JP2003223184A - Speech recognition device and method - Google Patents

Abstract

<P>PROBLEM TO BE SOLVED: To recognize a series of inputted speech data. <P>SOLUTION: A worker on the production line of a plant inputs facility management information (fault information) in voice from a portable telephone 10. The input speech is a series of speech of 'robot MMR5-1 encoder count error motor replacement', for example. A speech recognition server 14 first extracts a formal portion from the inputted speech, adds pauses before and after the form and distinguishes 'robot', 'MMR5-1' and 'encoder count error'. Besides, a sentence portion including a verb is extracted, pauses are added before and after such a portion and 'encoder count error' and 'motor replacement' are distinguished. After being divided into each of words, the input speech is collated with a database and each of the words are recognized. <P>COPYRIGHT: (C)2003,JPO

Description

Detailed Description of the Invention

[0001]

BACKGROUND OF THE INVENTION 1. Field of the Invention The present invention relates to a voice recognition device and method, and more particularly to a technique for recognizing a series of inputted voices.

[0002]

2. Description of the Related Art In recent years, with the progress of computer technology including IT, especially man-machine interface technology, technology for recognizing and operating a voice uttered by an operator has been developed. The operator does not need to use the keyboard to input data, and can efficiently input data or give necessary instructions to the computer. As an example,
It is assumed that a worker in a factory production line breaks down equipment on the line and repairs the failure. If such failure information (management information) in each line of the factory is stored in a database, even if similar failures occur in other parts of the line, just by accessing the database, the countermeasures can be quickly known. There are various advantages such as being able to perform, or statistically processing failure information. On the other hand, it is complicated and costly to provide each worker with a portable terminal, a notebook computer, or the like in order to make the management information into a database. Therefore, it is convenient if each worker can input the failure information by voice using a mobile phone such as PHS (registered trademark) that each worker already has for emergency contact, and can recognize this as voice to make a database. .

[0003]

However, in the conventional voice recognition, the operator has no choice but to input the voice while checking the management information one by one in accordance with the voice guidance. There was a problem that became complicated.

For example, assume a case where an abnormality occurs in an assembly robot on a production line. The type of robot is MMR
It is 5-1 and an encoder count error occurs, so that the operator replaces the motor. The items to be entered by the operator are a total of four items, that is, the robot that is the target equipment, the model that identifies the robot that is the target process, the encoder count error that is the content of the failure (failure), and the motor replacement that is the content of the countermeasure. According to the voice guidance, the worker first inputs the voice "robot" and causes the computer to recognize the voice. The computer repeats the recognition result, and the operator replies "Yes" or "No" with the validity of the recognition result. In the case of "no", the worker again inputs "robot" by voice to make the computer recognize it.

In the case of "yes", the operator next inputs the voice "MMARGONOICH (MMR5-1)". The computer recognizes this voice again and repeats the result. As described above, input-recognition-repeat-confirmation is repeated for each item to cause the computer to recognize, and it takes time and effort to input all management information.

[0006] This is because, for example, when a worker makes a voice input at once, such as "Robot MMR Gonoichi Encoder Count Miss Motor Kokan", there is no guarantee that the computer can accurately recognize each word division. There is a problem in that a section that should be recognized as an "encoder count error" cannot be recognized as a break from the preceding "Ichi", and it is mistakenly recognized as a "Ichi encoder count error", that is, a position encoder count error. Because you get it.

The present invention has been made in view of the above problems of the prior art, and an object thereof is to provide an apparatus and method capable of improving the recognition rate even when a series of voices is input. It is in.

[0008]

In order to achieve the above object, the present invention extracts a predetermined characteristic part from a voice input means and a series of voices inputted from the voice input means, It is characterized by further comprising processing means for dividing the series of voices before and after the portion and recognition means for recognizing the voices divided by the processing means by collating them with a voice database.

Here, it is preferable that the characteristic portion is a portion composed of alphabets and numbers.

Further, it is preferable that a series of voices including a machine type is input from the voice input means, and the characteristic portion is the machine type.

