JPH09120296A - Device and method for speech recognition, device and method for dictionary generation, and information storage medium - Google Patents

Abstract

PROBLEM TO BE SOLVED: To speed up speed recognition. SOLUTION: A word notation correspondence dictionary 5 is stored previously with speeches of recognition candidates while they are classified by the readings of the head parts. The readings of the head parts of '100' and '150' are 'hyaku'. When a voice is inputted to a voice input means 3, the speech whose head part has the same reading with its head part is detected by a speech recognition means 4 in the word notation correspondence dictionary 5. The reading to be processed for speech recognition is limited to the head part, so the number of head parts is decreased and the processing is completed fast.

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 for recognizing voice, a dictionary creation device and method for storing language and reading in a word notation corresponding dictionary of the voice recognition device, and a computer program in advance. It relates to a written information storage medium.

[0002]

2. Description of the Related Art Currently, there is a demand for practical use of a voice recognition device for recognizing voice. A general voice recognition device has a dictionary in which a reading is stored for each language of a recognition candidate, and when a voice is input, the reading of the dictionary is searched, and the voice is recognized as a matching language.

[0003]

The speech recognition device as described above can recognize speech as a language by searching the dictionary for a reading that matches the speech.

However, since the number of languages to recognize is enormous, processing is delayed if the readings are searched individually,
It becomes difficult to recognize continuously input voices in real time.

One of the techniques for solving such a problem is
"Two-stage LR parser considering meaning in free speech recognition" (Minami et al., Proceedings of ASJ, 3-4-10, 199)
It is disclosed in 3.3). This is intended for phone number inquiry tasks, and such inquiries are
Focusing on the fact that the items have meaning and the style has no meaning, the language of recognition candidates is reduced by considering the meaning.
More specifically, the LR table is created by classifying it at the meaning level, and a plurality of languages having the same meaning are put together.

However, if the tasks are extremely limited, the recognition candidate languages can be reduced as described above. However, in the case of general tasks, there are many languages having the same meaning. In such a case, the recognition candidate languages cannot be effectively reduced.

Another method for reducing the recognition candidate language is a "menu-based speech natural language input system".
(Yamamoto et al., 47th National Convention of IPSJ, 7M-2, 1993.1
It is disclosed in 9). In this method, the input unit of speech is a phrase, and the input is limited to the range where the system can execute the process with high probability to reduce the recognition target language.

However, since human utterances tend to be ambiguous at the end of sentences as compared to the beginning of sentences, erroneous recognition tends to occur when the input unit is a syllable as described above. In addition, in the case of utterances such as consecutive numbers, the first is often the same and the last is different, such as "sanju, sanjuni, sanbyaku", but when such a voice is recognized by the method described above,
Processing takes time with a low recognition rate.

In such a case, it is desirable that a plurality of recognition candidates including the correct answer be output in a short time process, rather than one incorrect recognition result being output in a long time process. In other words, even if there are multiple recognition candidates that are output as a result, if the correct answer is included, it is possible to narrow it down to one correct answer by another method. In such a case, the first process is quick. Required to be present. However, such a thing is not considered in the above-mentioned method.

[0010]

According to a first aspect of the present invention, there is provided a voice recognition device, wherein a voice input means for inputting a voice to be recognized and a language of a recognition candidate are stored in advance for each reading of the head part. It has a notation corresponding dictionary and a voice recognition means for detecting from the word notation corresponding dictionary a language in which the reading of the beginning matches the beginning of the voice input by the voice input means. When the voice to be recognized is input to the voice input unit, the voice recognition unit detects, from the word notation corresponding dictionary, a language in which the reading of the beginning matches the beginning of the voice. At this time, the reading to be processed by the voice recognition is limited to the head portion and the number of readings is reduced, so that this processing operation is executed at high speed.
Since human utterances tend to be clear at the beginning of speech, the rate of false recognition also decreases. It occurs when there are multiple recognition results, but since this includes a correct answer with a high probability, it is more practical than one in which one wrong recognition result is output by long-time processing. Since the head part here is a fixed part from the head, for example, in a short language, the entire reading may be the head part of the reading.

According to another aspect of the dictionary creating apparatus of the present invention, a general language dictionary in which various languages are read and stored in advance,
And a reading generation unit that stores the language extracted from the general language dictionary in the word notation corresponding dictionary for each head part of the reading. Since the reading generation means stores the language extracted from the general language dictionary in the word notation corresponding dictionary for each head portion of the reading, a plurality of languages having the same head but different tails are integrated into one reading. When the voice recognition device recognizes a voice by using such a word notation compatible dictionary, the number of readings to be processed by the voice recognition device is reduced. It should be noted that the general language dictionary mentioned here may be a general dictionary in which various languages are stored in advance together with reading, and for example, an existing language database or the like can be used.

According to the dictionary creating apparatus of the invention of claim 3,
A phonetic unit dictionary in which phonetic units for reading the language are stored in advance is provided, and the reading generation means generates a fixed phonetic unit portion from the beginning of the language as the beginning part of the reading, so reading is performed by simple processing. The length of the beginning part of the voice recognition becomes constant, and the characteristics of human utterance are well reflected in the reading to be processed by the voice recognition device.

According to another aspect of the dictionary creating apparatus of the present invention,
Since the reading generation means generates a portion of a fixed notation unit from the beginning of the language as the beginning of the reading, the length of the beginning of the reading becomes constant by a simple process, and the reading to be processed by the voice recognition device becomes The characteristics of human vocalization are well reflected.

