receives various data such as recognized voice data to / from other communication control devices via various communication networks 2 such as a telephone line network, a LAN, a personal computer communication network and the like. I do. The printer 16 is provided with a laser printer, a dot printer, etc., and is designed to print input data, the contents of recognized voice, and the like. The display device 17 includes an image display unit such as a CRT display or a liquid crystal display and a display control unit, and displays on screen the input data, the content of the recognized voice, and the operation instruction necessary for the voice recognition. Has become. The keyboard 18 is an input device for inputting parameter changes, setting conditions, and the like of the FFT device 21, and for inputting text, and a numeric keypad for inputting numbers,
Character keys for inputting characters, function keys for realizing various functions, and the like are arranged. A mouse 19 as a pointing device is connected to the keyboard 18.

A voice input device 23 such as a microphone is connected to the FFT device 21. The FFT device 21 converts analog voice data input from the voice input device 23 into digital data, and also performs spectrum analysis by discrete Fourier transform. By the spectrum analysis by the FFT device 21, a vector sequence by the power for each frequency is output for each time, and the vector sequence for each time is stored in the vector sequence storage area of the RAM 14. The figure reading device 24 is a CCD (Ch
a device for reading an image of characters, figures, etc. recorded on a sheet of paper, etc., and image data read by the image reading device 24 is stored in the RAM 14. It is like this.

FIG. 2 shows the configuration of the neuron element network 22. As shown in FIG. 2, the neuron element network 22 includes n self-associative neural networks (hereinafter, simply referred to as ANN) 1 to n. AN
The number n of N is the neuron element network 22 for the input data.
It is provided by the number of specific meanings to be distinguished by. For example, in the case of speech recognition according to the present embodiment, the number of phonemes to be recognized with respect to the input speech data is a value of n, and if recognition is performed for 80 phonemes, n =
80. Phonemes are associated with these ANN1 to ANNn, and learning of the corresponding phonemes is performed in learning of the neural network. That is, ANN1 learns about the phoneme "a", ANN2 learns about the phoneme "i", and ANN3 learns about the phoneme "u" in self-associative learning.
Each of the NN4 to NN is also designed to perform self-associative learning on the corresponding phoneme.

The ANN comprises three layers, an input layer I, an intermediate layer H, and an output layer O. The input layer I includes a number p of neuron elements I1 to Ip corresponding to the number p of input data arbitrarily selected corresponding to various processes such as voice recognition and graphic recognition. The intermediate layer H includes a number p of neuron elements H1 to Hq smaller than the number p of neuron elements of the input layer H.
(Q <p). The output layer O includes the same number p of neuron elements O1 to Op as the input layer H. As described above, the self-associative NN uses a so-called bottleneck neural network. This is because when the number of neuron elements in the middle layer is the same as the number of neuron elements in the input layer and the output layer, when self-recognition learning is performed, the connections between layers become all 1 and appropriate learning is performed. This is because it cannot be done.

Each neuron element H1 to Hq of the intermediate layer H
Is the connection weights W11 to Wpq that can be changed at the time of learning with all the neuron elements of the input layer I (hereinafter, this set is W
It is completely connected with. Each of the neuron elements H1 to Hq of the intermediate layer H has thresholds θ1 to θq that can be changed in the learning stage. Each of the neuron elements H1 to Hq of the intermediate layer H is adapted to output an output value due to forward propagation activity based on the input data input to the input layer I, the connection weight W, and the threshold value. Also, the output layer O
Each of the neuron elements O1 to Op is completely connected to all the neuron elements H1 to Hq of the intermediate layer H with variable connection weights w11 to wqp (hereinafter, this set is represented by w) during learning. . Then, each neuron element O1 to Op
Outputs the self-associative NN output value from the output value of the hidden layer H and the connection weight w.

The neuron element network 22 includes a memory (not shown), and stores a connection weight table corresponding to each self-associative NN. The connection weight table stores the connection weight W of the input layer I and the intermediate layer H, the threshold value θ of the intermediate layer, and the connection weight of the intermediate layer H and the output layer O. Note that the threshold value θ may be set for each neuron element in the input layer and the output layer. Then, the learning of the neuron element network 22 is executed by the CPU 11 changing these connection weights and threshold values according to a predetermined backpropagation rule.

