JPS6386654A - Secret call system - Google Patents

Abstract

PURPOSE:To effectively inform the information, whose contents a third person is not wanted to hear, by telephone by processing to recongize characters inputted as hand written characters and the characters inputted as images, synthesizing a speech as for a character code string shown according to recognized results and transmitting the synthesized speech instead of an input voice. CONSTITUTION:A telephone terminal providing a transmission part A and a reception part B and having a function for communicating voice information by telephone through a communication part C, for instance a work station, provides a means D synthesizing a speech as for the information expressed by inputted character code strings and communicates the voice data synthesized in terms of speech instead of input voice by telephone. Especially, the character code strings are made to given, for instance, after character pattern inputted as hand writing through a tablet E is processed to be recognized (a character recognition part F), or the character pattern read and inputted through an image.scanner G is processed to be recognized (the character recognition part F), so that the information inputted through these means can be transmitted in the voice synthesis. Thus the contents of information, which is not desired to be made to transpire to a third person, can be informed by telephone.

Description

DETAILED DESCRIPTION OF THE INVENTION (Field of Industrial Application) The present invention relates to a secret telephone communication system that allows effective telephone communication of content that is undesirable to be intercepted in the vicinity of the caller.

(Prior Art) With the development of telephone systems, telephones are being used for various types of information communication. However, depending on the content of the communication, there may be cases where the user does not want a third party in the vicinity to hear the content of the call. For example, when notifying a bank account number or the amount of a transaction over the phone, there are cases where the contents are not known to a third party. There are many cases like this, especially when carrying out office work.

Conventionally, people have generally taken measures such as talking in a low voice, but this has the problem that it becomes difficult for the person on the other end of the call to hear the voice.

(Problem to be solved by the invention) In this way, in the past, communication content was transmitted to a third party near the caller.
It was difficult to make the telephone communication without being overheard. SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a secret communication system that can accomplish the purpose of notifying a person over the phone of information that should not be leaked to a third party.

[Configuration of the Invention (Means for Solving the Problems)] The present invention includes a transmitter section A and a receiver section B, the schematic configuration of which is shown in FIG. A telephone terminal, such as a workstation, which is equipped with a function to communicate information over the telephone, is equipped with a means for speech-synthesizing the information indicated by the input character code string, and uses this speech-synthesized speech data in place of the input speech. It was designed for telephone communication.

In particular, the above-mentioned character code string is processed, for example, by recognition processing of a character pattern input by hand via a tablet E (character recognition unit F), or by recognition of a character pattern read and input via an image scanner G. The information inputted through these means is then processed (by the character recognition unit F) and transmitted as synthesized speech.

(Function) According to the present invention, when a caller wants to make a call about content that they do not want a third party around them to hear, for example, by inputting the content information in writing via a tablet, the input information can be transmitted by voice. Combined and output.

Therefore, it is no longer necessary to speak in a low voice as in the past, and this eliminates problems such as difficulty for the person on the other end of the phone call to hear the telephone voice.

(Example) Embodiments of the present invention will be described below with reference to the drawings.

FIG. 2 is a schematic diagram of an intelligent workstation according to an embodiment of the present invention. This intelligent workstation is
It is composed of the following parts.

Bus 1: Used to transfer necessary information between each section described below.

Control unit 2: Mainly composed of a microprocessor, it controls the operations of each part of the intelligent workstation.

Image input device 3; consists of a camera, scanner, OCR, etc., and inputs various image information.

Position coordinate input device 4: Consists of a tablet, mouse, etc., and inputs specified position coordinate information.

Audio input unit 5: Consists of a microphone, etc., and inputs audio information.

Keyboard section 8: Equipped with a number of keys, it is used for inputting character/symbol codes, control codes, etc.

IC card unit 7: As described later, an IC card is installed therein, and necessary information is input/output between the IC card unit and the IC card.

Bus controller 8; controls information transfer between each unit via bus 1.

Audio output unit 9: Consists of a speaker, etc., and outputs audio information.

Display section 10. It consists of a CRT display, liquid crystal display, etc., and displays characters, figures, images, etc.

Image output device 11. It consists of a fax machine, a color printer, etc., and prints out various image information.

communication device 12.13; the workstation and the telephone;
Or, it communicates information with other workstations, terminals, etc. installed in remote locations.

Switching device 14: Switches and uses a plurality of communication devices.

Timer part 15: Provides time information and time information to the workstation.

Encryption processing unit 16: Encrypts various information.

Voice verification unit 17. Verification processing is performed to determine whether the given voice information is a specific voice.

Image verification unit 18: Performs verification processing to determine whether the given image information is a specific image.

Speech recognition unit 19. Verify and process the given audio information.

Speech analysis unit 20: Analyzes the voice input from the voice input unit 5 etc. by extracting the characteristics of the voice.

Character recognition unit 21: Recognizes character/symbol patterns input from the image input device 3 or the like.

Image recognition unit 23: Recognizes graphic images and the like input from the image input device 3 and the like.

Output format selection unit 24: Selects and controls the format of information output from the workstation.

Work environment data collection unit 25; collects and inputs information such as the functional status of the workstation and the work environment in the office.

Speech synthesis unit 26; generates synthesized speech according to the processed data.

Image compositing unit 27; performs compositing processing on a plurality of pieces of image information, and executes image editing processing according to processing data.

Graphic composition processing unit 28: Composes various graphics and executes editing processing such as adding and deleting graphics according to processing data.

Audio compression/expansion unit 29: Compresses and decodes audio data, and decompresses and decompresses compressed audio data.

Image compression/expansion unit 30; compresses and encodes image data, and restores and expands compressed image data.

Signal processing unit 31. It executes a series of signal processing such as encoding and compressing various signal information, decompressing and decompressing it, and adding necessary information.

Database unit 32: Classifies various information into a plurality of relations and stores them as a database. still,
This database is constructed not only as code information but also as images, sounds, etc.

The intelligent workstation according to the present invention is basically configured with each of the above-mentioned parts, and each of the above-mentioned parts effectively utilizes its respective functions to exhibit an intelligent function as a whole. .

Next, the IC card section 7, encryption processing section 16, etc., which are not common like the aforementioned keyboard section 5, etc., but have characteristic functions in this intelligent workstation, will be explained in more detail.

First of all, an IC card has a semiconductor circuit such as a microprocessor and a memory circuit built into the card body 7a, which is the size of a business card, for example, as shown in FIG. It is constructed by providing an interface section 7b for connection to the station main body and a display window section 7c.

The display window section 7c is formed by embedding a transparent polarizer, and its position is set so as not to overlap with the interface section 7b or the semiconductor circuit. Also, the card body 7a
Only the portion corresponding to the display window 7C may be transparent, or the entire substrate may be transparent.

Specifically, as shown in an exploded perspective view of FIG. 4, the IC card has a pair of cover substrates 7d.

7c, an embedded substrate 7f1, a core sheet material 7g1, a printed circuit board 7h, sandwiched between these cover substrates 7d and 7e.
It is constructed by integrally bonding them together by thermocompression.

An input/output terminal 71 is provided on the second printed circuit board 7h at a position facing the interface portion 7b, and a liquid crystal display device 7j is provided at a position facing the display window portion 7c. Furthermore, a semiconductor integrated circuit 7k is mounted on the printed circuit board 7h.
is provided. Further, the cover substrate 7c is provided with a metal foil 7o+ for dissipating heat generated in the printed circuit board 7h.

Note that the holes drilled in the cover substrates 7d and 7e, the embedded substrate 7r, and the core sheet material 7g are the printed circuit board 7.
The semiconductor integrated circuits 7j and the like integrated in the semiconductor integrated circuits 7j and the like are provided at positions facing each other. The semiconductor integrated circuit 7, etc. are fitted into these holes, and the cover substrate 7d,
7e, embedded board 7'', core sheet +47g, and printed circuit board 7h are laminated and integrated to form an IC card.The input/output terminal 71 is exposed through the hole drilled in the cover base !27d, It constitutes an interface section 7b that is electrically connected to the workstation main body.

The liquid crystal display device 7j has a liquid crystal layer between a pair of polyether sulfon film substrates provided via a spacer, as shown in FIG. A transparent conductive film is formed on the inner surface of the film substrate, and a polarizer or a reflector is provided on the lower film substrate. If the liquid crystal display device 7j is constructed using a polyether sulfone film substrate in this way, it is easy to reduce the thickness to Q,fiIU or less, compared to the case where the liquid crystal display device is constructed using a glass substrate. This allows the IC card itself to be made thinner.

Further, the driving power for this IC card may be supplied from the workstation main body via the interface section 7b, or it may be built into the card. In this case, it may be incorporated as a sheet-like battery using, for example, a polymer film.

The semiconductor integrated circuit 7 includes, for example, a CPU 7p and a PROM 7 which is a data memory, as shown in Figure Ts6.
q, an E2 PROM 7q, and a selection unit 78 for these memories.

FROM 7Q is a large-capacity nonvolatile memory that cannot be erased or rewritten, and stores control programs for the CPU 7p, information to be permanently recorded, and the like. Also E2P
The ROM 7r is a rewritable, small-capacity nonvolatile memory, and stores information that is updated when used, such as information transaction numbers and information serial numbers.

These memories are selectively driven under the control of the selection section 7s and input/output information to/from the CPU 7p. The CPU 7p executes necessary information processing using these memories, and outputs information from its interface section to the intelligent workstation main body via the terminal section 71 mentioned above.

The IC card unit 7 is equipped with such an IC card and outputs information to and from the IC card.

Incidentally, it goes without saying that the IC card is not limited to the configuration described above, and it goes without saying that the IC card section 7 can be configured depending on the configuration.

Next, the encryption processing section 16 will be explained.

For example, as shown in FIG.
It is configured to include a public key file section 16d1 and a key update section 16c.