According to another aspect of the dictionary creating apparatus of the present invention,
The length of the leading part of the reading is preset in the length setting dictionary for each language classification, and the reading generation means changes the length of the leading part of the reading of the generated language for each classification. For example, if the reading length of only a specific language is extended, the recognition accuracy of a specific voice is improved without increasing the time required for the entire processing of the voice recognition device.

According to the dictionary creating apparatus of the invention described in claim 6,
The reading of each digit of the number is stored in advance in the digit-corresponding dictionary for each number of digits, and the reading generation means, when generating the reading of the leading part by using the plurality of digits of the language as the language, The reading is detected from the general language dictionary, and the reading of a predetermined digit at the beginning is detected from the digit correspondence dictionary and combined. If the general language dictionary consists of a general database, etc., it is highly likely that a single-digit number will be stored, but a multi-digit number will not be stored. It is generated by a simple process.

According to another aspect of the dictionary creating apparatus of the present invention,
The reading of each digit that changes depending on the combined number is stored in advance in the reading change dictionary, and the reading generation means refers to the reading change dictionary to generate a reading when the reading is generated using a plurality of digits as a language. Correct the reading of. Readings that are unnatural in a simple combination are corrected to a natural form.

According to another aspect of the dictionary creating apparatus of the present invention,
The language classification is preset in the language classification dictionary, and the reading generation unit classifies the language corresponding to the generated reading according to the setting of the language classification dictionary, and the number of the classified languages is a preset reference value. If it does not exceed, the reading is not restricted to the beginning. For example, if the reading of a language in which voice recognition frequently appears is limited to the beginning and the reading of a language in which appearance frequency is low is not limited to the beginning, the processing time of the speech recognition apparatus is shortened and the recognition accuracy is reduced. Is improved.

According to the dictionary creating apparatus of the invention of claim 9,
The language classification is preset in the language classification dictionary, and the reading generation means classifies the language corresponding to the generated reading according to the setting of the language classification dictionary, and the number of readings in this classification is set to a preset reference value. The length of the leading part of the reading of the generated language is changed so that it does not exceed the limit. Even if there are many languages in one category, the number of readings is constant, and if the number of languages in one category is small, the reading is not limited to the beginning.

In the dictionary creating apparatus according to the tenth aspect of the present invention, consecutive conditions of a plurality of languages are preset in the condition setting dictionary, and the reading generation means refers to the condition setting dictionary and a plurality of consecutive languages. Generate the reading of and limit the reading to the beginning only in the language located at the end. Multiple languages that are expected to be consecutive are combined in advance and treated in the same way as one language.

In the dictionary creating apparatus according to the eleventh aspect of the present invention, the reading generation means reads the reading of the language located at the end so that the total reading of a plurality of consecutive languages does not exceed a preset reference value. Since the length of the head part is limited, even if the head language of a plurality of languages to be combined is long, the total length is constant.

In the dictionary creating apparatus according to the twelfth aspect of the present invention, the condition setting dictionary is configured such that the reading length of the language located at the end is preset together with the continuous conditions of a plurality of languages. , Limit the reading of the last language to a set length. Since the reading length of the last language is variable according to the classification, the longer the first language, the shorter the last language can be.

The speech recognition method according to the invention of claim 13 is
The language of the recognition candidate is stored in advance in the word expression corresponding dictionary for each reading of the head part, and the language in which the head reading matches the head part of the speech to be recognized is detected from the word expression corresponding dictionary. . When the speech to be recognized is input, a language whose reading at the beginning matches the speech is detected from the word notation corresponding dictionary. At this time, the reading to be processed by the voice recognition is limited to the head portion and the number of readings is reduced, so that this processing operation is executed at high speed. Since human utterances tend to be clear at the beginning of speech, the rate of false recognition also decreases. It occurs when there are multiple recognition results, but since this includes a correct answer with a high probability, it is more practical than one in which one wrong recognition result is output by long-time processing.

The dictionary creating method according to the invention of claim 14 is
Various languages are taken out from a general language dictionary that is stored in advance with reading, and the taken out languages are stored in the word notation corresponding dictionary for each leading part of reading. Since the languages extracted from the general language dictionary are stored in the word notation corresponding dictionary for each head part of the reading, a plurality of languages having the same head but different tails are aggregated into one reading. When the voice recognition device recognizes a voice by using such a word notation compatible dictionary, the number of readings to be processed by the voice recognition device is reduced.

An information storage medium according to the invention of claim 15 is
In an information storage medium in which computer-readable software is written in advance, a word notation-compatible dictionary in which a recognition candidate language is stored in advance for each reading of the beginning portion and a reading of the beginning portion in the beginning portion of the speech to be recognized are stored. A program for causing the computer to detect a matching language from the word notation corresponding dictionary is written. When the software of this information storage medium is read by a computer and operated, the computer functions as a voice recognition device that detects, from the word notation corresponding dictionary, a language in which the reading of the beginning matches the beginning of the voice. At this time, the reading to be processed by the voice recognition is limited to the head portion and the number of readings is reduced, so that this processing operation is executed at high speed. Since human utterances tend to be clear at the beginning of speech, the rate of false recognition also decreases. It occurs when there are multiple recognition results, but since this includes a correct answer with a high probability, it is more practical than one in which one wrong recognition result is output by long-time processing.

The information storage medium of the invention according to claim 16 is
In an information storage medium in which a program that a computer reads and executes a corresponding operation is written in advance, take out a language from a general language dictionary in which various languages are read and stored in advance. A program for causing the computer to execute is stored in the word notation compatible dictionary. When the computer reads the program of the information storage medium and operates the computer, the computer functions as a dictionary creating device that creates a dictionary corresponding to word notation used in the voice recognition device. At this time, since the language extracted from the general language dictionary is stored in the word notation corresponding dictionary for each head part of the reading, a plurality of languages having the same head but different tails are integrated into one reading. When the voice recognition device recognizes a voice by using such a word notation compatible dictionary, the number of readings to be processed by the voice recognition device is reduced.