Next, the operation of the embodiment configured as described above will be described. Outline of operation Self-associative neural networks ANN1 to ANNn
Corresponds to each phoneme to be recognized,
Only learn about the corresponding phonemes. In the learning, self-associative learning is performed by a vector sequence for each phoneme obtained by the spectrum analysis of the FFT device 21,
Independent of learning from other ANNs. On the other hand, in the case of performing voice recognition, the vector sequence of the voice spectrum-analyzed by the FFT device 21 is input to all ANN1 to ANNn. Then, the similarity between the input vector sequence and the output vector sequence is calculated for each ANN. For example, when the vector string of the phoneme "a" is input as the recognition target, the similarity is highest because only the corresponding ANN1 is learning about the phoneme "a". on the other hand,
Since the other ANN2 to ANNn have not learned "a", the degree of similarity is extremely small. Therefore, the phoneme corresponding to the ANN having the highest degree of similarity can be recognized as the phoneme forming the input voice.

Details of Learning of Neural Network As described above, each ANN1 to ANN1 of the neuron element network 22 is
Each ANNn corresponds to a specific phoneme, for example, ANN1 independently performs only self-associative learning of a phoneme "a", and ANN2 only self-associative learning of a phoneme "i". It is done independently. Further, the other ANNs 3 to n are also configured to independently perform only self-associative learning for the corresponding phonemes.
As described above, each ANN only needs to learn about a specific phoneme corresponding to each ANN, which makes it possible to reduce the size of the input / output layer and the intermediate layer (reduce the number of neuron elements). Easy to learn and recognize.

When learning about the neural network, the learning mode is designated by first operating the keyboard 18 or operating a predetermined key displayed on the display device 17 with a mouse. After designating the learning mode, the characters corresponding to the predetermined 80 phonemes are sequentially input from the keyboard 18, and then the voice for the phoneme is input to the voice input device 23.
The phonemes to be input may be displayed on the display device 17 to sequentially notify the phonemes to be uttered.
In the voice input device 23, for example, for the phoneme “a”, when an analog signal as shown in FIG. 4A is input, the analog signal is supplied to the FFT device 21. In the FFT device 21,
The supplied analog audio data is sampled at 22 KHz, A / D converted into 16-bit PCM data, and stored in a storage unit (not shown). The sampling interval is not particularly limited to 22 KHz,
Other intervals may be used as long as the intervals are ½ or less with respect to the utterance time of a phoneme having a short utterance time such as “ku”, “tsu”, and “p”. Further, the PCM data is not limited to 16 bits, but may be 32 bits, 10 bits, 8 bits.
It may be bits, 6 bits, 4 bits or the like.

Next, in the FFT device 21, a rectangular window, a Hamming window, and a Hanning.
Digital audio data “a” is processed by fast Fourier transform (FFT) processing at each time tn (n = 1, 2, ...) According to the shape of a time window such as a window and parameters such as the number of points.
The spectrum analysis is performed. That is, the FFT device 21 calculates the power P (tn) for each frequency (F1 to F30) of the audio data at each time tn, as shown in FIG. 4 (b). Power P of each frequency
The vector sequence by (tn) is stored in the vector sequence storage area of the RAM 14 every time as shown in FIG.

The learning data for each phoneme is generated as described above, but a plurality of learning data are generated in order to make the self-associative learning by each ANN1 to ANNn universal. The generation will be described below by taking the phoneme "a" as an example. For the phoneme "a" to be learned, the phoneme (onset phoneme) when uttered at the beginning of the word is represented by "a", and the phoneme (tail tail phoneme) when uttered at the end of the word is " It is represented by "A", and a phoneme (phoneme in phoneme) when it is uttered in the middle of a word is represented by "A". For example, "a" is taken from aki (autumn),
"A" is taken from denwa (phone), "A" is tom
Take from ari (night stay). In the following description, the phoneme "a" is divided into three parts, "a", "a", and "A".
The learning of the phoneme "a" by the pattern will be described as an example, but the learning is performed by 3 to 30 patterns, preferably about 100 patterns for each phoneme.