Further, a series of voices including a target equipment, a process, a trouble, and a countermeasure content is input from the voice input means, and the processing means extracts the step as the characteristic part from the series of voices, The series of voices is separated before and after, and the measure content is further extracted from the series of voices as the characteristic part, and the series of voices is divided before and after the measure content, so that the series of voices is the target equipment,
Dividing into process, trouble, and countermeasure content, the recognition means divides the target equipment, process, trouble, and the like by the processing means.
It is preferable to collate the countermeasure contents with the voice database. In this device, the voice input means is preferably a mobile phone.

The present invention also provides a method for recognizing a series of inputted voices. This method comprises the steps of extracting a predetermined characteristic part from the series of voices, dividing the series of voices into a plurality of words based on the characteristic part, and dividing each of the divided words into a voice database. And a step of recognizing by collating with.

Here, the characteristic part may be a part consisting of alphabets and numbers, and is preferably a machine type.

Further, the series of voices are voices including target equipment, process, trouble, and countermeasure contents. In the extracting step, the process and countermeasure contents are extracted, and in the dividing step, the process and By dividing the series of sounds before and after the countermeasure content, the target equipment, process,
It is preferable to divide the problem and countermeasure content.

Also in this method, it is preferable that the series of voices be input from a mobile phone.

[0016]

BEST MODE FOR CARRYING OUT THE INVENTION Embodiments of the present invention will be described below with reference to the drawings by taking as an example a case where an operator in a production line of a factory inputs management information (fault information) by voice.

FIG. 1 shows a block diagram of a voice recognition system including a voice recognition apparatus according to this embodiment. Each worker on the production line uses a mobile phone 1 such as PHS (registered trademark)
0 is held, and when a failure occurs in the production line, voice input is performed using the mobile phone 10. Mobile phone 1
The voice input from 0 is supplied to the voice recognition server 14 via the public network 12.

The voice recognition server 14 recognizes the voice input from the mobile phone 10 and stores the recognition result in the database 16
Register with. Speech recognition server 14 and database 16
Are connected by, for example, a LAN.

FIG. 2 shows the voice recognition server 1 shown in FIG.
4 is a configuration block diagram. Speech recognition server 1
Reference numeral 4 denotes an interface (I / F) for inputting a voice (a series of voices in this embodiment) from the public network 12, a CPU, and an R.
It has a memory such as OM and RAM, and further has a voice database and a voiceprint database.

The voice database comprises a voice database for each management item, and specifically includes a target facility voice database, a process voice database, a fault voice database, and a countermeasure content voice database. Each voice database stores model data consisting of multiple consecutive phonemes, and the input voice is collated with the model data in the database, and the input voice is input by specifying one of the words in the database. Recognize voice.
A technique for collating input voice data with model data, or a waveform pattern matching technique is known.

The voiceprint database is a database for storing voiceprint data unique to each worker. The voiceprint database holds not only the voiceprint of each worker but also the "habit" at the time of voice input of each worker as data. The “habit” includes a tendency of utterance as well as a sequence of items for inputting management information. For example, a certain worker A inputs the target equipment, the process, the trouble, and the countermeasure contents in this order, while another worker B inputs the trouble, the process, the target equipment, and the countermeasure contents in that order. The input order data of the management information can be acquired by learning. As is well known, the spectrum obtained by frequency analysis of voice shows the characteristic peculiar to the person like a fingerprint, and the owner of the voice can be specified by comparing the voice data with the voiceprint data. You can also In the present embodiment, a voiceprint database is used to identify a worker from the input voice, the input voice is corrected based on the worker's habit, and collated with standard model data. When recognizing each item of the management information, the voice input order of the operator identified based on the voice print is taken into consideration.