[0026]

FIG. 1 shows a first embodiment of the present invention.
This will be described below with reference to FIG. First, as shown in FIG. 1, the voice recognition device 1 and the dictionary creation device 2 illustrated here are integrally formed. As shown in FIGS. 2 and 3, the hardware is a data processing device. It has a computer system 100. This computer system 100 has a CPU (Central Processing Unit) 101 as a main body of a computer, and this CPU 101 is provided with a bus line 102.
ROM (Read Only Memory) 103, RAM (Random Acce
ss Memory) 104, HDD (HD Drive) 105 with built-in HD (Hard Disk ... Not shown), FD (Floppy Disk) 10
FDD (FD Drive) 107, CD (Compact
Disk) -CD-ROM drive 109 in which ROM 108 is loaded, keyboard 11 to which mouse 110 is connected
1, a display 112, a microphone 113, a communication I / F (Interface) 114, etc. are connected.

In the computer system 100, a program or the like for causing the CPU 101 to execute various processing operations is set in advance, and the software such as the program is, for example, the RAM 104 which is an information storage medium.
And is previously written in the HD (not shown) of the HDD 105. In the computer system 100 as described above, the CPU 101 executes various processing operations in accordance with the programs stored in the RAM 104 or the like, whereby the voice recognition device 1 and the dictionary creating device 2 of the present embodiment are realized. There is.

As shown in FIG. 1, the voice recognition device 1 of the present embodiment supports a voice input unit 3 which is a voice input unit, a voice recognition unit 4 which is a voice recognition unit, and a word notation which is a word notation compatible dictionary. It has a table 5, a result output section 6 which is a result output means, and a result selection section 7 which is a result selection means, and the voice recognition section 4 has the voice input section 3, the word notation correspondence table 5, and the result. The output unit 6 and the result selection unit 7 are connected.
The dictionary creating device 2 of the present embodiment has a word dictionary 8 that is a general language dictionary and a reading generating unit 9 that is a reading generating unit. The reading generating unit 9 includes the word dictionary 8 and the reading dictionary. The word notation correspondence table 5 is connected.

The voice input unit 3 has the microphone 113 as hardware and converts a voice uttered by a human into an electric signal. Correspondence table 5 for word notation
The word dictionary 8 and the word dictionary 8 have an information storage medium such as the RAM 104. In this case, numerical words are stored in advance as a recognition candidate language. The word dictionary 8 is composed of, for example, an existing database of numerical words corresponding to a task of voice recognition, and as shown in FIG. 4, various numerical words that are recognition candidates are stored together with readings in advance.

In the word notation correspondence table 5, as shown in FIG. 5, the recognition candidate numbers are stored in advance for each reading of the head part. The reading of the leading part of the numeral is generated as a part of a fixed notation unit from the beginning of the reading of the numeral, and specifically, the number of characters as the notation unit of the numeral is three or less. Is restricted to. For this reason, "hyakujugo," which is the reading of the number "115", is shortened to "hyaku" at the beginning, but "100""hyaku" remains "hyaku."

The voice recognition unit 4 has the CPU 101 and the like, and when the voice input unit 3 inputs a voice to be recognized, the head of the voice is read with a numeral that matches the reading of the head. It is detected from the word notation correspondence table 5. In this case, since the leading part of the reading stored in the word notation correspondence table 5 is three characters as described above, only the leading three characters of the input voice are processed.

The result output unit 6 includes the display 1
12 and the like, and outputs the recognition result of the voice recognition unit 4. The result selection unit 7 uses the keyboard 111.
When there are a plurality of output results of the result output unit 6, one of them is selected in response to the user's manual operation.

The reading generation unit 9 has the CPU 101 and the like, and stores the numerical words extracted from the word dictionary 8 in the word notation correspondence table 5 of the voice recognition device 1 for each head of the reading. As described above, in the word notation correspondence table 5, the readings of the numerical words are stored for every three characters at the beginning, so that the reading generation unit 9
Generates a three-letter part, which is a fixed notation unit, from the beginning of the reading of the number taken out from the word dictionary 8 as the beginning.

The above-mentioned voice recognition device 1 and dictionary creation device 2
The respective units are realized by utilizing the hardware such as the keyboard 111, the display 112, the microphone 113, etc., if necessary.
This is realized by the CPU 101 operating in accordance with software written in the M104 or the like.

The software written in the RAM 104 in this way is the word dictionary 8 composed of software readable by the CPU 101, the language is taken out from the word dictionary 8 and the word notation correspondence table is provided for each leading part of reading. 5, the control program of the CPU 101 stored therein, the word notation correspondence table 5 composed of software readable by the CPU 101, and when the voice to be recognized is input to the voice input unit 3, the head portion is read at the beginning. CPU 10 for detecting a language in which is matched from the word notation correspondence table 5
1 control program, etc.

In such a configuration, the voice recognition device 1
Recognizes human voices. More specifically, as shown in FIG. 6, when a voice uttered by a human is input to the voice input unit 3, the voice recognition unit 4 extracts the three characters at the beginning of the voice and sets the start point to the beginning. Numerals whose readings match are detected from the word notation correspondence table 5 by the fixed spotting. Since the number output detected in this way is output from the result output unit 6, if there are a plurality of detected number output, the result selection unit 7 manually selects one number.