FIG. 6 shows these three types of "A", "A",
For “A”, the FFT device 21 uses each time t (t = 1,
2, ...) Represents data obtained by spectrum analysis by FFT processing. The FFT device 21 is shown in FIG. 6 for each phoneme “A”, “A”, and “A”.
As shown in (a), (b), and (c), at each time t, each frequency of the audio data (the number of frequency divisions corresponds to the number p of neuron elements in the input layer I of the ANN, F1 to Fp
Value of the power (P) for each of
Then, the vector sequence for each time with the power P (t) of each frequency is stored in the own vector sequence storage area of the RAM 14 for each phoneme.

Now, as shown in FIG. 6A, the vector sequence of the power P (1) at the time t = 1, which is spectrally analyzed for the phoneme "A", is set to A1, and the time t =
The vector sequence of power P (2) in 2 is A2, and similarly, although not shown, the vector sequence at time t = n is A. In addition, as shown in FIG. 6B, the phoneme "A"
The vector sequence of the power P (1) at the time t = 1 and the vector sequence of the power P (2) at the time t = 2, which are spectrally analyzed for
Although not shown, the vector sequence at time t = n is an.
Further, as shown in FIG. 6C, the power P at time t = 1, which is spectrally analyzed for the phoneme “A”.
The vector sequence of (1) is A1, the vector sequence of power P (2) at time t = 2 is A2, and similarly, although not shown, the vector sequence of time t = n is An.

By the vector sequence composed of the power P (t) spectrally analyzed for each of these phonemes,
Learning of ANN1 is performed at each time t. That is, the vector sequence A1, A1, A of each phoneme at the same time, for example, t = 1 is used as the input data of the input layers I1 to Ip of ANN1 and is used as the teacher signal of the output layers O1 to Op. Then, learning is performed for each vector sequence at each time t.

FIG. 7 shows input data and teacher signals in the learning of the self-associative type N27. In FIG. 7, an example is shown in which learning is performed based on the vector sequence of power for each phoneme shown in FIG. As shown in FIG. 7, learning is performed with each time t (t = 1, 2, ... N) as a unit. For example, in the case of time t1, input data A1, A1 and A1 with the teacher signal A1.
Then learning the input data A1 and A1 and A1 with the teacher signal A1.
With the teacher signal as A1, learning is performed on input data A1, A1 and A1. Further, learning is performed by all combinations of the input data and the teacher signal by A2, A2, and A2. In the same manner, learning about other points t, t, and At is performed.

As described above, even if the same phoneme, a plurality of phonemes (a head phoneme, a middle phoneme, and a tail phoneme) by a plurality of people are used.
Using the training data “A”,
Not only in the case of performing the self-associative learning as, the self-associative learning is performed by using other "A", "a", etc. belonging to the same phoneme as a teacher signal. As a result, it is possible to learn about universal phonemes included in the category of the phoneme "a" for the same phoneme.

On the contrary, the same pattern may be used as the input data of the input layer I and the teacher signal of the output layer O instead of the combination of each pattern of each phoneme. That is, for the learning data "a",
The self-associative learning may be performed only for the teacher signal “A” having the same pattern. Although not shown in FIG. 7, not only when learning is performed at the same time t, for example, “A1”, “A1”, and “A1” are compared with “A2” with respect to the learning data. , "A2", "A2" may be used as the teacher signal. That is, the data of tn and tn + 1 may be used as the teacher signal for learning the data of each time tn.

In the learning of ANN1, when the input of the vector sequence to the input layer I and the input of the teacher signal to the output layer are completed, the CPU 11 shows the input layer I, the intermediate layer H, and the intermediate layer H for the ANN1 shown in FIG. Learning is performed using the connection weight W between each neuron element of the output layer O and the threshold value θ, and each connection weight is updated to the value after learning. Although the self-associative learning of the corresponding phoneme “a” has been described above for ANN1, the self-associative learning for the corresponding phonemes is similarly performed for ANN2 to ANNn.