As the concept is shown in Figure 8, a given original communication text can be encrypted according to an encryption key to generate the encrypted message, or conversely, a given encrypted communication can be decrypted according to the encryption key to generate the encrypted message. Execute the process to obtain the original text.

Private key file part lee and public key file part led
is for storing keys used for this encryption/decryption, and the key updating section 16e is in charge of updating these files of keys.

Here, the secret key is a key known only to the workstation that controls the encryption processing section 1B, and is kept secret from other workstations.

On the other hand, the public key is paired with each private key set in each workstation, and is given to each other workstation and made public. The public key file section led stores the public keys published by these plurality of workstations in correspondence with each workstation.

As shown in FIG.
8i and an encryption type adding section 16j.

When attempting to encrypt the original communication text and communicate information, the original communication text is encrypted using the public key published by the workstation of the communication partner, and information indicating the type of encryption is added to the encrypted communication text. It creates communication information and communicates it. In addition, the information on the type of encryption is as follows.
For example, it includes information indicating that "0° is not encrypted,""1" indicates that it is encrypted, information indicating the encryption method, etc.

Further, the decryption unit 16b inputs the encrypted message encrypted and communicated by the workstation using the public key published by the self-workstation, and decrypts it using the private key corresponding to the private key. 10th
As shown in the figure, a ciphertext division section 18k, a cipher type determination section 16m1 and a switching section 16n.

13p and an R3A processing section 16q.

The ciphertext dividing section 18 is for dividing the communication information communicated in the above-mentioned format into the above-mentioned cipher type information and the encrypted message, and the cipher type determining section 16m divides the communication information communicated in the above-mentioned format into the cipher type information and the encrypted message. It is determined whether the message is encrypted or not. If the message is not encrypted, the message is outputted via the switching units 18n and 16p,
If the message is encrypted, the message is sent to the RSA processing unit 1.
It leads to 69. The encrypted message is decrypted in the RSA processing section 16Q using the private key and outputted via the switching section lap.

The RSA processing sections 16i and 1Bq are configured to include a block division section 16s, an exponentiation/remainder calculation section tet, and a block concatenation section 16u, as shown in FIG. 11, for example.

Here, the block division section 16s divides the given signal sequence into blocks M of a constant length, and the exponentiation/remainder calculation section tet divides each block M.

At each time, a signal sequence N of N − M, k (mod n) is obtained using an encryption key k. However, n is a fixed value. This signal series Ni is sequentially connected and outputted via the block connection unit 18u.

In the encryption process, the signal sequence M is the original communication text, and the encrypted communication text is obtained from the original communication text as the signal sequence Ni. In the decoding process, the signal sequence M1 is an encrypted message, and the original communication text decrypted from this encrypted message is obtained as the signal sequence N1.

The key k responsible for such encryption/decryption is the aforementioned public key and private key, which are set as a pair.

Therefore, each workstation can encrypt communication information according to a public key made public by another workstation, but the encrypted communication can only be decrypted using the secret paired with the public key. Only certain workstations can know the key.

A workstation that wishes to encrypt and communicate information thus encrypts the original communication text in accordance with a public key made public by the workstation with which it communicates.

The communication information can only be decrypted by the communication partner's workstation that has the private key.

Note that it is not necessary to store all the public keys published by other workstations in the public key file 16d. For example, public keys published by each workstation are stored in a file corresponding to each workstation in a public key file memory provided separately for the system. When information communication becomes necessary, the public key of the other party is read out from the public key file memory 1.
It may be stored in .

The above is the basic configuration of the encryption processing section 16 and its functions.

Next, the image matching section 18 will be explained.

The image matching unit 18 receives image information inputted from the image input device 3, for example, an image of an individual's face, and performs individual identification.

FIG. 12 shows a schematic configuration of this image matching section, where 1lla is an image storage section, 18b is a normalization circuit, and 1lla is an image storage section;
8c is a binarization (thinning) circuit, and 18d is a feature data extraction circuit. Further, 18e is a data storage section that stores image data, 18f' is a search circuit, 18g is a collation circuit, and 18h is an output section.

The image storage section 18a stores image information input through the image input device 3, and performs image verification processing. The normalization circuit 18b performs normalization processing on the image information stored in the image storage section 18a, and the binarization circuit 18c performs binarization processing on the image information.

Specifically, here, the normalization circuit +8b normalizes the size of the individual's face in order to identify the individual based on the image of the individual's face. The binarization circuit tse performs, for example, edge line segment detection and line thinning processing on the normalized face image to obtain a binary image of the image.

The feature data extraction circuit 18d performs normalization and
This method extracts feature data from binarized image information. That is, in the matching process using a face image, for example, as shown in FIG. 13, the contour of the face is extracted as one feature, and then features such as the eyes, nose, and mouth in that image are also extracted. . Specifically, the contour features of the face are classified as code information, and the distance between the eyes 110
The size m10 of the image, the distance from the mouth, iln, etc. are extracted as numerical data as features of the image.

Therefore, the data storage unit 18e stores, for example, the 14th facial image feature data obtained for each individual in advance.
It is registered as shown in the figure. That is, the above-mentioned facial image characteristic data is registered for each individual using the individual's name as an identification name, and the facial image data are connected by pointers.

The search circuit 18r searches the theta section +8e based on the feature data extracted by the feature data extraction circuit 18d. And the search data is collation circuit 18g
, and is compared with the feature data obtained by the feature data extraction circuit 18d.

This matching process is performed, for example, when the feature data of the input image obtained by the feature data extraction circuit 18d is X + (t is the type of feature), and the feature data of the image registered in the data storage section 18e is Y. , D''''Σ IX, -Y, 1 (1) This is performed by performing the following calculations, and identifying the person with the smallest value as the result of the calculation as that individual. This identification result is output via the output section 18'h.

The image matching unit 18 basically performs matching processing on the input image in this manner, and performs, for example, personal identification of the input image.

Next, the speech recognition section 19 will be explained.

The speech recognition unit 19 is configured as shown in FIG. 15, for example. The audio input circuit 19a inputs the audio signal input from the audio input section 5 or the audio signal received by the communication device 12.13 via the public telephone line, and converts this input audio signal into an appropriate form. Amplifiers that amplify the signal to a certain level, bandpass filters that limit the bandwidth, and A/
It is composed of a D converter and the like. The input audio is limited to signals in a frequency band of, for example, 30 to 3400 Hz, and is quantized into a 12-bit digital signal at a sampling period of 12 KFIz in the audio input circuit 19a.

The acoustic processing section 19b is composed of, for example, a product-sum circuit constructed from dedicated hardware. Basically, it operates at high speed in a pipeline manner in synchronization with the audio input circuit 19a.

The acoustic processing here is performed by two types of bandpass filter groups. One of them is a 16-channel filter bank through which the spectral changes of the input row signal are extracted.

The other is a gloss filter that divides the same band into four channels, and the acoustic features of the input voice are extracted through this gloss filter.

These two types of filter groups (filter bank and gross filter) are configured, for example, as fourth-order cyclic digital filters. Then, for example, the filtering output is obtained every 10 m5ec. still,
Control of this sound processing section is performed using a microprogram method.

Therefore, the preprocessing/recognition unit 19c is a high-speed processor 19.
d, a pattern matching processing section 19c, a word dictionary memory 19f1, and a buffer 7 memory 19g.

The buffer memory 19g receives the audio signal filtered by the audio processing unit +9b, and stores, for example, 1.8 seconds worth of audio data at maximum. The high-speed processor 19d processes the data stored in the buffer memory 19g by performing sound section detection, resampling, labeling, recognition processing using a transition network,
and executes its comprehensive logic judgment process. The high-speed processor 19d also performs communication with the host computer and controls the overall operation of the voice recognition section 19.

The pattern matching processing unit 19c performs matching processing such as composite similarity calculation on the speech data processed by the high-speed processor 19d with standard pattern data of word speech registered in the word dictionary memory 19r,
We are looking for recognition candidates.

For example, speech words to be recognized are uttered discretely.

Therefore, the high-speed processor 19d detects the input section of word speech using input speech energy calculated every 10 m5ec during acoustic processing, for example.

Specifically, as shown in FIG. 16, a threshold E is adaptively calculated from the background noise level and the input audio level, and the input audio signal level exceeds the threshold Eo for a certain period of time or more. , the point in time when the threshold value E is exceeded is detected as the start point S of the spoken word. Thereafter, the level of the input audio signal reaches the threshold E. When the threshold value E continues to be lower than the threshold value E for a certain period of time or more. The point in time when the value falls below is detected as the end E of the spoken word.

By the way, speech recognition can be considered as a type of pattern recognition. However, there are variations in the sound pattern, individual differences due to the speaker's gender, shape of the vocal organ, vocalization method, etc., as well as noise generated by the speaker himself and the surrounding environment. In this case, there are problems with level differences and noise caused by going through public telephone lines. Therefore, the problem is how to take these into account, absorb the above-mentioned variable factors, and produce spoken words 1 accurately and stably.

Therefore, the preprocessing/recognition unit 19c employs a recognition method called a hybrid structure matching method, which combines the pattern matching method and the structural analysis method in two stages.

That is, when a word speech section is detected as described above, the speech section (S, E) is first divided into 15 equal parts, and each of the 16 points is set as a resample point. Then, the spectrum at each resample point is extracted from the 16 channels of audio data (spectral time series) that have been acoustically processed as described above. Note that if there is a deviation between the sample point of the audio data and the resample point, the spectrum of the nearest point to the resample point may be extracted.

By this resampling process, 16X 1B (-256
) dimension speech pattern vector X is determined. That is, when the j-th (j-1, 2, 3, ~ 1B)-th glue sample point is r, the subvek J of the 16 channels at r is
J torque data as S(,)−(S,S,~S,)rJ
Obtain lrJ, 2rJ, l[irJ, respectively, and rearrange these S, irJ to obtain X-(S S -S -3)Llrl,
lr2. 2r1. 1f3rl[i
Find the vector X of the voice pattern. However, t indicates the transposition of the matrix.