For example, when "Hyakuju" is input as a voice, the reading is "100," which is the number of "Hyaku".
Since three of 110 and 115 "are output, the user selects" 110 "as desired.

In the voice recognition device 1 described above, since the word notation correspondence table 5 stores the numerics for each reading of three characters, the number of readings to be processed is reduced and the processing load is reduced. The time required is shortened. Moreover, since the reading retrieval processing is executed with only three characters in this way, the processing load is reduced and the required time is shortened.

Human utterances tend to be ambiguous at the end of a sentence as compared to the beginning of a sentence, but since the voice recognition apparatus 1 described above targets only the beginning of a voice, erroneous recognition is less likely to occur. In this case, there are many cases where the recognition result becomes plural as described above, but the plural recognition candidates include the correct answer with a high probability and are output in a short time. Even if the second process selected in step 1 is executed, the overall required time is short, and it is more practical than the case where one erroneous recognition result is output in a long time process.

The word notation correspondence table 5 of the voice recognition device 1 stores the number for each special reading as described above, which is mechanically created by the dictionary creating device 2. That is,
As shown in FIG. 4, in the word dictionary 8, various numbers as recognition candidates are pre-stored together with the reading, so that the reading generation unit 9 extracts the numbers extracted from the word dictionary 8 for each reading head part. And stored in the word notation correspondence table 5 of the voice recognition device 1.

More specifically, when the number "100" is first retrieved from the word dictionary 8, the reading is "hyaku", so this "hyaku" is "1" as the beginning of the reading.
00 ”is stored in the word notation correspondence table 5. Next, when the number“ 110 ”is taken out, the reading is“ hyakujyu ”, so the head part is“ hyaku ”, and this number“ 1 ”
10 "is stored at the reading position of" hyaku "in the word notation correspondence table 5 together with" 100 "described above.

For this reason, in the word notation correspondence table 5, a large number of numerical words are allocated and stored in a small number of readings, and a plurality of numerical words having the same head but different tails are likely to cause misrecognition. Are summarized in. Since the word notation correspondence table 5 of the voice recognition device 1 is mechanically created by the dictionary creation device 2 from the existing word dictionary 8, it is not necessary for a person to perform this work. In this way, since the reading of the number of words is limited to the head portion according to the number of characters that are the notation units, the length of the head portion of the reading can be made constant in common by a simple process.

The present invention is not limited to the above-described embodiment, but allows various modifications. For example,
Here, it is illustrated that the voice recognition device 1 and the dictionary creation device 2 are integrated, but this is a separate device, and the dictionary creation device 2 is used when the voice recognition device 1 is developed and shipped as a product. The voice recognition device 1 may not include the dictionary creation device 2.

In the present embodiment, the CPU 101 operates according to the program stored in the RAM 104 or the like as software, whereby the voice recognition device 1
It is exemplified that each unit of the dictionary creating device 2 is realized.
However, it is also possible to manufacture each of these units as its own hardware, and to store some of them as software in the RAM 104 or the like and to manufacture some of them as hardware. Further, it is possible to manufacture the RAM 104 or the like in which predetermined software is stored and the hardware of each unit as firmware, for example.

Further, in this embodiment, when the computer system 100 is started up, the software stored in the HDD 105 is copied to the RAM 104, and the software stored in the RAM 104 is stored in the CPU 101.
However, it is also possible to allow the CPU 101 to use such software stored in the HDD 105 or to write it in the ROM 103 or the RAM 104 in advance.

Further, it is also possible to write the software in the FD 106 or the CD-ROM 109, which is an information storage medium that can be handled by itself, and install the software in the RAM 104 or the like from the FD 106 or the like.
FD106 without performing such installation
It is also possible for the CPU 101 to appropriately read the software written in, for example, to execute data processing.

Further, it is also possible to realize a program for realizing each part of the voice recognition device 1 and the dictionary creation device 2 by combining a plurality of software,
In that case, only the minimum necessary software needs to be stored in the information storage medium which is a single product. For example,
When the application software is provided to the computer system 100 in which the operating system is installed by an information storage medium such as the CD-ROM 108, the software that realizes each unit of the voice recognition device 1 and the dictionary creation device 2 is the application software. Since it is realized in combination with the operating system, the software of the part that depends on the operating system can be omitted from the information storage medium of the application software.

The method of supplying the software thus written in the information storage medium to the computer is not limited to loading the information storage medium directly into the computer. For example, it is also possible to write the above-mentioned software on an information storage medium of a host computer, connect the host computer to a terminal computer via a communication network, and supply the software to the terminal computer by data communication from the host computer.

In this case, it is possible for the terminal computer to execute stand-alone data processing while the software is downloaded to its own information storage medium. However, the real-time data communication with the host computer is possible without downloading the software. It is also possible to perform data processing. In this case, the entire system in which the host computer and the terminal computer are connected by the communication network is the voice recognition device 1 of the present invention.
And the dictionary creation device 2.

Next, a second embodiment of the present invention will be described with reference to FIG.
This will be described below with reference to FIG. Regarding the second embodiment, the same parts as those in the first embodiment described above are designated by the same names and reference numerals, and detailed description thereof is omitted.

First, as shown in FIG. 7, a voice recognition device 11 and a dictionary creating device 12 illustrated here are also integrally formed, and the dictionary creating device 12 has a syllable table 13 which is a phonetic unit dictionary. Has been added. This syllable table 13 is R
It has a storage device such as an AM, and as shown in FIG. 8, syllables, which are the phonetic units for reading the numbers, are stored in advance. The reading generation unit 9 refers to the syllable table 13 when storing the numbers extracted from the word dictionary 8 in the word notation correspondence table 5 for each beginning of the reading, and identifies the two syllables from the beginning of the reading of the numbers. It is generated as the head part.