The CPU divides a series of voices input via the interface I / F into each item, that is, into each word having a meaning as management information, and the voice data of each divided word is used as a voice database. Recognize by collating.
When dividing a series of voices into meaningful words, the CPU extracts a characteristic portion from a series of input voices, and divides the extracted characteristic portion before and after the characteristic portion. The characteristic part is a part that can uniquely specify which of the four items, specifically, a process part. The process is dictated by the machine type, which consists of a combination of letters and numbers. For example, "M
MR5-1 "and the like. There is no item consisting of such a combination of letters and numbers other than the process. Therefore, by extracting the alphabet and number parts from a series of voices, the extracted part can be specified as a process, and by separating the voice data before and after the characteristic part, it is possible to separate other items from the process. Can be correctly recognized.

FIG. 3 shows a processing flowchart of the voice recognition server 14 in this embodiment based on the above principle. It is assumed that the worker uses the mobile phone 10 to suddenly input "Robot MMR Gonoichi (meaning MMR5-1) Encoder Count Miss Motor Change (meaning motor replacement)" at once.

First, the voice recognition server 14 extracts a portion including alphabets and numbers from the voice data collectively input, and specifies this as a type word (S10).
1).

Specifically, "MMARGONOICH"
Part of. Needless to say, in order to recognize that it is a part of the alphabet and the numbers, it is necessary to collate it with the voice data of the alphabet and the voice data of the numbers in the voice database. Alphabets and numbers are relatively easy to recognize.

After extracting the model part, the voice recognition server 1
In step 4, a series of voice data is divided before and after the model part is sandwiched (S102). Specifically, the "robot" is divided into "robot MM Argo Noichi" before "M MU Argo Noichi".
And "MMARGONOICH", and in the section of "MMARGONOICH Encoder Count Mistake" after "MMARGONOICH", it is separated into "MMARGONOICH" and "Encoder Count Miss". Divide into.

Next, the voice recognition server 14 extracts a sentence containing a verb from a series of voice data and recognizes it as a sentence (contents of countermeasure) (S103). Specifically, "motor kokan" including the verb "kokan" is extracted from the input voice. The verb can be extracted by comparing the input voice data with the verb model data stored in advance in the countermeasure content voice database in the voice database. The verb model data is composed of verb groups that can be considered as countermeasures, such as “exchange”, “replacement”, and “replenishment”. A document containing a verb has only countermeasure contents as an item, and by extracting a document containing a verb, the countermeasure contents item can be uniquely extracted from a series of voices.

After recognizing the sentence (contents of the countermeasure), a series of voices are divided before and after the contents of the countermeasure (S104). Specifically, the "encoder count miss motor change" part is divided into "encoder count miss" and "motor change". The "Motorkoukan" exists at the end of the series of voices, so it is not separated after that.

As described above, after a series of voices is divided before and after the model and the contents of countermeasures, recognition is performed at each division (S105). That is, S101 to S1
By the processing of 04, a series of voices are "robot", "MMARGONOICH", "encoder count error",
It is divided into four parts (words) "Motorkoukan", and the voiceprint of the input voice can tell the word order by the voice of any worker, so it is possible to know which item each word corresponds to (type and countermeasure content). Has already been known, and it is clear that the remaining two items are the target equipment and malfunction, respectively.) By comparing the model data of each item in the voice database with the input voice, the voice data of each item is obtained. Can be recognized. For example, if it is possible to specify that the voice input by the worker A is input from the voiceprint of the input voice, and the worker A inputs in the order of the target equipment, the process, the defect, and the content of the countermeasure, the “robot” and the “encoder count error” are Each can be identified as a target equipment and a failure, and the robot is verified by matching the "robot" with the model data in the target equipment voice database, and the "MMARGONOICH" is compared with the model data in the process voice database. Recognizes that it is MMR5-1, matches the "encoder count error" with the model data in the defective voice database, recognizes that it is an encoder count error, and identifies "Motor Kokan" as the countermeasure data voice database model data. It is confirmed that it is a motor replacement by collating with. The steps and countermeasures are S101 and S, respectively.
It is also possible to recognize at 103.