Input speech pattern vector X obtained in this way
The degree of similarity between this and the standard pattern of the li-word sound registered in the li-word dictionary memory 191' is calculated by, for example, a composite similarity method.

Here, the standard pattern of word sounds pre-registered in the word dictionary memory 19 is, for the word category ωk, (ψ ψ ~ ψLk) 1 to '2' (λ λ ~ λ) 1k '2 to' Lk However, , (λ ≧λ ≧~≧λLk) 1k 2 is prepared. Furthermore, ψ λ is the category, ni'
, are the eigenvectors and their eigenvalues in the dispersion matrix I (of the pattern vector +1 X II in the formula
is the norm of vector X.

Such composite similarity calculations are performed for all categories, and the top similarity value and the resulting category name are determined as a pair.

Pattern matching using such a composite similarity method enables recognition processing that eliminates many pattern variations. However, for similar patterns or patterns with added noise, the similarity value of 2 may become small between different categories.

Therefore, as mentioned above, the following structural analysis method has been introduced as a supplement to the pattern matching method. This structural analysis performs recognition processing based on the differences in the sounds that make up word speech, and consists of two types: a phoneme label series and an acoustic feature series.
It uses two time series.

That is, the phoneme label sequence is 10ffl from the input audio signal.
The degree of similarity with the phoneme dictionary is calculated using the spectrum of 16 channels calculated for each scc, and phonemes having a degree of similarity of a certain value or more are labeled and determined. Note that this 'g element label is composed of six types, for example, five vowels and a nasal sound.

At this time, the phoneme dictionary is divided into male and female voices! ?
It is better to have X in stock.

Compared to vowels, which are pronounced relatively stably, it is difficult to label consonants individually as phonemes. Therefore, the consonant is labeled with its acoustic characteristics,
This is taken as feature information. Specifically, acoustic features are extracted from the output of a four-channel gross filter and audio energy obtained through acoustic processing. The acoustic features extracted and labeled in this way are, for example, the 17th
As shown in the figure in comparison with the characteristics of the output of the gloss filter, there are 12 types such as silence, voicelessness, friction, bursting, and energy dip. .

Thus, the phoneme/acoustic label sequence obtained for the input speech is as follows over the range including the speech section (S, E):
The information is input to a transition network created for each word category, as shown in FIG. 18, for example.

The presence or absence of the specified phoneme label or acoustic feature is checked for each node of this transition network. Then, if there is no one, reject it, and if there is one, move to the next node,
When the feature sequence ends, the input sequence that has reached the goal of the transition network is accepted, and its category is determined. Note that the direction of sequence checking can be selected from forward to reverse for each network.

The comprehensive judgment logic is a logic that synthesizes the candidate categories ordered by pattern matching as described above and the recognition results obtained by the transition network, and makes a final judgment.

That is, in this comprehensive judgment logic, when the maximum similarity obtained by pattern matching is set to 81, this is compared with a predetermined threshold value θ. In the case of (Sl×θ), this is rejected as noise.

In addition, in the case of (S1≧θ), another threshold value Δθ is used to
Categories with a degree of similarity greater than or equal to Sl−Δθ) are extracted as candidates. And the fin of the extracted category is 1
If so, this is extracted as the age recognition result. Also?
When k fi categories are extracted, only the categories accepted by the transition network are extracted with reference to the analysis results by the transition network. Then, the category with the greatest degree of similarity is determined as the recognition result.

In addition, if the categories extracted by threshold processing do not include those accepted by the transition network,
It is assumed that it is impossible to judge.

I-! As described in step 2, the pattern recognition processing result based on the composite similarity method and the recognition result using the transition network are integrated to recognize the input word speech.

Figure 19 shows the flow of the word speech recognition processing procedure in this speech recognition unit. After speech segment detection processing, resampling processing is performed to perform pattern matching, and at the same time, labeling processing is performed to check using a transition network. After that, these recognition results are integrated to perform comprehensive judgment logic processing. Such processing is executed under a processing sequence by the high-speed processor 19d.

By the way, in order to recognize words in continuously uttered speech instead of discretely uttered Qi speech, the following procedure may be used. That is, in this case, the input speech may be divided into various sub-intervals, and word similarity may be determined by performing rltl recognition for each sub-interval.

Specifically, for example, as shown in FIG. 20, all analysis frames in an input speech section are set as boundary candidates for partial sections, and the input speech section is divided into a plurality of subintervals. At this time, since the maximum time length D and the minimum time length D can be set for the duration of the word to be recognized, only a partial section ln within the range needs to be targeted for recognition processing.

In the example shown in FIG. 20, two partial sections are calculated assuming that the number of words in the continuously uttered voice is two. However, since the number of words in the input speech is generally unknown, it is sufficient to assume that 2 to n words exist as word candidates and to detect each partial interval. Then, calculate the lj word similarity for each detected subinterval, compare the connected relationships of the similarity results with each other to find the boundary of the most reliable subinterval, and calculate the subinterval separated by the boundary. What is necessary is to obtain the recognition results for each Il1 word.

However, when subintervals are obtained in this manner and 411 word similarity calculations are performed, the number of subintervals becomes enormous, which hinders speeding up of processing. Therefore, in practice, considering the speedup of processing, for example, the input (number of words is 2 to 51
11 words, duration of one word 128-640ffls
The ratio of f1 word length in ec11 utterances is 2.5 or less,
It is a good idea to limit the frame period to 16IOsec (extract one word out of every two in a period of 8IIlsec) and detect partial sections.

In this way, it becomes possible to effectively recognize each word in continuously uttered speech.

By the way, learning of the dictionary (word dictionary) used for such speech recognition processing is performed as follows.

This learning process is roughly divided into (1) a process for finding the characteristic kernel from the vowel pattern and the consonant pattern, and (2) a process for finding the eigenvalue and eigenvector for the characteristic kernel. Then, N eigenvalues and eigenvectors are determined in descending order of the eigenvalue. This process is generally called KL expansion.

First, the process of finding the characteristic kernel will be explained. The characteristic kernel of the input speech pattern (learning pattern) is found as follows, where S is the vertical vector of the learning pattern.

Here, 5-(S S -S)t m Ill'm2' llIn Note that this learning pattern S is given as a 64-dimensional vertical vector in the case of the consonant pattern.

In the case of a vowel pattern, it is given as a 16-dimensional vertical vector.

Therefore, the characteristic FmK is created by multiplying the vertical vector S of m learning patterns by the horizontal vector S obtained by transposing this vertical vector S.
It is obtained by averaging each component of the m matrix over the m learning patterns. Therefore, the number of elements of the characteristic kernel is the square of the number of elements of the vector.

Note that in order to obtain the characteristic kernel K that reflects the pattern distribution of the category through such processing, a sufficient amount of learning patterns are required. For this reason, it is necessary to store a predetermined number of learning patterns in the learning pattern memory in advance.

However, in the case of vowels, it is sufficient to QQ learning patterns of at least 6 categories in 16 dimensions, but in the case of consonants, there are 101 categories and it is necessary to obtain data in 64 dimensions. For this reason, it is undeniable that a huge amount of memory capacity will be required if the system is left as it is.

Therefore, in order to obtain a characteristic kernel K that reflects the pattern distribution using a small number of learning patterns, we perform the following process of updating the characteristic kernel, and gradually improve the characteristic kernel to reflect the pattern distribution through sequential calculations. Make it.

That is, the arithmetic process '+wSS L n is repeatedly executed on K-. However, W is a floating coefficient at the time of updating the characteristic kernel. This weighting coefficient W takes a positive or negative value, and if it is positive, it increases the degree of similarity to the human cover pattern of the characteristic kernel matrix, and if it is negative, it has the effect of decreasing the degree of similarity.

Further, ′ indicates the characteristic kernel before learning the learning pattern S, and K indicates the characteristic kernel updated by ■ learning of the learning pattern S.

After that, for the characteristic kernel obtained in this way,
Processing is performed to obtain the eigenvalues and eigenvectors, and based on these eigenvalues and eigenvectors, a standard pattern used for the above-mentioned N-combination similarity inverse calculation is created.

The standard pattern is obtained by performing KL expansion on the characteristic kernel, for example, by performing KL expansion using power east.

Now, the characteristic kernel K has eigenvalues λ, λ2. ~λ has I
n, and the corresponding eigenvectors ξ, ξ2. 〜ξ. In this case, the arbitrary vector U is
The above-mentioned unique “vectors ξ , ξ , ~o
The linear combination of 1 2 ξ is expressed as U ″ Σ ai ξI. At this time, since the relationship Kv,−Sen ξ1 holds, it becomes.

Here... 〉1λ 1 1λ 1〉1λ2 I〉 1 n[λ, /λ
l ] > 1 (i=2, 3. to .n), so when S becomes sufficiently large, the second term in the above equation converges to 0.

Therefore, the above formula K u mal λ1 ξ1 It can be made into a stone statue.

sol This shows that the ratio between (K u ) and (K u ) is the eigenvalue λ1. Also (K8
u) is shown to be proportional to the eigenvector ξl.

By the way, in a calculation process based on such a theory, the results often scale out immediately during the calculation. Therefore, let U be an arbitrary vector, for example, the 111th-order vector, and use v8. , -Kusu = (v)/(b)sol
sol sol(sa-0, l, 2...
・) Execute the calculation a. Here (b)s
ol is the element with the largest absolute value of the vector (V) and so
l Yes. At this time, u = (v)/(b)sol
sol 5o1=(Ku)/(b) s sol? =(Kv)/(b-b)s
sol s S + t − (K u) / (b ...... b ) o
Since sol s , from this, λ , b , ξ .

l sol I Finding U becomes n ■ ability.

sol Once the eigenvalue λ1 and the eigenvector ξ1 with the largest absolute values are obtained in this manner, the eigenvalue λ2 and the eigenvector ξ2 with the next largest absolute values are similarly obtained.