With such a configuration, the voice recognition device 11 of the present embodiment also recognizes a voice uttered by a human.
At this time, the voice recognition unit 4 extracts the two syllables at the beginning of the voice, and detects from the word notation correspondence table 5 a number that matches the reading of the two syllables. Since the voice recognition device 11 executes the process of recognizing a voice in accordance with the syllable which is a phonetic unit, it is possible to favorably reflect the characteristics of human utterance in the process of voice recognition.

Then, the dictionary creating device 12 also creates the word notation correspondence table 5 of the voice recognition device 11 as described above. When the reading generation unit 9 stores the number extracted from the word dictionary 8 in the word notation correspondence table 5 for each head of reading, the syllable table 13
Refer to to limit reading to two syllables. In this way, since the reading of the number is limited to the head portion according to the number of syllables that are phonetic units, the length of the head portion of the reading can be made constant in common by a simple process.

Next, a third embodiment of the present invention will be described with reference to FIG.
This will be described below with reference to FIGS. Note that, regarding the third embodiment of this embodiment, the same parts as those of the second embodiment described above are denoted by the same names and reference numerals, and detailed description thereof is omitted.

First, as shown in FIG. 10, a voice recognition device 21 and a dictionary creation device 22 illustrated here are integrally formed, and the dictionary creation device 22 has a recognition word table 23 which is a language classification dictionary. Has been added.

The recognition word table 23 has a storage device such as a RAM, and as shown in FIG. 11, the classification of numbers as languages is set as "product A, product B". The length of the leading part of the reading is set as "2, 3". Here, it is assumed that the product A is a copy machine and the product B is a facsimile, and the corresponding numeral is the model number of the product.

Then, the reading generation unit 9 recognizes the length of the leading portion of the reading to be generated when limiting the reading to the leading portion when setting the numerics taken out from the word dictionary 8 in the word notation correspondence table 5. It is changed for each classification according to the setting contents of the word table 23. Therefore, as shown in FIG. 12, the word dictionary 8
The classification of "product A, product B" is also set for each of the numbers stored in the table, and as shown in FIG. 13, the reading of the numbers of the product A consists of two syllables. Consists of three syllables.

With such a configuration, the voice recognition device 21 of the present embodiment also recognizes a voice uttered by a human.
At this time, the voice recognition unit 4 extracts one syllable at the beginning of the voice, and identifies the number whose reading is consistent with this as a word notation correspondence table 5.
Detect from. For this reason, the number "hyaku" classified as the product A is recognized at the second syllable, while the number "nyak" classified as the product B is recognized at the third syllable.

In the voice recognition device 21, since the reading length of the recognized voice is changed according to the classification of the language, for example,
You can improve the recognition accuracy of a specific voice without increasing the reading time for a specific language and increasing the overall processing time, and adjust the voice recognition accuracy and speed according to the language classification. be able to.

Then, the dictionary creating device 22 creates the word notation correspondence table 5 of the voice recognition device 21 as described above. The reading generation unit 9 refers to the recognition word table 23 to change the reading length for each classification when storing the numbers extracted from the word dictionary 8 in the word notation correspondence table 5 for each reading head. In this way, since the reading length of the number is changed for each classification, the word notation correspondence table 5 of the speech recognition device 21 in which the accuracy and speed of the speech recognition are adjusted according to the classification of the language as described above, can be simply processed. Can be created with.

Next, a fourth embodiment of the present invention will be described with reference to FIG.
This will be described below with reference to FIGS. With regard to the fourth embodiment, the same parts as those in the first embodiment described above are designated by the same names and reference numerals, and detailed description thereof will be omitted.

First, as shown in FIG. 14, the voice recognition device 31 and the dictionary creation device 32 illustrated here are also integrally formed. A recognition word table 33, which is a recognition word dictionary, and a digit correspondence table 34, which is a digit correspondence dictionary, are added to the dictionary creating device 32.

The recognized word table 33 has a storage device such as a RAM, and as shown in FIG. 15, as a recognized language, a numeral having a plurality of digits is "400, 410, ...".
However, no reading is set for these numbers. On the other hand, although the number dictionary is stored together with the reading in the word dictionary 8, the storage content is simplified, and as shown in FIG. 16, the one-digit number "4,5, ..." Is stored. However, multi-digit numbers are not stored. The digit correspondence table 34 also has a storage device such as a RAM, and as shown in FIG. 17, the reading of each digit of the numeral is stored in advance for each digit number.

Then, when the reading generation unit 9 generates a reading by using the plural-digit number extracted from the recognition word table 33 as a language, it generates a certain phonetic unit portion from the beginning of the language as the beginning portion of the reading. Therefore, here, the reading of plural digits is generated from the reading of the first digit. The leading digit of the number and the number of digits are determined, the reading of this number is detected from the word dictionary 8, the reading of the number of digits is detected from the digit correspondence table 34, and these are combined and written as a word. Correspondence table 5
To be stored. For this reason, the word notation correspondence table 5 of the voice recognition device 1 stores the numerals in the reading corresponding to only the first digit of the head portion, and the voice recognition unit 4 outputs the voice of the number only in the first digit of the head portion. recognize.

In such a configuration, the voice recognition device 31 of the present embodiment also recognizes a voice uttered by a human as a number stored in the word notation correspondence table 5. In this word notation correspondence table 5, since the numbers are stored in the reading corresponding to only the first digit of the head part, the voice recognition unit 4 recognizes the voice of the number only in the first digit of the head part.