In the present embodiment, the learning to be performed is the learning based on the back propagation rule. The learning formula is Δw (t) = [S (t) / [S (t-1) -S
(T)]] × Δw (t−1), and the details of the equation and the learning algorithm (The Quickprop Alg
orithm) is Carnegie Mellon University 1988 9
Issued monthly, S.M. Technical report #CMU by Fahlman
-CS-88-162, "An Imperial
Study of Learning Speedin
Back-Propagation Network
s ". Also, Elman (JL El
man)) Finding structure
in time, Cognitive science
e, 14, pp. 179-211 (1990), learning that applies the backpropagation rule of the feedforward network to the recurrent network of discrete time may be applied. Further, learning is not limited to the above method, and other learning method may be used.

The spectrum data for each phoneme to be learned is the input device 23 and the FFT device 21.
It is also possible to input from the communication control device 15 data on various phonemes that have been spectrally analyzed in advance by another device and store them in the vector sequence storage area of the RAM 14 instead of generating them at the time of learning.

Details of voice recognition When the voice to be recognized is input from the voice input device 23 after the learning of each ANN1 to ANNn is completed, the FFT device 21 performs spectrum analysis,
The power for each frequency is stored in the vector sequence storage area of the RAM 14 as a vector sequence at each time t. Then, the CPU 11 stores the vector sequence P (tn) at the time tn in the RAM 1 for the voice data to be recognized.
It is read out from No. 4 and input to the input layer I of ANN1. Then, the output vector sequence O (tn) for ANN1 is obtained from the learned connection weight table stored in the memory of the neuron element network 22. Then, the CPU 11
The similarity S1 (tn) between the output vector sequence O (tn) of this ANN1 and the input vector sequence P (tn) is calculated. Similarly for the other ANN2 to ANNn, the CPU 11 obtains the output vector sequence when the vector sequence P (tn) is input and the similarity S2 (tn) to Sn (tn) of the output vector sequence with respect to the input vector sequence. calculate. The similarity is the difference between the input and output, the Euclidean distance,
It is calculated by the least squares value and various other methods.

The CPU 11 has all the ANN1 to ANNn.
After calculating the similarities S1 (tn) to Sn (tn), the phoneme corresponding to the ANN having the highest similarity is recognized as the phoneme at the time t1 of the input voice, 2 is stored in the RAM 14. That is, when the similarity S2 of ANN2 is the largest, the phoneme at time tn is recognized as "i". In this way, each ANN1-A
Since the NNn learns only the corresponding phonemes, the AN corresponding to each phoneme of the input speech data.
The similarity of N is extremely high, and the similarity of other ANNs is extremely low, and it becomes possible to specify a phoneme from the value of the similarity S.

Similarly, the vector sequence P after the time tn + 1 is sequentially read from the RAM 14 and the ANN1 to ANN1 are sequentially read.
The phoneme at that time is recognized from the maximum value of the similarities S1 to Sn of ANNn and sequentially stored in the RAM 14.

Since the phoneme is specified every time the vector sequence P (tn) is input to each of the input layers I of ANN1 to ANNn in time series, the RAM 14 stores a plurality of phoneme sequences. For example, the voice "iro" is input, and the phoneme sequence "iiiiiirrrooooo" recognized at each time is stored in the RAM. The CPU 11 recognizes the input voice as “iro” from the phoneme string stored in the RAM 14. Then, when there is an input instruction from the keyboard 18, the CPU 11 converts the recognized voice into a text by the Japanese conversion system. The converted text is displayed on the display device 17 and stored in the RAM 14. Further, according to an instruction from the keyboard 18, data is transmitted to various communication control devices such as a personal computer and a word processor via the communication control device 5 and the communication network 2.

Even if the phoneme has the maximum similarity, it may be erroneously recognized if the similarity does not exceed a predetermined threshold value. In this case, the vector input to the input layer I is removed from the recognition target. This is likely to occur in the case of a vector sequence that is spectrally analyzed in the middle of changing from each phoneme to a phoneme. That is, in the case of a spectrum between phonemes, all the similarities S of ANN1 to ANNn may be low, and in such a case, the phoneme at the time t cannot be recognized. However, the voice can be easily recognized by the phoneme that is continuously specified thereafter.
For example, for the voice "Iro", "iii? Rr? O"
It is assumed that the output "o" is made. Even if there is a vector sequence (vector sequence represented by?) That is removed from the recognition target due to the low degree of similarity in this way, the phonemes that make up the input speech are recognized before and after that.
The input voice "color" can be recognized as a whole.
Therefore, continuous speech recognition can be easily performed.