Now, if we consider -λlξ1ξ1 for K' -, then from ξ11ξl-0 (i-2, 3, ~, n), for K' ξ - we have ξl-λ1ξ1ξl ξl■ -λ1ξl-λlξ, -0(+-1) K′ ξ − to ξ, −λ ξ, ξ, L ξ1
1 1 l + 1-λ1 ξ1
(1≠1). Therefore, ' in the above is 1λ21〉...〉;λr 1〉...〉1λn 1〉
It can be seen that it has an eigenvalue of 0. It is assumed here that ξ1 has been normalized.

Such processing can be achieved by repeatedly performing the above-mentioned processing on the characteristic kernel converted into '-K-λξ fist ξ 1'. Through this processing, eigenvalues with large absolute values and corresponding private vectors are sequentially determined, and dictionary learning is performed.

FIG. 21 shows a dictionary learning process executed based on such a calculation algorithm.

Next, the character recognition section 21 will be explained.

This character recognition unit 21 is composed of a first character recognition block that recognizes characters read by a scanner or the like, and a second character recognition block that recognizes character information input online via a tablet or the like. .

For example, as shown in FIG. 22, this first character recognition block includes an image memory 21a that stores image data read and input by a scanner or the like;
An area detection unit 21b detects an area in which a character to be recognized is written from among the image data stored in the image memory 21a, and character data to be recognized from among the image data stored in the image memory 21a according to the area detection result. The character extraction unit 21c extracts each mark character pattern of the recognition target character registered in advance in the standard pattern dictionary 21d, and l
, and an identification section 21e that performs character recognition by individually comparing the character patterns extracted by the td character extraction section 21c.

This character recognition block is set at a predetermined position on the FAX transmission manuscript paper 211' as shown in FIG. 23, for example, and recognizes the characters written in a character frame 21g in which the transmission destination is written. The manuscript paper 21f' on which such a transmission destination is written is used as the first manuscript (one sheet 11) when the transmission manuscript consists of a plurality of sheets. The image data read and input from this single-sheet original is stored in the image memory 21a for character recognition processing.

The area detection unit 21b detects the character frame 21g from the predetermined format information of the FAX transmission manuscript paper 21r.
The system obtains position information and detects the area where the character to be recognized is written. The character extraction unit 21c uses this area detection information and the projection pattern information of the image information to individually extract image data of the characters written in the character frame 21g, for example, as shown in FIG. 24. There is.

As disclosed in, for example, Japanese Patent Publication No. 12778/1983, the identification unit 21e extracts the characteristics of the character pattern from the extracted character image, and identifies the extracted character pattern and the characteristics registered in the standard pattern dictionary 21d. Pattern matching is performed with the standard pattern of each character. Then, through this pattern matching, the character category of the standard pattern that has been successfully matched is determined as the recognition result.

It goes without saying that the pattern matching method can be modified in various ways.

By the way, the second character recognition block that recognizes character information input online via a tablet or the like is configured as shown in FIG. 25, for example.

This second character recognition block includes a coordinate detection circuit 21h that sequentially detects a series of position coordinates indicating the writing stroke of a character input online via a tablet or the like.

The time series data of the position coordinates detected by the coordinate detection circuit 21h is input to the preprocessing circuit 21i, and after removing minute noises such as detection errors in the tablet 4,
The coordinate series storage circuit 21j sequentially stores the coordinates and subjects them to character recognition processing. Incidentally, in this preprocessing circuit 21i, for example, 1
When a number of characters are input, the size of the characters is normalized.

Further, the stroke number detection circuit 21 detects the number of strokes, that is, the number of strokes of the character pattern, from, for example, breaks in the writing strokes (separations in the time series of position coordinate data).

Accordingly, the recognition processing unit 21n selectively extracts a standard pattern with the corresponding number of strokes from among the mark patterns of the recognition target character category registered in the mark character pattern memory 210 according to the information on the number of strokes. . The characteristics of each stroke of this standard pattern and the coordinate series storage circuit 2tj
The stroke characteristics of input character patterns stored in the input character pattern are compared with each other (matching process). Answer determination circuit 21
p determines the matching processing result, and obtains a recognition target character category having a stroke that corresponds to the stroke characteristics of the input character pattern as the recognition result.

In other words, e) the input character pattern is recognized by matching the characteristics of the stroke with the stroke characteristics of the mark character pattern according to the characteristics of the written stroke of the character pattern input on the line.

Note that information on positional coordinates of end points, intersection points, break points, etc. when a written stroke is approximated by a broken line may be used as the stroke characteristics.

By the character recognition unit 21 having the functions as described in 1- below,
Character information read and input via a scanner or the like or character information input online via a position coordinate input device such as a tablet is recognized.

Next, the figure recognition section 22 will be explained.

This figure recognition section 22 is configured as shown in FIG. 26, for example. The input unit 22a stores, for example, a captured and input graphic image, and subjects it to graphic recognition processing.

The contour tracing unit 22b divides the tracing direction of the line segment into eight directions, for example, as shown in FIG. There is. Specifically, for example, as shown in FIG. 28, a triangular figure is tracked clockwise, and the information on the direction of the tracking is stored as, for example, r1.2. ~2,3.4. ~4,5,7.
~7'' is obtained as a series of direction codes.

The segmentation unit 22c extracts singular points, such as curved parts, from the sequence of direction codes obtained in this way, and divides the contour of the guard figure into a plurality of characteristic parts according to the singular points.

The matching section 22d recognizes the input figure by matching the information on the figure outline segmented in this way with the feature information of various figures registered in the dictionary memory 22c.

For example, when the figure shown in FIG. 29 is given, from the series of direction codes obtained by contour tracing, for example, three mutually adjacent contour points (+-1), (+
) and (11), the sum of the direction codes is obtained in order, and this is smoothed as the direction code at the central contour point i. This smoothing process removes noise components.

Thereafter, the segmentation unit 22c detects end points that are characteristic points of the contour, that is, points where the curve is steep, and divides the contour around the end points.

Then, each divided contour portion is compared with the dictionary memory 22c to obtain the recognition result.

Through the above processing, as shown in Figure 30, nine figures have no end points, triangular figures have three corner points detected, and quadrilateral figures have four corner points. Each figure is identified and recognized. At this time, information such as that each of the end points is convex or that the contour connecting the end points is a straight line or a curved line may be used for figure identification.

On the other hand, the image recognition section 23 is configured as follows.

FIG. 31 shows a schematic configuration of this image recognition section 23, which includes an original image memory 23a, a binarization device 23b1
It is composed of a processed image memory 23C, a thinning device 23d1, and a code conversion device 23e.

The image memory 23a stores a given image to be recognized, and the binarization device 23b binarizes it and stores it in the image memory 23c.

This binarization level is variably set, for example, while monitoring the binarized image on a display.

The line thinning device 23d performs line thinning processing on the binarized image and converts the image into a line figure. The image memory 23c is rewritten with the thinned image and subjected to recognition processing.

The code conversion device 23c is configured, for example, as shown in FIG. 32, and first divides the thinned image into a plurality of segments in a segment dividing section 23f'. This segment division is performed, for example, by dividing the line figure at its end points, branch points, or intersections. Curvature converter 23g
calculates the curvature of each of the segments thus divided.

Straight line/curve dividing section 23h1 curve dividing section 231. The inflection point dividing unit 23j and the inflection point dividing unit 2311 further divide each segment divided as described above according to the information on its curvature, and thereby divide the segments into inflection points, switching points between straight lines and curves, Points of inflection, points of radius change in the curve, etc. are detected. Through such segment division and feature point detection, information on each part constituting the image line figure is extracted.

The approximate information creation unit 23m synthesizes information on these divided segments and feature points in the segments to express the image figure, such as the position coordinates of the starting point and ending point of each segment, and the type of the segment. Obtain code information that specifies.

For example, when an input image is given as shown in FIG. 33(a), the image line figure 23n in the input image
is thinned and extracted, and divided into segments as shown in FIG. 3(b). In this example, an image line figure 23n in which a circle figure and a square figure are so-called skewered by straight lines is input. As shown in FIG. 33(b), the image line figure 230 of the lever is divided at the intersection, and segmented into two semicircles, two U-shaped figures, and four straight lines.

The curvature converting section 23g calculates the curvature of each divided segment as shown in FIG. The refraction point dividing section 23j and the inflection point dividing section 23h detect feature points of each segment from the curvature change points. Specifically, in the example shown in FIG. 34(a), the curvatures at the bending points of the two straight lines increase sharply, so it is possible to detect the bending points from the change in the curvatures. Also, Fig. 34(b)
In the example shown in , since a change in curvature is detected at the transition portion from a straight line to a curve, the feature point can be detected from this change in curvature.

Similarly, in the examples shown in FIGS. 34(c) and 34(d), it is possible to detect the feature point in the segment from the point of change in curvature.

In this manner, the image recognition unit 23 segments the given image figure and detects the feature points of each segment. Then, the image line figure is approximately expressed and recognized as code information indicating each type of segment of the number of stages and its position coordinates.

Now, the voice verification section 17 is configured as follows.

This voice verification unit (γ) recognizes the speaker who input the voice as an individual (
For example, it is configured as shown in FIG. 35.

That is, the speech given through the speech input section 17a is filtered by the phoneme filter 17b and the personal filter 17c, and the speech characteristics are extracted.

Each band of the channels of the number of stages of the phoneme filter 17b is set by equally dividing the sound frequency band, for example, as shown in FIG. 36(a).

The phonological filter 17b having such filter characteristics extracts feature parameters indicating the phonological characteristics of the input speech. Note that the bandwidth of each channel may be set by dividing the audio frequency band logarithmically.