The dictionary creating device 32 creates the word notation correspondence table 5 of the voice recognition device 31 as described above. The reading generation unit 9 extracts from the recognition word table 33 the plural-digit number stored in the word notation correspondence table 5, determines the leading one-digit number and the number of digits of the plural-digit number, and determines the number. Is read from the word dictionary 8 and the reading of the number of digits is detected from the digit correspondence table 34 and combined. For example, if "400" is extracted as a plural digit number, the first digit of the number is "4" and the number of digits is "3", so the reading of this number is "Yon".
And the reading of the number of digits "hyaku" are combined, and the reading at the beginning is "yonhyaku".

As described above, when the model number of a product is used as a language to be recognized, such a numeral generally has a large number of digits. However, when the word dictionary 8 is composed of a general database or the like, there is a high possibility that a single-digit number is stored but a multi-digit number is not stored. Even in such a case, the above-described dictionary creating device 32 can generate a reading of a plural-digit number by a simple process, and the voice recognition device 31 can accurately read the plural-digit number and the word notation correspondence table 5 Can be earned.

Moreover, as described above, the dictionary creating device 32
Generates a reading from the first digit of a multi-digit numeral, so the length of the reading can be made constant with simple processing.
Since this processing is executed for each phonetic unit of the language, it is possible to favorably reflect the characteristics of human speech in the processing operation of the voice recognition device 31. In other words, for ideographic characters such as numbers and kanji, multiple syllables are set for one character, but it is common for utterances to be separated by the ideographic character notation unit, so if this is reflected in reading, recognition accuracy will be improved. Can be improved.

Next, a fifth embodiment of the present invention will be described with reference to FIG.
This will be described below with reference to FIGS. With regard to the fifth embodiment, the same parts as those in the fourth embodiment described above are designated by the same names and reference numerals, and detailed description thereof will be omitted.

First, as shown in FIG. 19, the voice recognition device 41 and the dictionary creating device 42 illustrated here are also integrally formed, and the dictionary creating device 42 has a reading change table 43 which is a reading change dictionary. Has been added.

This reading change table 43 has a storage device such as a RAM, and as shown in FIG. 23, readings of digits that change depending on the combined numeral are stored in advance. Then, when the reading generation unit 9 generates a reading using a plurality of digit numbers extracted from the recognition word table 33 as a language,
The one-digit number reading of the leading part of the number is detected from the word dictionary 8 and the number reading is detected and combined from the digit correspondence table 34. In this case, the reading change table 43 is used to correspond. Correct the reading of the number.

In such a configuration, the dictionary creating device 4
2 creates a word notation correspondence table 5 of the voice recognition device 41. The reading generation unit 9 generates a reading of a plural-digit number by a combination of the reading of the leading one-digit number and the number of digits,
The reading change table 43 is referenced to correct the reading of the corresponding numeral. For example, if "300" is taken out as a plural-digit number, the first digit of the number is "4" and the number of digits is "3", so the reading of this number is "san" and the reading of the number is "hyaku".
When and are simply combined, it becomes “sanhyaku”,
This is corrected to a natural reading, "sanbyaku."

The dictionary creating device 42 described above can generate a reading of a plural-digit number by a simple process, and can correct a reading that is unnatural with a simple combination into a natural form. The voice recognition device 41 can acquire a plural-digit number in the word notation correspondence table 5 together with an accurate reading.

Next, a sixth embodiment of the present invention will be described with reference to FIG.
This will be described below with reference to FIGS. With regard to the sixth embodiment, the same parts as those in the third embodiment described above are designated by the same names and reference numerals, and detailed description thereof will be omitted.

As shown in FIG. 25, the recognition word table 23, which serves as a language classification dictionary, has the classification of the numerical numbers as the language set as "commodity A, commodity B". The length is not set. Then, when setting the numbers read out from the word dictionary 8 in the word notation correspondence table 5 as described above, the reading generation unit 9 corresponds to the readings to be generated when the reading is limited to the head portion by the number of syllables or the like. If the number of the classified languages does not exceed a preset reference value such as "5", the reading is not limited to the head part. Therefore, as shown in FIG. 26, the classification of “product A, product B” is also set for each of the numbers stored in the word dictionary 8, and as shown in FIG.
Although the reading of the quantifiers of the product A set to 1 is restricted to the leading two syllables, the reading of the quantifiers of the product B is not restricted.

In such a configuration, since the dictionary creation device creates the word notation correspondence table 5 of the voice recognition device, the reading generation unit 9 thereof writes the numbers extracted from the word dictionary 8 into the word notation for each head part of the reading. Store in correspondence table 5. At this time, the language corresponding to the generated reading is classified according to the setting of the recognition word table 23, and if the number of the classified languages does not exceed the preset reference value, the reading is not limited to the head part.

For example, when "100" is taken out as a number, the category of this number is detected as "commodity A" from the recognition word table 23, and it is also detected that the number of this number is seven. Since this is more than the standard value of five, the quantifiers of the classification of "product A" are limited to the leading two syllables, and all are set as "hyaku" here. On the other hand, when "200" is taken out as a number and the classification is detected as "commodity B", the number of the classification, two, is less than the reference value of five,
The reading of the quantifiers of the category of “product B” is not limited to the beginning.

As described above, in the speech recognition apparatus in which the number is set, the reading is restricted to the head portion in many languages with the same classification, so that the language having a high appearance frequency is recognized in the head portion of the reading. Therefore, the processing time can be shortened. On the other hand, since the readings of the numbers having the same classification with a small number are not limited to the leading part, the language having a low appearance frequency is recognized in the entire reading, and the recognition accuracy can be improved. Then, the dictionary creating apparatus of the present embodiment can set the above-mentioned languages in the word notation correspondence table 5 by a simple process, so that a high-performance speech recognition apparatus can be created.