The reason why the phoneme cannot be specified when each phoneme changes is that the learning is performed for each phoneme unit at the learning stage, and the effect of each phoneme on the spectrum is not the object of learning. Is considered to be.

According to the present embodiment, each of ANN1 to AN is
Since Nn specializes in learning only corresponding phonemes, it is possible to obtain a high recognition rate by performing abundant learning about corresponding phonemes (learning phonemes in the case of multiple people by multiple people). You can Therefore,
Unspecified speaker recognition can be performed.

Further, in the case of performing speech recognition on a phoneme basis, it has been a problem in the past how to accurately determine the starting point of a phoneme to be recognized. According to this embodiment, " , Etc. is sampled at intervals of 1/2 or less than the utterance time of a phoneme having a short utterance time.
With respect to the data, it is not necessary to specify the starting point of one phoneme. Also, when performing continuous speech recognition in phoneme units,
The voice can be recognized regardless of the utterance time of each phoneme having a large individual difference. For example, as a voice, "Haru"
In this way, even when the voice "ha" is extended and uttered, "hhhhhh ... aaaaaaaaaaaaaaaaa"
Phoneme "a", such as "rrrr ... uuuuu ..."
It is possible to easily recognize the voice "Haru" simply by specifying many.

Further, in the present embodiment, the speech recognition in the unit of phoneme has been described, but the speech recognition may be performed in the unit of word. In this case, a code string representing the word as a specific meaning represented by the vector string is used as a teacher signal.

Further, in this embodiment, the neuron element network 22 is learned by the CPU 11 according to the learning program stored in the ROM 13, and the neuron element network 2 after learning is learned.
Although the voice recognition by 2 is performed, it is possible to perform continuous voice recognition of an unspecified speaker with a high recognition rate.
Less need for re-learning. Therefore, the speech recognition device does not necessarily have to have a learning function, and has the connection weight W and the threshold value θ obtained by learning of another device.
You may make it use the neuron element network which consists of 1-ANNn. In this case, the neuron element network may be configured by hardware having learned connection weights.

In the embodiment described above, the FF
Although the spectrum analysis was performed on each phoneme during learning and the speech during speech recognition by the fast Fourier transform in the T device, the spectrum analysis may be performed by another algorithm. For example, spectrum analysis by DCT (discrete cosine transform) or the like may be performed.

The above-described ANN1 to ANNn of FIG.
In the above, the case where the layers I, H, and O are completely bonded has been described, but the present invention is not limited to this. For example, the connection state may be determined according to the number of neuron elements in each layer and the learning ability.

Next, a second embodiment will be described. In the above-described first embodiment, FFT is performed in speech recognition.
The vector sequence spectrally analyzed by the device 21 is input to the input layer I.
In the second embodiment, the cepstrum data is input to each input layer I to perform learning and voice recognition. FIG. 9 shows a system configuration of the neural network according to the second embodiment. As shown in this figure, in the neural network, the system of the first embodiment shown in FIG. 1 is further provided with a cepstrum device 26. Since the other parts are the same as those in the first embodiment, the same numbers are given and the description thereof is omitted.

The cepstrum device 26 is the FFT device 21.
The cepstrum data is obtained by performing an inverse Fourier transform on the logarithm of the short-time amplitude spectrum of the spectrum analyzed in the above. With this cepstrum device 26, the spectral envelope and the fine structure can be approximately separated and extracted.

Here, the principle of the cepstrum will be described. Now, when the Fourier transform of the impulse response of the sound source and the Fourier transform of the sound path are respectively expressed by G (ω) H (ω), the relationship of X (ω) = G (ω) H (ω) is expressed by the linear separation transmission circuit model. can get. When the logarithm of both sides of this equation is taken, the following equation (1) is obtained. log | X (ω) | = log | G (ω) + log | H (ω) | ... (1) Further, when the inverse Fourier transform of both sides of this equation (1) is taken, the following equation (2) is obtained. This is the cepstrum. c (τ) = F ⁻¹ log | X (ω) | = F ⁻¹ log | G (ω) + F ⁻¹ log | H (ω) | (2) Here, the dimension of τ is from the frequency domain. Since it is an inverse transformation, it takes time, and it is called kefrenshi.