On the other hand, each bandwidth of the index channel of the personal filter 17c is set by dividing the audio frequency band exponentially, as shown in FIG. 36(b). With the personal filter 17c having such filter characteristics, the voice characteristics from the low frequency range to the middle frequency range of the input voice are
This feature is extracted more often than the features on the high frequency side. The filter output of each of these channels is obtained as a feature parameter for personal verification.

Therefore, the word recognition unit +7d has the phonological filter 17b.
The word indicated by the input speech is recognized from the phonetic feature parameters obtained through the phonological feature parameter with reference to the word dictionary 17e. This word recognition function is performed by the voice recognition unit 1 mentioned above.
9, and the function of the voice recognition unit 19 may be used as is.

Then, according to this word recognition result, a dictionary to be used for personal verification of the personal dictionary 17 (') is selected.This personal dictionary IN
is a result of the analysis by the personal filter 17c of a specific word uttered in advance by an individual who is a partner for speaker verification, and is classified and registered for each word.

The speaker matching unit 17g calculates the degree of similarity between each feature parameter of the corresponding word selected from the personal dictionary 171' and the feature parameter of the input voice found in the personal dictionary 17c, and The degree values are discriminated using predetermined threshold values. Then, by comparing these discrimination results with each other, for example, a personal category with the highest similarity value and a feature parameter with a sufficient difference from the next highest similarity value is determined to be the speaker of the input voice. Personally identified.

Here, the characteristics of the personal filter 17e will be explained in more detail.As mentioned above, the characteristics are set to be different from those of the phoneme feature filter 17b. Considering the distinctiveness of the individuality of this voice, the distinctiveness can be evaluated, for example, by the F ratio given as F ratio - (inter-individual variance)/(intra-individual variance).

Now, when considering the F ratio of each channel output of the filter characteristics set in the phonetic filter 17b, an exponential trend is shown as a solid line in the diagram T537.

For this reason, in the past, individual verification has been performed exclusively using voice feature information on the high frequency side.

However, if it is possible to identify individuals using voice features in the entire frequency band rather than using only high-frequency features of the voice, it is thought that the matching accuracy will improve significantly. That is,
If the value of the F ratio is between 1 and 1 in all frequency bands, and the inter-individual variance is 1-times the intra-individual variance, more accurate individual matching becomes possible.

Therefore, here, as mentioned above, the personal filter 17c
The characteristics of the voice are determined exponentially, and the characteristics of the high frequency range where individual characteristics are prominent are roughly extracted, and the audio characteristics of the low frequency range are extracted finely by increasing the number of channels allocated to the low frequency range. That's what I do.

Specifically, since the change in F ratio of each channel shows an exponential tendency, a filter bank in which the bandwidth of the high frequency channel is increased exponentially compared to the bandwidth of the low frequency channel. and configure this as personal filter 17c
It is said that

According to the output of each channel of the filter 17C configured in this manner, the F ratio is as shown by the broken line in FIG. 37, and a significant improvement in the F ratio in the middle range is recognized. As a result, it becomes possible to carry out individual verification by actively utilizing not only the voice characteristics in the high frequency range but also the voice characteristics in the mid-range, and it becomes possible to improve the verification accuracy.

That is, in addition to the features used in the 111-word recognition of the input speech, the speech matching section 17 uses a filter bank to extract feature information that clearly reveals the individuality of the input speech.

As a result, the phonological recognition for input speech is I'1. It is now possible to perform personal identification of the speaker, that is, personal verification, with high precision.

Next, the speech synthesis section 2B will be explained.

The speech synthesis section 26 includes a discriminator 26a as shown in FIG.
, a decoder 26b1, a rule parameter generation device 26C1, and an audio synthesizer 26d.

The discriminator 2f3a determines whether the input code string is a character string or a No. 71 sequence indicating analysis parameters for speech synthesis. This information discrimination is performed, for example, by determining the sleuthing information added at the beginning of each input code string. When it is determined that the code string is an analysis parameter, the code string is provided to the decoder 26b, and it is decoded to obtain its phonetic parameters and prosodic parameters.

If it is determined to be a character string, the character string data is provided to the rule synthesis parameter generation devices 2B and 2c to generate its phonetic parameters and prosodic parameters.

The speech synthesizer 28d processes the sound source wave via the vocal tract approximation filter to generate a synthesized speech wave according to the phonological parameters and prosodic parameters thus obtained by the decoder 28b or the rule synthesis parameter generation device 28c. is being generated.

To further explain the rule synthesis parameter generation device 28e, the device 26c is configured as shown in FIG. 39. The character string analysis unit 28c refers to the language dictionary 26, identifies each word in the input character string, and obtains grammatical information such as accent information, word/clause boundaries, part of speech, and conjugation for the word. .

Then, phonological rules and prosodic rules are applied to the analysis results, and control information thereof is generated.

Here, the phonological rules provide information on the pronunciations of the analyzed words, realize phenomena such as rendaku and disuse caused by word concatenation, and generate phonological symbol strings.

The speech parameter generation unit 26g inputs this phoneme symbol string, sequentially obtains syllable variations from the CV file 211ih according to the syllable unit, and performs interpolation combination. The phonetic parameter generation unit 28g generates a phoneme parameter sequence from the 1-diagrammatic symbol sequence.

In addition, the prosodic rules determine the boundaries of utterances and breath positions according to grammatical information such as C1 word and clause boundaries, and determine the duration of each 1η, pause length, etc.

At the same time, this prosodic rule generates a prosodic symbol string that is based on the basic accent of each word and takes into account the clause accent. The prosodic parameter generation unit 261 inputs this prosodic symbol string and generates a prosodic parameter string representing a temporal change pattern of pitch.

On the other hand, when the input code string is a code string indicating analysis parameters for speech synthesis, the decoder 2Gb functions as follows.

That is, when the code string of the analysis parameter indicates the cepstral coefficient of the Cv file, the code string 26m is generally the fourth
As shown in figure 0, the parameters P (pitch) and C, C, ~
Bits are assigned to C (cepstrum coefficient) o I II and information is compressed.

Therefore, in the decoding W 26b, parameter conversion table 2 [
In is used to convert and decode the compressed analysis parameters into a number of bits suitable for the speech synthesizer 26d. For example, each parameter is converted to 8 bits, and a phonological parameter sequence (cepstrum coefficient) and its prosodic parameter sequence (pitch) are determined.

The speech synthesis unit u26d includes, for example, a voiced sound source 26q and an unvoiced sound source (M-sequence generator) 26'' as shown in FIG. , or shoreless sound source wave (P-0)
are occurring selectively.

This sound source wave is input to a preamplifier 26s, subjected to signal control according to the cepstrum coefficient C of the phoneme parameter, and input to a logarithmic amplitude approximation digital filter 28t. This logarithmic amplitude approximation digital filter 281 constitutes an oscillator circuit that approximates the vocal tract characteristics according to the cepstrum coefficients C, ˜C of the phonetic parameters, and filters the sound source wave. The filter 26L synthesizes and outputs speech data indicated by the phoneme parameters and prosody parameters.

Then, the signal synthesized by the logarithmic amplitude approximation digital filter 28t passes through the D/A converter 28u, and then L P
The signal is filtered through F26v and output as a synthesized audio signal (analog signal).

In the speech synthesis section 2B configured as described above, the speech indicated by the data series is synthesized according to the rules from the input data series and is output.

Next, the image composition section 27 will be explained.

As shown in FIG. 42, the image synthesis section 27 includes a control calculation section 27a, a display file memory 27b, an image synthesis circuit 27c, an image memory 27d, and, if necessary, a display 27c.

Note that this display 27c may be the display unit 10 provided in the 7fi workstation.

The image synthesis circuit 27 reads parameters of vectors, polygons, and arcs written in the display file 27b under the control of a dedicated control calculation unit 27a, generates line figures indicated by the parameters, and stores them in the image memory. 27
Writing to the specified address of d. The image generation function of the image synthesis circuit 27 constructs a designated line graphic image on the image memory 27d.

This line graphic image is displayed and monitored on the front display 27c under the control of the control calculation section 27a.

Further, the image generation circuit 27b has a special processing function for image generation and a filling processing function. This special processing function includes, for example, functions such as erasing hidden lines and clipping processing for overlapping image figures of a number of stages. The filling function consists of filling a partial area of an image figure with a specified color.

Through the functions of the image synthesis circuit 27b, various image figures are created and their synthesis processing is performed.

By the way, the synthesis of image figures generated as described above and natural pictures can be roughly divided into the following two types. One is a process of embedding and synthesizing an image obtained through arithmetic processing into a natural picture such as a landscape photograph, and the other is a process of embedding and synthesizing an image obtained by arithmetic processing into a natural picture such as a landscape photograph. It consists of the process of embedding and synthesizing a natural image within a two-dimensional image.

In the case of embedding an image in the former natural image, for example, as shown in FIG. This can be achieved by superimposing this on a natural image and compositing it. Then, in the image area to which the "transparent color" code is given, the information of the natural image will be displayed as is, and in the other areas, the figure generated by the control ap unit 27a will be displayed. As a result, image synthesis using a natural image as a background is realized. This technique is called overlay.

On the other hand, as the concept is shown in FIG. 44, a natural image may be written in the image memory, and a figure generated by the control calculation unit 27a may be written on top of it (in the front). . This method is called the two-buffer method,
This can be implemented relatively easily together with the overlay method described above.

By the way, in the latter case, in which a natural image is inserted and synthesized within a plane shown as an internal model of the control calculation unit 27a, high-speed processing is performed as follows.

The coordinate transformation necessary to embed the natural image on the shell in a space facing an arbitrary direction in three-dimensional space is given by the following equation.

CX0C5Y0CG CX0C5Y106 However, X and Y are the coordinates on the display surface, and u and v are the coordinates on the natural image.

If this coordinate transformation process is to be executed as is, six multiplications and two divisions will be required every time a pixel is displayed, resulting in an enormous amount of calculation and 21° calculation processing time.