Next, a seventh embodiment of the present invention will be described with reference to FIG.
This will be described below with reference to FIGS. Note that, regarding the seventh embodiment of the present embodiment, the same parts as those of the sixth embodiment described above are denoted by the same names and reference numerals, and detailed description thereof is omitted.

As shown in FIG. 28, the recognition word table 23, which is a language classification dictionary, has the classification of the numerical numbers of the languages set as "product A, product B". As shown in FIG. The classification of “product A, product B” is also set for each of the numbers stored in the word dictionary 8. Reading generator 9
When limiting the readings of the numbers read from the word dictionary 8 and set in the word notation correspondence table 5 to the head part, the languages corresponding to the generated readings are classified according to the setting of the recognition word table 23, and the readings in this classification are The length of the leading part of the reading of the language to be generated is changed so that the number does not exceed a preset reference value such as "2".

More specifically, first, all languages of one classification are extracted from the word dictionary 8 according to the recognized word table 23, and the readings thereof are increased by one character from the beginning. When the number of readings exceeds the reference value, the reading immediately before that is adopted. Therefore, as shown in FIG. 30, the word notation correspondence table 5
, The reading of the number of the product A is limited to the leading two syllables, but the reading of the number of the product B is not limited.

In such a configuration, since the dictionary creation device creates the word notation correspondence table 5 of the voice recognition device, the reading generation unit 9 thereof writes the numbers taken out from the word dictionary 8 into the word notation for each head part of the reading. Store in correspondence table 5. At this time, the language corresponding to the generated reading is classified according to the setting of the recognition word table 23, and the reading length is changed so that the number of readings of the classified language does not exceed the reference value.

For example, "100, 110, 120, 125, 127, 130,
When "170" is taken out, these readings are less than the standard value of two because one reading is one "hi". Even if the number of readings is increased by one,
Up to "Hyaku", the number is one and less than the standard value.
However, if the number of reading characters is increased to four, the reading will be five, which is "hyaku", "hyakuji", "hyakuni", "hyakusa", and "hyakuna", which exceeds the standard value of 2. ing. Therefore, in this case, it is adopted just before the number of readings exceeds the reference value, and the reading of the number in the classification of "product A" becomes "hyaku".

On the other hand, "2" is used as a quantifier for classifying "commodity B".
When "00,220" is extracted, the number is two and the same number as the reference value. Therefore, even if the number of reading characters is increased by one, the number does not exceed the reference value. These readings are not limited to the leading part, and "niyaku" and "niyaku niju" are set for each.

Since the number of readings is limited for one classification in the voice recognition device in which the number is set as described above, the number of readings is constant even if there are many languages in one classification, and the processing time is long. Can be shortened. On the other hand, when the number of languages in one category is small, the reading is not limited to the beginning, so that the language with a low appearance frequency is recognized in the entire reading, and the recognition accuracy can be improved.
Then, the dictionary creating apparatus of the present embodiment can set the above-mentioned languages in the word notation correspondence table 5 by a simple process, so that a high-performance speech recognition apparatus can be created.

Next, an eighth embodiment of the present invention will be described with reference to FIG.
A description will be given below with reference to FIGS. With respect to the eighth embodiment, the same parts as those in the third embodiment described above are designated by the same names and reference numerals, and detailed description thereof will be omitted.

As shown in FIG. 31, the recognized word table 23 serving as a condition setting dictionary has continuous conditions set in a plurality of languages. This condition is that the product type name "A type, B type" of the language and the product type number classification "Product A, Product B" of the language are individually indicated by "+" indicating the continuity of the language. They are linked and set as "A type + product A, ...". As shown in FIG. 32,
In the word dictionary 8, the classification and the reading of “product A, product B” are set for each number, but the language of the name “A type,
The reading is set for each of the "B type".

The reading generation unit 9 refers to the recognition word table 23 to read the readings of a plurality of continuous languages when limiting the readings of the numerical words taken out from the word dictionary 8 and set in the word notation correspondence table 5 to the leading part. Is generated, and the reading is restricted to the beginning only in the language located at the end. More specifically, the recognition word table 2
Since the condition of "A type + product A, ..." is set in 3, the reading of multiple consecutive languages that match this is "Etaipu Hyaku", etc., and the number "hyaku" located at the end of it. "Only reading is restricted to the beginning part by the number of syllables. Therefore, as shown in FIG. 33, the word notation correspondence table 5 is set to a continuous language such as “A type 100”, but the reading of the rear number is limited to two syllables. ing.

In such a configuration, since the dictionary creation device creates the word notation correspondence table 5 of the voice recognition device, the reading generation unit 9 thereof writes the numbers taken out from the word dictionary 8 for each head part of the reading. Store in correspondence table 5. At this time, the reading generation unit 9 refers to the recognition word table 23, recognizes the continuous condition of a plurality of languages, and generates the continuous reading of a plurality of languages according to this condition.

For example, the reading of a plurality of consecutive languages that matches the condition of "A type + commodity A, ..." is "Etaipu Hyaku""EtaipuHyakuju""EtaipuHyakuniju" Three are generated. However, the reading of the number "Hyaku, Hyakuju, Hyakuniju" located at the end of the number is limited to the first two syllables.
The reading of a continuous language of is one of the "Etaipu Hyaku".