Next, the extraction of the basic period and the envelope will be described. The first term on the right side of the equation (1) is the fine structure on the spectrum, and the second term is the spectrum envelope. There is a big difference between the inverse Fourier transforms of the two, the first term is a peak of high kefflenency, and the second term is concentrated in the low kefflenency portion of about 0 to 2 to 4 ms. The logarithmic spectrum envelope is obtained by performing Fourier transform using the high-keflency part, and the spectrum envelope is obtained by subjecting it to exponential transformation. The degree of smoothness of the obtained spectrum envelope changes depending on how many components in the low Keffency portion are used. The operation of separating the kefrenshi component is called lifter ring.

FIG. 9 shows the structure of the cepstrum device 26. The cepstrum device 26 includes a logarithmic transformation unit 261, an inverse FFT unit 262, and a cepstrum window 2
63, a peak extraction unit 264, and an FFT unit 265. The cepstrum window 263, the peak extraction unit 264, and the FFT unit 265 use the inverse FFT unit 2 as data to be supplied to the voice input layer 32 of the neuron element network 22.
It is not necessary when using the cepstrum data obtained in 62, and is necessary when using the spectrum envelope as input data of the neuron element network 22. Also, FF
The T section 265 is not always necessary, and the FFT
The device 21 may be used.

The logarithmic transformation unit 261 calculates the mathematical expression (1) from the spectrum data X (ω) supplied from the FFT device 21.
Logarithm conversion is performed according to to obtain log | X (ω) |
The inverse FFT unit 262 is supplied. In the inverse FFT unit 262,
Inverse FFT is further performed on the supplied value to obtain c (τ)
By calculating, the cepstrum data is obtained. Reverse F
The FT unit 262 supplies the obtained cepstrum data to the input layers I of ANN1 to ANNn described in the first embodiment as input data In (vector sequence) for learning or recognizing voice data. It is like this. Input data I input to ANN1 to ANNn
As for the number of n, the same number as the number of neuron elements p of the input layer I, which is arbitrarily selected in accordance with the voice recognition, is selected. Therefore, the keffency (τ) axis is divided into p parts, and the power value for each keffency is supplied to each ANN1 to ANNn as the input data of the neuron elements I1 to Ip. The cepstrum data obtained by the inverse FFT unit 262 is ANN1.
It is the first example in the second embodiment that supplies to each input layer I of ANNn.

Next, a second example of the second embodiment will be described. In this second example, the cepstrum window 26
By performing lifter ringing on the cepstrum data obtained in step 3, the keflenency component is separated into a high kefflenency portion and a low kefflenency portion. In the FFT unit 265, the separated low-keflency portion is Fourier-transformed to obtain a logarithmic spectrum envelope, and further exponentially transformed to obtain a spectrum envelope curve. From this spectrum envelope data, the frequency axis is divided according to the number of neuron elements, and the power value (vector sequence) for each frequency is input to each input layer I of ANN1 to ANNn.
To supply.

It should be noted that the cepstrum data of the low-keflency portion separated by the cepstrum window 263 may be supplied to each input layer I as input data. Alternatively, the peak extraction unit 264 may extract the fundamental period from the separated cepstrum data of the high-keflency portion, and this may be used as one of the input data together with the spectrum envelope data obtained by the FFT unit 265. In this case, since the number of neuron elements in the input layer I is p, the input data In1 to (p-1) of the spectrum envelope data
I (p-1) n is input to each input layer I, and input data Inp is input to the input layer I from the data of the basic period.

As described above, according to the second embodiment, by using the cepstrum data of the voice data, the learning and recognition of the phoneme can be performed by the data in which the features of the voice are more captured than the power spectrum. As a result, the recognition rate is improved.

Although voice recognition has been described in the first and second embodiments, image recognition may be performed using cepstrum data of image data. The image data in this case may be either the image data read by the graphic reading device 24 or the image data received by the communication control device 15.