Therefore, here, the calculation described in - is divided into two conversion processes and executed using the following intermediate coordinates (s, t). This arithmetic processing is executed at high speed using, for example, affine transformation.

u-Ca S+(22t+a3)/l (1)
vllllI(α7 s+α8 t+α9)/ls”C3
Praise X-C4Y (2) C4X
+C5Y+CG In other words, perspective transformation is performed using equation (1) above, and then two-dimensional affine transformation is performed using equation (2) to perform perspective transformation to an arbitrary plane at high speed. There is.

Here, since the denominator of equation (1) is the coordinate t itself, it is easy to perform the calculation at high speed simply by slightly improving the conventionally known affine transformation circuit.

In this way, the image compositing section 27 executes extensive image compositing processing at high speed.

Next, the output format selection section 24 will be explained.

The output format selection section 24 is activated upon receiving a media selection request signal, and selects which medium is to be used to output data. In other words, it selects which of various media should be used to transmit information.

FIG. 45 is a schematic diagram of the output format selection section 24, in which the media selection control section 24a, the input media determination section 24
b, a partner media determination unit 24C, a media conversion table 24d, and a self-media function table 24e. Also, FIG. 46 shows this output format selection section 24.
This shows the flow of processing. The functions of the output format selection section 24 will be explained along the flow of this processing procedure.

When the media selection request signal is given, the media selection control section 24a requests the control section 2 to provide input media information necessary for the media selection operation. Then, it issues a media information detection request and a media function identification request to the input media determining unit 24b.

The input media determination unit 241) is composed of a media detection unit 24 and a media identification unit 24g, and upon receiving an information request from the bipedal media selection control unit 24a, the input media determination unit 241)
It is designed to expose the input media given from the source and identify and determine the function of the detected media. This input media determination unit 24b determines that the input media is M.
If it is r, the function of the medium is determined to be ADPCM, etc.

Thereafter, the media selection control section 24a informs the control section 2 whether the destination of the data output is another functional block of its own terminal (within the workstation) or another function block connected via a communication line or the like. Determine whether it is a workstation or communication terminal.

When an instruction is given to output data to another workstation or communication terminal, the media selection control unit 2
4a requests the control unit 2 for identification information regarding the destination station. In response to this request, information regarding the partner station that outputs data is input to the partner media determining section 24c.

The partner media determining section 24c is a partner station identifying section 24h.

The partner station media identification unit 2411 is configured to include a function identification unit 24j, and operates upon receiving an identification information determination request from the media selection control unit 24a. Then, from the identification information for the partner station, first, at t14':f, the station is identified, and the media of the partner station is identified. Then, the function of the partner station's media is identified.

Specifically, for example, it is determined that the destination station to which data is to be outputted (transmitted) is an automatic FAX, its communication medium is an image, and its function is Gm type. The identification of this partner station is determined by the negotiation (
This may be done based on information sent using a handshake function. In addition, if there is no negotiation function, the media detection function is used by the function identification unit 24j.
It's good to have it in hand. In this way, it becomes possible to identify the function according to the media information signal from the other party.

FIG. 47 shows the flow of this partner station identification processing procedure. As shown in this flow, for example, it is determined whether the eight communication parties are telephones or not, and if they are telephones, FAX is sent.
Determine whether a signal arrives.

If the other party's station is a telephone and a FAX signal arrives, it is sufficient to identify this as the other party's device being a fax machine. Further, if it is determined that the device is a telephone and no FAX signal arrives, it may be determined that the other party's device is a normal telephone. Further, if it is determined that the device is not a telephone, it may be determined that the other party's device is a communication device other than a telephone.

When the media of the communication partner station is identified in this way, the media selection control unit 24a then refers to the media conversion table 24d configured as shown in FIG. 48, and selects the input media and input function. , phase T', equipment, F
11 Obtain media conversion selection information corresponding to the device media and the functions of the partner device.

For example, if the input media is for sound, its function is ADP.
If it is a CM and the destination device is a Gm type FAX, it must be the media or image of the destination device, and whether the main media conversion function is (voice) to (code character) (code character) to (image) The following are required. At the same time, it is required that the conversion function can be realized by (ADPCM, voice) to (GIII; FAX). At this time, if dependent media reduction information exists, this information is also obtained at the same time.

The media conversion information obtained in this way is sent to the control unit 2.
and selectively specifying the format of the data output.

Note that when data output is to be performed inside the own workstation, the media selection control section 24a refers to the self-meso9 function table 24e to find an output format in which the data can be output. According to this information, the media selection control section 24'a refers to its own media conversion table of the media conversion table 24d, similarly obtains media conversion information, and provides this to the control section 2.

According to the media conversion information obtained in this way,
For example, the above-mentioned phonetic synthesis section 26 may be used to convert text information given in a character code series to speech information and output as data, or the sound recognition section 19 may be used to convert speech information into character code series information. The data will be output after conversion.

Next, the database section 32 will be explained.

The database unit 32 organizes and stores various types of data such as codes, images, and VI voices, and provides this data to various application systems. FIG. 49 shows a schematic configuration of this database unit 32, which includes an interface unit 32a that executes command analysis processing, a data operation unit 32b that executes database search processing, etc., and a storage medium that stores various data. magnetic disk device 32e
, optical disk device 32d1 and its additional function section 320
It is composed of

Various data are classified and organized into relations of one ship according to the type of data, and a database is constructed by registering each relation.

The database unit 32 will be explained below by dividing it into four parts: its logical structure, stored data, physical structure, and additional functions.

The logical structure indicates how various data are stored when this database section 32 is viewed from the application system side. Here, data is handled as a logical 1+ structure according to the relational model, for example, as an image of a table as shown in FIG.

A table (relation) has several columns (attributes)
are provided, and predetermined +1-rank data is stored in each of these columns. Unit of data (Tarapur)
is defined as a set of values to be stored in each column. One lane is constructed by the tV-valued number of attributes that store such data.

However, in this model, data is stored in the database by specifying a relation name and giving values for each of its attributes. Database searches also specify relations and attributes, and determine whether the value stored there satisfies a predetermined condition with the specified value or the value stored in another attribute. This is done by extracting Tarapuls that satisfy the conditions.

This search condition is given as the values being equal, unequal, smaller, larger, etc. At this time, it is possible to designate search conditions for each ha number attribute and perform logical processing (AND, OR, etc.) on the results of the condition determination. In history, it is also possible to search a database by specifying multiple relations and finding a predetermined Tara pull from among multiple relations based on conditions such as the value of the attribute of one relation being equal to the value of the attribute of another relation. It is.

Furthermore, data deletion from the database is basically performed at t, 1, the same as the above search, but instead of extracting the data, it is performed by deleting the data.

Furthermore, data updating is also performed by changing the value of the specified attribute of the obtained tuple and storing it.

In addition, in each relation, information (person name, person in charge code), etc. of the person who is permitted to read, add, or change data for each attribute is entered, and data protection measures are taken. Note that instead of taking this data protection measure for each attribute, it is also possible to take this measure for each relation.

It should be noted that there may be more than one person's information listed here.

In the relation example shown in FIG. 50, the data is shown as a character string, but the data stored in each relation may also be a simple dot string.

In other words, the data stored in the relation may be not only character strings, but also image information, audio information, and the like.

Now, the data stored in this database includes the "personal schedule" shown in Figure 50 mentioned above, as well as the "address book" shown in Figure 51, for example.
“Individual work and its agents” “Operation history” “Human resources”
It consists of various sections such as ``Meeting 1 Lecture Room'', ``Conference Room Reservation'', and ``Meeting''.

As shown in this example, relations consist of those that are mainly used for personal use and those that are commonly used by many users. Personal glue rations are set up for each workstation used by each individual.
A common ration is provided at a workstation that is common to one (1) number of users.

Furthermore, the hardware of a Jl:tsu workstation is not necessarily different from other workstations! It doesn't mean depending. It goes without saying that a personal workstation may also serve as a common workstation. Furthermore, the number of common workstations is not limited to one, but may be provided in several units depending on the hierarchical level of the system. A common workstation is set up as a common workstation.

Here, the data (1η structure) of the "Personal Schedule" relation shown in FIG. 50 will be briefly explained.

This relation indicates that the relation name is "Personal Schedule J" and that it was created by "ΔΔΔΔ". This relationship creator “△△△△
” allows all data operations for the relation.

Also, according to the data protection function added to this relation, everyone is allowed to read data, but only "O○○O" and "persons belonging to the engineering department" are allowed to add data. ing. Incidentally, this "person belonging to the technical department" is determined by, for example, checking the ration of "r people'μ". Additionally, data changes are only permitted for those whose "person level" value is "5" or higher. this"
"Person level" refers to the person=lG relationship, for example (manager: 8) (deputy manager: 7) (section manager: 6) (+
5) to represent the position.

Attributes such as ``start time'', ``end time'', ``type'', ``name'', and ``location'' are set for this relationship, and data is written to each of them.

Next, the physical N4 structure for actually storing the above-mentioned six types of transcriptions in this database section 32 will be explained.

The information storage unit (storage unit) mainly stores data, and can read and write any part of the data at relatively high speed.The magnetic disk device 32c and the optical disk device 32g described above are not very expensive. used.

To store the database in this information storage section, the storage area of the information storage section is divided into sections of a specific size (for example, several kilobytes) (determined according to the tumble length and the speed of the computer), and each section is divided into sections. Managed as a vane. As shown in FIG. 52, for example, page management information is stored in the 0th page, relation list information is stored in the 1st page, and page information in use is stored in the 2nd page r6.

This list of relations indicates the location of various relations in the database.

For example, the actual data stored on the 9th and 11th pages is stored on the 5th page, and is sorted based on the ration attribute (main attribute) and according to the index page information stored on the 10th page. It has become. This index page information indicates in which page the number of attribute values is stored.