After recognizing all four items, the voice recognition server 14 stores the recognition result in the database (management information database) 16 (S106). Database 16
In this case, the management information is classified into four items and accumulated, and the worker can access the database 16 as necessary to obtain the necessary information. The information provided from the database 16 to the worker can be appropriately processed. For example, it is possible to output in the format of a target equipment list in which the motor has been replaced, or in the format of a list of defects that occurred in the past in a certain target equipment. It would also be preferable to output statistical results such as a failure occurrence rate in a specific process in a specific target facility.

As described above, in the present embodiment, attention is paid to the model item consisting of the alphabet and the numbers as the characteristic part, the attention is paid to the measure content item including the verb, and a series of voices is put before and after these items. Is correctly divided into words, and it is not necessary to input each word as in the past and input the next word while checking the recognition result each time, and efficiently input the management information by voice. You can

Further, in the present embodiment, since a series of voice inputs is performed using the mobile phone 10 which each worker already holds for emergency contact, a new mobile terminal or personal computer is provided to the worker. There is no need to give it, and it is possible to reduce the cost required for infrastructure development. Familiar mobile phone 1
There is also an advantage that you can input voice from 0.

Although the embodiment of the present invention has been described above, the present invention is not limited to this, and various modifications can be made. For example, in the present embodiment, a model consisting of alphabets and numbers is extracted as the characteristic portion, but the characteristic portion can be determined according to the nature or category of a series of input voices. The attribute that the characteristic part should have is that it is composed of phoneme data that is clearly distinguished from other items and can be uniquely determined.

[0034]

As described above, according to the present invention, the recognition rate in the case of inputting a series of voices can be improved and the voice input efficiency can be increased.

[Brief description of drawings]

FIG. 1 is a configuration diagram of a voice recognition system according to an embodiment.

FIG. 2 is a block diagram of a configuration of a voice recognition server in FIG.

FIG. 3 is a processing flowchart of an embodiment.

[Explanation of symbols]

10 mobile phone, 12 public network, 14 voice recognition server, 16 database.

Claims (10)

[Claims] 1. A voice input means, a processing means for extracting a predetermined characteristic portion from a series of voices input from the voice input means, and dividing the series of voices before and after the characteristic portion, and the processing. A voice recognition device, comprising: a recognition unit that recognizes voices separated by the means by collating each voice with a voice database. 2. The voice recognition device according to claim 1, wherein the characteristic portion is a portion including alphabets and numbers. 3. The voice recognition device according to claim 1, wherein a series of voices including a model of a machine is input from the voice input unit, and the characteristic portion is a model of the machine. 4. The apparatus according to claim 1, wherein a series of voices including target equipment, processes, troubles, and countermeasures are input from the voice inputting means, and the processing means outputs the characteristic portion from the series of voices. By extracting the step, dividing the series of voices before and after the step, further extracting the measure content as the characteristic portion from the series of voices, and dividing the series of voices before and after the measure content A series of voices is divided into the target equipment, steps, defects, and countermeasures, and the recognition means collates the target equipments, steps, troubles, and countermeasures separated by the processing means with the voice database. And a voice recognition device. 5. The apparatus according to claim 1, wherein the voice input unit is a mobile phone. 6. A method for recognizing a series of input voices, the method comprising: extracting a predetermined characteristic part from the series of voices; and converting the series of voices into a plurality of words based on the characteristic part. A voice recognition method comprising: a step of dividing; and a step of recognizing each of the divided words by collating each word with a voice database. 7. The voice recognition method according to claim 6, wherein the characteristic portion is a portion including alphabets and numbers. 8. The voice recognition device according to claim 6, wherein the characteristic portion is a machine type. 9. The method according to claim 6, wherein the series of voices is voice including target equipment, process, trouble, and countermeasure content, and in the extracting step, the process and countermeasure content are extracted, In the step of dividing, the target equipment by dividing the series of voices before and after the process and the content of the countermeasure,
A voice recognition method characterized by dividing processes, defects and countermeasures. 10. The voice recognition method according to claim 6, wherein the series of voices is input from a mobile phone.