In the above-described first and second embodiments, the case of voice recognition has been described. However, the present invention can be applied to character recognition, figure recognition, prediction of occurrence of a time series pattern of motion, and the like. You can In the case of character recognition, an ANN having the number of characters to be recognized is provided, and each ANN is
Self-associative learning is performed for each corresponding character. As learning data, feature points such as intersections, stroke order, and stroke count are used.

[0057]

Therefore, according to the neural network of the present invention, an input layer having a plurality of neuron elements,
A plurality of bottleneck neuron element networks each having an intermediate layer having fewer neuron elements than the input layer and an output layer having the same number of neuron elements as the input layer, each having a different specific meaning associated with each other, Input means for inputting a vector sequence to each data input layer of the neuron element network at the bottleneck, and the similarity between the output vector sequence of each bottleneck neuron element network and the input vector sequence due to the input of the vector sequence by this input means Outputting means for outputting similarity meaning calculating means and a specific meaning corresponding to the bottleneck neuron element network having the highest similarity calculated by the similarity calculating means as the meaning of the vector sequence input to the input means Since the above is provided, a high recognition rate and the like can be obtained with a simple configuration. Further, since self-associative learning using the corresponding vector string of a specific meaning as the input data and the teacher signal is performed for each bottleneck neuron element network, the learning can be easily performed in a short time.

[Brief description of drawings]

FIG. 1 is a system configuration diagram of a voice recognition device using a neural network according to an embodiment of the present invention.

FIG. 2 is a configuration diagram of a neuron element network of the voice recognition device.

FIG. 3 is an explanatory diagram showing a connection weight table that stores the connection weights and threshold values of each ANN of the neuron element network 22.

FIG. 4 is an explanatory diagram explaining a state of spectrum analysis of voice by the voice recognition device.

FIG. 5 is an explanatory diagram showing a vector sequence for a voice spectrum-analyzed by the FFT device of the voice recognition device.

FIG. 6 is an explanatory diagram showing data obtained by spectrum analysis of three types of “A”, “A”, and “A” by the FFT device of the voice recognition device.

FIG. 7 is an explanatory diagram showing a relationship between input data and a teacher signal during learning of ANN1 in the neuron element network of the voice recognition device.

FIG. 8 is a system configuration diagram of a neural network according to a second embodiment of the present invention.

FIG. 9 is a configuration diagram of a cepstrum device according to a second embodiment.

[Explanation of symbols]

 11 CPU 12 Bus Line 13 ROM 14 RAM 15 Communication Control Device 16 Printer 17 Display Device 18 Keyboard 21 FFT Device 22 Neuron Element Network 23 Voice Input Device 24 Graphic Reading Device 26 Cepstral Device

Claims (6)

[Claims] 1. An input layer having a plurality of neuron elements, an intermediate layer having a smaller number of neuron elements than the input layer, and an output layer having the same number of neuron elements as the input layer, each having a different specification. A plurality of bottleneck neuron element networks having associated meanings, input means for inputting a vector sequence to each data input layer of the neuron element network with the bottleneck as the bottleneck, and each bottleneck neuron by inputting a vector sequence by the input means The similarity calculation means for calculating the similarity between the output vector sequence and the input vector sequence of the element network, and the specific meaning corresponding to the bottleneck neuron element network having the highest similarity calculated by the similarity calculation means, Output means for outputting the meaning of the vector sequence input to the input means. Network. 2. The plurality of bottleneck neuron element networks have been subjected to self-associative learning using a corresponding vector string of a specific meaning as input data and a teacher signal. Neural network. 3. A learning means for carrying out self-associative learning for each of the bottleneck neuron element networks, using a corresponding vector string of a specific meaning as input data and a teacher signal. Neural network described. 4. The neural network according to claim 2, wherein self-associative learning is performed according to the back propagation rule. 5. The specific meaning is a phoneme constituting a voice, and the vector sequence input to the input layer is a vector sequence representing a feature amount of a phoneme analyzed in time series. The neural network according to any one of claims 1 to 4. 6. The neural network according to claim 5, wherein at least one of speech spectrum data and cepstrum data is used as a vector representing a feature amount for the specific meaning.