When searching data using an attribute other than the second main attribute, use the second main attribute for that attribute.
Via the sub-index of page 0, first the 21st
Obtain the sub-data shown on the page and the 22nd page. This sub-data contains only the attribute values and the above-mentioned main attribute values, and the actual data is determined using the attribute values determined here.

If the actual data is enormous, such as image data or audio data, and some bit errors in the data are not a problem, these actual data may be stored on another inexpensive device such as the optical disk device 32d. The file may be stored in a suitable information storage device. In this case, the 9th page or the 11th page
The actual data page such as the page may store information to that effect and the storage location information of the actual data in the device.

However, additional functions for the database constructed in this way consist of automatic disposal of unnecessary data3, for example. This automatic discarding of unnecessary data is performed by giving additional information to the relation, such as [discard; i+1/unii1] and [discarding method], and by operating an erase command for each relation at predetermined intervals.

Incidentally, erasing of the table can be performed by determining, for example, whether the end time of the meeting information is before the current time, or whether the end time is before the current time. Therefore, there is no need to add any special functionality to erase such a Tara pull.

Another additional feature is data security. This data preservation function prevents data from becoming corrupted or lost due to, for example, hardware failure or power outage. Specifically, this data preservation function is achieved by duplicating information and writing data onto magnetic tape.

In this way, the database section 32 sorts and organizes various data according to relations, manages them in 11 page units, and provides them to various application systems.

Next, the work environment data collection section 25 will be explained.

The work environment data collection unit 25 collects data on past operation history for the workstation, and provides operation guidance based on this data.

Here, the work environment data collection unit 25 has a command that associates a command corresponding to a function of the information processing system with a command corresponding to r,+j functions of another information system, as shown in FIG. 53, for example. There are tables available.

Specifically, when the information processing system is A1 and the other information processing systems B, C, D, etc., the commands of the other systems corresponding to the command "DELETE" in system A are "DEL", "ERASE", etc. ” ”REM
OVE", as indicated by the command correspondence table.

Figure 54 analyzes the command input by the user,
This figure shows a schematic configuration of a work environment data collection unit 25 that executes predetermined operations and various guidances.

In this work environment data collection section 25, first, the command inputted from the command input section 25a is sent to the command analysis section 25.
The command correspondence table 25c is analyzed by referring to the command correspondence table 25c. Specifically, it is checked whether the input command is registered in the command correspondence table 25c according to the procedure flow shown in FIG. That is, when a command is input, it is first checked whether the input command is from system A or not. When the input command is determined to be a system A command, the command resolution resolving section 25b supplies the input command to the command execution section 25d, and causes the command execution section 25d to execute a predetermined operation based on the command.

On the other hand, if the input command is not from system A, it is checked whether it corresponds to a command from another system, and if there is a command associated with it (if there is one, the corresponding command is displayed on the screen). It is displayed on the display section 25c.

In other words, if a command is used in another system (system B), for example "DEL", the corresponding command "DELETE" of system A is obtained and displayed on the screen display section 250 as operation guidance. I will do it.

Note that if the command corresponding to the input command does not exist in the command correspondence table 25e, the screen display section 2
A command error message is displayed in step 5c.

Specifically, the command input is processed as follows. Now, assume that a user who has experience operating systems B and C tries to operate system A (the information processing system). Here, if the user enters a command to erase data "ABC", conventionally system A
°DELET for data deletion according to the instruction manual of
It is necessary to search for the command E'' and enter it.

However, in this case, the user can use the data erasure command °ER that he used in System C according to his past experience.
Input ASE ABC' as shown in FIG. 56(a).

Then, the work environment data collection unit 25 analyzes this input command, obtains the system A command "DELETE" corresponding to the input command "ERACE" from the command correspondence table 25e, and displays this as a guide. As a result, even when operating System A for the first time, the user knows that the data deletion command is "DELETE" and can erase the data by inputting that command according to the guide. It becomes possible.

In addition, in order to display a list of file names,
) command “DIR” in system B as shown in
When inputting ``CATA'', the corresponding command "CATA" in the system A is similarly determined and displayed as a guide. As a result, following this guide the command “CATA
” will display a list of file names.

By utilizing the functions of the work environment data collection unit 25 in this manner, by inputting a command used in a system with which the user has past operating experience, corresponding commands in that system are displayed as a guide. Therefore, system users are able to operate the system by making maximum use of their past knowledge of fW. Then, it becomes possible to easily know the commands of the information processing system.

Therefore, the user is freed from the trouble of checking the operating manual of the information processing system each time. Therefore, effects such as being able to significantly shorten the time required to learn how to operate the system can be expected.

Incidentally, when a command corresponding to an input command is obtained and displayed as a guide, the command may be executed upon receiving a judgment input of its pass/fail.

That is, the procedure flow is shown in FIG. 57, and the display example is shown in FIG. 58.
ERASE", and when the corresponding deletion command "DELETE" of system A is requested, it is checked whether this is correct or not.

When an instruction input of 1E(Y) is received, it is determined that the input command is smaller than "DELETE", and this is sent to the command execution unit 25d to execute the process.

In this way, the correspondence between commands is guided and at the same time the desired process is executed in accordance with the input command, so there is no need to input a new command. In other words, the process of automatically converting an input command into a corresponding command is executed. Therefore, it is possible to improve the operability in the future.

1. The corresponding commands depend on the type of system (see 11).
f=R&i may exist. In short, it is sufficient to store them in association with each other in the command correspondence table 25c. It goes without saying that the command is not limited to the character string format described above.

Next, the collection of system proficiency data by the work environment data collection unit 25 will be explained. Inside the work environment data collection unit 25, an external storage device and a control device are placed as hardware for executing this system proficiency data collection process.

FIG. 59 is a flowchart showing the system proficiency data collection process.

When a user enters their identification code (user number, password, etc.)
When inputted, the work environment data collection unit 25 obtains the proficiency skin chart corresponding to the identification code from the external storage device and sets it inside the device. This proficiency table stores the degree of proficiency of each user with respect to various functions of the system, and is structured as shown in FIG. 60, for example.

In other words, this proficiency surface is the usage 1M for each usage function.
On the other hand, it is composed of information such as the proficiency class for the function declared by the user at the time of the last publication date, the proficiency class when the function was used last time, and the complexity of the function.

Here, the degree of complexity increases as the function to be used requires specialized knowledge, and as the function becomes more advanced than the basic function.

However, such a proficiency table is set up for each user,
Each is stored in an external storage device.

Note that for a user who uses the system for the first time, a learning curve for that user is created by setting a new identification code, and is registered in the external storage device.

Note that, in addition to the above-mentioned proficiency chart, messages for each function used corresponding to the proficiency class are registered in the external storage device, as shown in FIG. 61, for example. The lower the proficiency level of the message, the easier it is to understand the message, including the background explanation. In addition, classes with a high level of proficiency have advanced content that includes simple explanations and introductions to specialized functions.

Also, proficiency classes are, e.g. A; Riverside class B; Intermediate class C; Expert class The classification is set as follows.

When the proficiency table corresponding to the input identification code is obtained, a menu is displayed to allow the user to select the function to be used. For this menu, the user inputs, for example, a number corresponding to the function to be used.

Then, the control device determines whether the input information is the end r signal or the selection signal of the function to be used, and if the function to be used is selected fj, the operation is as follows.

That is, when a usage function selection signal is input, first, the proficiency chart for the user is referred to, and information such as usage frequency, last usage date, declared proficiency class, etc., corresponding to the selected usage function is determined. is required.

Then, weighting is performed according to this information, and the current proficiency class is determined.

To determine this proficiency class, for example, the usage frequency is Pl, the last usage date and time is T, the current usage date and time is T, , f
llll child declaration proficiency class 1st c
1. When the usage proficiency class is X2@fA, B, CI, the complexity is P 1, and the discriminant function is F, C "F -, P, +2 (To-To) + Ka c
+ [x + ] + 4 G2 [X9] + 5P.

required as. However, in equation 1, 2°K,
4 is an evaluation of A, B, and C, which are set to appropriate values through experiments and the like. These 5.0 price market values are set at appropriate values through experiments, etc., based on the relationship Y x Y x Y, Zl x Z2 x Z3.

Here, G [X, ] means that it takes the value Y when X, -A, and takes the value Y2 when X 2 "" B. Moreover, (T-T') is the number of units used from the last used Ce from 1 o'clock in the month of the year to the present time converted into time.

The class determination is performed in the following manner based on the value of the discriminant function Fr described above.

F <N...A class I N≦F<N...B class l r
2 N2≦F, ...C class Furthermore, the judgment thresholds N and N are appropriate based on experiments, etc.!
2.

When the proficiency class is determined in this way, a guide message or an error message corresponding to the determined proficiency class and corresponding to the specified function to be used as described above is obtained from the external storage device.

Thereafter, the currently determined proficiency class is compared with the previous proficiency class stored in the proficiency table. If there is a change in the proficiency class, a message indicating that the proficiency level has changed is added to the guide message and written.

This proficiency class change message consists of four types of messages as shown in FIG. 62, for example. Then, it is determined according to the form of the class lady and displayed together with the guide message, etc.

The user performs processing operations according to the various messages displayed in this manner.

Specifically, when it is determined that the user is in the beginner class (class A) for the function used to store created data in a file, a message as shown in FIG. 63 is displayed.

If the user makes a mistake in inputting information despite this message, an error message as shown in FIG. 64 is displayed, and guidance on how to use the function is provided.

Further, if the user's proficiency level is determined to be intermediate class (B class), the message shown in FIG. 65 is displayed. If the +1 user makes a mistake in inputting information despite this message, an error message as shown in FIG. 66, for example, is displayed, and instructions for operating the function to be used are provided. Similarly, if the user's proficiency level is determined to be in the expert class (C class), a message as shown in Figure 67 will be displayed, and if there is an error in the information input, a message as shown in Figure 68 will be displayed. An error message as shown in the figure is displayed to guide the operation of the function to be used.

When data is entered into the blank field of the guide message displayed as described above, the control device adds (+1) to the corresponding usage frequency of the proficiency table of the corresponding user as described above, and Updates will be made at 1 o'clock in the month of use and the previous use familiarization class. Then, the system prompts the user to execute the function, assumes that the function has ended, and returns to the menu display operation for selecting the ill function described above.

If +lr usage function selection signals are input here, the above-described process will be executed +lr times.

However, when the end r selection signal is input, the proficiency table created and updated as described above is written into the proficiency file of the external storage device together with the identification code of the corresponding user, and this is saved. Then, the series of processing procedures is completed.

In this way, the work environment data collection unit 25 collects data on the proficiency level regarding the L'-work of the system.
It provides guidance to the AUJ in its operations according to the collected data.

The above is the basic f+'ls configuration and its functions of this workstation.

Next, the secret communication 3 functions achieved in the workstation configured as described above will be explained.

Telephone communications at the workstation are carried out by communication device 12.
．． The audio data inputted through the audio input section 5 is transmitted to other workstations or 7U phone terminals connected via the audio input section 13, and the received audio data is transmitted through the audio output device 9. This can be achieved by outputting the voice. Therefore, & voice input section 5 is a transmitter, voice output device 9
will be used as a telephone receiver.

Therefore, when a command to select the secret call function is input from the keyboard unit 6, or when a switch to select the secret call function is turned on, the voice synthesis unit 26
is started. By activating the speech synthesis section 26, information indicated by the given character code string is synthesized into speech, and preparations are made for transmitting the speech in place of the sounds input from the speech input section 5.

In this state, when character information is inputted using the position input device 4, for example, a tablet, the series of input position coordinate data is given to the character recognition/ia section 21, and the input character pattern is recognized. The character code string indicated by this recognition result is provided to the speech synthesis section 26 for speech synthesis. Therefore, by handwriting inputting telephone information into a tablet, the handwritten information will be synthesized into speech and transmitted.

On the other hand, if you want to notify numerical data etc. written on a form etc. by telephone without being heard by a third party, set the form etc. in which the information is written on the image input device 3, and enter the information into the image input device 3. do.

Then, the input image information is sent to the image recognition section 23.
For example, the character pattern is subjected to verification processing, and then the character is recognized. The character code string indicated by the recognition result of the sentence t' is then given to the speech synthesis section 26, where it is synthesized into speech and transmitted.

Note that, without using such a character recognition processing function, communication information may be directly input as a character code string from the keyboard B, and this may be voice-synthesized and transmitted.

(If such a secret call function is used, the person making the call does not have to say the information he or she wants to communicate over the phone into the handset, so the content can be heard by a third party nearby.) In addition, since the information at the other end of the call is reliably output using synthesized voice, there are no inconveniences such as, for example, the fact that it is difficult to hear what is being said because it is spoken in plain white as in the past. There is no invitation.

Therefore, it is possible to reliably notify the user by telephone of information contents that the user does not want a third party to hear.

Note that the present invention is not limited to the embodiments described above. Although the secret call function has been described here as one of the functions possessed by the workstation, it is also possible to provide this function in a general conversation terminal. In this case, for example, an old-fashioned character input section using a tablet, its character recognition section, and a voice synthesis section may be incorporated. It is also possible to switch between the input voice and the synthesized voice by switching the input using a switch in the transmitter. In addition, the present invention can be implemented with various modifications without departing from the gist thereof.

[Effects of the Invention] According to the present invention, characters input by hand or images are recognized, a character code string indicated by the recognition result is synthesized into speech, and this is transmitted in place of input speech. Therefore, it becomes possible to effectively notify by telephone of information that the speaker does not want third parties around him to hear.

[Brief explanation of the drawing]

The figures show one embodiment of the present invention, and FIG. 1 is a diagram showing a schematic configuration of a secret call function according to the present invention, FIG. 2 is a schematic configuration diagram of a workstation to which the present invention is applied, and FIG. The figure is an external view of the IC card attached to the workstation, and Figure 4 is an exploded perspective view showing the structure of the IC card.
ff Figure 15 is a diagram showing the structure of the printed circuit board part of the IC card, Figure 6 is a diagram showing the configuration of the semiconductor integrated circuit part of the IC card, Figure 7 is a diagram showing the configuration of the encryption processing part in the workstation, Figure 8 is a diagram showing the concept of encryption/decryption, Figure 9 is a block diagram of the encryption unit, W, Figure 10 is a diagram of the configuration of the decryption unit, Figure 11 is a diagram of the R3A processing unit, and Figure 12 is a diagram showing the configuration of the encryption unit. 13 is a diagram showing the configuration of the image matching unit in the workstation, FIG. 13 is a diagram showing an example of a face to be image processed, FIG. 14 is a diagram showing the structure of image data, and FIG. 15 is a diagram showing the structure of the image data in the workstation. 16 is a diagram showing an example of an input speech pattern, FIG. 17 is a diagram showing acoustic characteristics of a consonant, FIG. 18 is a diagram showing an example of a transition network, and FIG. 19 is a diagram showing an example of a transition network. The figure is gishi 1
Figure 20 is a diagram showing the continuation of the process, Figure 20 is a diagram to explain partial interval detection for input speech, Figure 21 is a diagram showing the learning processing procedure of a speech recognition dictionary, Figure 22 is character recognition on a workstation. FIG. 23 is a diagram showing an example of a FAX transmission manuscript paper on which characters to be recognized are written, and FIG. 24 is a diagram illustrating a process for cutting out characters to be recognized. Figure 25 is a diagram showing the configuration of the second character recognition block in the character recognition unit, Figure 26 is a diagram showing the configuration of the figure recognition unit in the workstation, and Figures 27 to 30 are diagrams showing the configuration of the second character recognition block in the character recognition unit. Figure 31 is a diagram showing the configuration of the image recognition unit in the workstation; Figure 32 is a diagram illustrating the configuration of the code conversion device;
Figure 33 is a diagram showing an example of processing for an input image;
Figure 4 is a diagram showing feature point detection in a segment, Figure 35 is a diagram showing the configuration of a voice verification unit in a workstation, Figure 36 is a diagram showing an example of band division of a filter bank, and Figure 37 is a diagram showing filter characteristics. Figure 38 is a diagram showing the configuration of the speech synthesis section in the workstation, Figure 39 is a diagram showing the configuration of the rule synthesis parameter generation device, Figure 40 is a diagram showing the leveling of conversion of the pitch parameter, and Figure 41 is a diagram showing the configuration of the speech synthesis section in the workstation. Figure 42 is a diagram showing the configuration of the image synthesis unit in the workstation, Figures 43 and 44 are diagrams showing the concept of image synthesis processing, and Figure 45 is the output form in the workstation. 46 is a diagram showing the flow of the output format selection processing procedure, FIG. 47 is a diagram showing the flow of the partner station identification processing procedure, and FIG. 48 is a diagram showing the structure of the media conversion table. , Fig. 49 is a diagram showing the configuration of the database section in the workstation, Fig. 50 is a diagram showing the data structure of the database, Fig. 51 is a diagram showing an example of beam ration, and Fig. 5 is a diagram showing the data structure of the database.
Figure 2 shows the ration (, 'η), Figure 53 shows the structure of the command correspondence table, and Figure 5 shows the structure of the command correspondence table.
Figure 4 is a diagram showing the configuration of the work environment data collection unit in the workstation, Figures 55 to 58 are diagrams for explaining the processing of the command unit, and Figure 59 is a diagram showing the flow of the system proficiency data collection process. 60 is a diagram showing the structure of the learning skin table, and FIGS. 61 to 68 are diagrams for explaining the processing of the work environment data collection section. l...Bus, 2...Control unit, 3...Image input device, 4...Position input device, 5...Audio input unit, 6
. . . Keyboard section, 7. IC card section, 9. . . Bus controller, 9. Audio output device, IO.
H...Communication device, 14...Switching device, 15...
Part of timer, 16... Encryption processing unit, 17... Voice verification unit, 18... Image verification unit, 19... Voice recognition unit, 20... Voice analysis unit, 21... Character recognition part,
22... Figure recognition unit, 23... Image recognition unit, 2
4... Output format selection unit, 25... Working environment data collection unit, 26... Voice synthesis unit, 27... Image synthesis unit, 28... Graphic synthesis unit, 29... Audio compression - Decompression section, 3... Image compression/expansion section, 31... Signal processing section, 32... Database section. Applicant's representative Patent attorney Takehiko Suzue Figure 1 Figure 4 Figure 5 Figure 7 Figure 13 Figure 14 Figure 19 Figure 20 Figure 21 Figure 26 Figure 27 Figure 28 Figure 29 Figure 30 Figure 31 Figure 32 (a) (b) Figure 33 Figure 34 Figure 38 Fig. 39 Fig. 40 Fig. 41 Fig. 42 Fig. 43 Fig. 44 Fig. 46 Fig. 48 Fig. 49 Fig. 50
ν,. Figure 52Figure 53Figure 54Figure 55Figure 56Figure 58Figure 60Figure 61Figure 63Figure 64Figure 65Figure 66

Claims (4)

[Claims] (1) A means for inputting voice, a means for transmitting the input voice data by telephone communication, a means for outputting the voice data received through telephone communication, and a means for transmitting the information indicated by the input character code string by voice. 1. A secret communication system comprising: means for synthesizing the voice; and means for communicating over the telephone using the voice synthesized voice data instead of the input voice. (2) The secret call system according to claim 1, wherein the character code string is obtained by recognition processing of a character pattern input by hand via a position coordinate input device. (3) The secret call system according to claim 1, wherein the character code string is obtained by recognition processing of character information read and inputted through an image input device. (4) The secret call system according to claim 1, wherein the character code string is input via a keyboard device.