The continuous languages such as "A type 100" and "B type 200" described above are assumed to be "commodity model + model number", and such a language is continuously breathed in a specific combination. Often vocalized. Therefore, in the voice recognition device of the present embodiment, the recognition accuracy is improved by preliminarily combining a plurality of languages as described above and processing them as one language, and the reading of only the last language is limited to the leading part. Improves the processing speed.
Then, the dictionary creating apparatus of the present embodiment can set the above-mentioned languages in the word notation correspondence table 5 by a simple process, so that a high-performance speech recognition apparatus can be created.

Next, a ninth embodiment of the present invention will be described with reference to FIG.
This will be described below with reference to FIGS. Note that, regarding the ninth embodiment of the present invention, the same parts as those of the eighth embodiment described above are denoted by the same names and reference numerals, and detailed description thereof is omitted.

As shown in FIG. 34, the recognition word table 23, which is a condition setting dictionary, is set with continuous conditions in a plurality of languages. The name A and the product name B "and the category" product A and product B "of the model number of the language are individually connected by" + "indicating the continuity of the language. As shown in FIG. 35, the word dictionary 8 includes "copy," which is the language of the model name.
The classification and reading of "product name A, product name B" are set for each of "facsimile", and the quantifier "100," of the model number is set.
The category and the reading of "commodity A, commodity B" are set for each of "...".

The reading generation unit 9 refers to the recognized word table 23 to read the readings of a plurality of continuous languages when limiting the reading of the numerical words taken out from the word dictionary 8 and set in the word notation correspondence table 5 to the leading portion. Is generated, and the length of the leading part of the reading of the language located at the end is limited so that the entire reading does not exceed the reference value set in advance such as eight characters. More specifically, the recognition word table 23 includes “product name A + product A, ...”
Since the following conditions are set, the readings of multiple consecutive languages that match this will be, for example, “kopyyaku”, and the number at the end of the reading will be “hakuyaku” so that the entire reading does not exceed eight characters. The reading of "is restricted to the beginning.
Therefore, as shown in FIG. 36, the word notation correspondence table 5 is
The reading of "copy 100" is still "kopyyaku", but the reading of "facsimile 100" is restricted to eight characters "fakushimirihya".

In such a configuration, since the dictionary creation device creates the word notation correspondence table 5 of the voice recognition device, the phonetic generation unit 9 thereof writes the numbers taken out from the word dictionary 8 into the word notation for each head part of the phonetic reading. Store in correspondence table 5. At this time, the reading generation unit 9 refers to the recognition word table 23, recognizes the continuous condition of a plurality of languages, and generates the continuous reading of a plurality of languages according to this condition.

For example, when reading a plurality of consecutive languages that match the condition of "commodity name A + commodity A", three words "kopyhyaku", "kopysanbyaku", and "kopysanbyakuniju" are generated. However, since the reading of the last number is limited to the beginning so that the whole number does not exceed eight characters, the entire reading of the above-mentioned plural consecutive languages is “kopyhyaku” or “kopysanbyaku”. It becomes one.

In the speech recognition apparatus of this embodiment, the recognition accuracy is improved by preliminarily combining and processing a plurality of continuously uttered languages, and the reading of only the last language is restricted to the beginning. It improves the processing speed. At this time, since the reading of the language does not exceed the standard value,
Even if the leading language is long, the overall length is constant, and the processing speed is steadily improving. Then, the dictionary creating apparatus of the present embodiment can set the above-mentioned languages in the word notation correspondence table 5 by a simple process, so that a high-performance speech recognition apparatus can be created.

Next, a tenth embodiment of the present invention will be described with reference to FIG.
This will be described below with reference to FIGS. Note that, regarding the tenth embodiment of the present invention, the same parts as those of the ninth embodiment described above are designated by the same names and reference numerals, and detailed description thereof will be omitted.

As shown in FIG. 37, in the recognition word table 23 which is a condition setting dictionary, continuous conditions in a plurality of languages are set such as “product name A + product A, ...”, and further,
The length of the leading part of the reading is set as "2, ..." In this end classification. As shown in FIG. 38, in the word dictionary 8, the classification and the reading of “product name A, ...” are set for each of the “copy ,. The classification and the reading of "commodity A, ..." are set for each of "100, ...".

The reading generation unit 9 refers to the recognition word table 23 to limit the reading of the numerical words taken from the word dictionary 8 and set in the word notation correspondence table 5 to the reading of a plurality of consecutive languages. Is generated, and the length of the beginning part of the reading of the language located at the end is limited to the length set in the recognition word table 23. More specifically, since the condition “product name A + product A” is set in the recognition word table 23, a plurality of consecutive languages that match this condition are read as “kopi nihyaku” or the like. , The reading length of the last number is set to 3 characters, so this reading is “kopynihya”
Is limited to On the other hand, under the condition of "product name B + product B", the reading length of the last number is set to one character, so the reading of "Fakushimari nihyaku" is limited to "Fakushimi ni" .

Therefore, as shown in FIG. 39, in the word notation correspondence table 5, the reading of "copy 100" is "kopihyaku".
However, the reading of "copy 200" is limited to "copying" and the reading of "facsimile 200" is limited to "faximili".

In such a configuration, since the dictionary creation device creates the word notation correspondence table 5 of the voice recognition device, the reading generation unit 9 thereof writes the numbers extracted from the word dictionary 8 into the word notation for each head part of the reading. Store in correspondence table 5. At this time, the reading generation unit 9 refers to the recognition word table 23, recognizes the continuous condition of a plurality of languages, and generates the continuous reading of a plurality of languages according to this condition.

For example, when reading a plurality of consecutive languages that match the condition of "commodity name A + commodity A", three words "kopihyaku", "kopyunihyaku", and "kopyunihyakuniju" are generated. It However, the final number is restricted to the beginning so that it does not exceed three letters, so the total reading of the above-mentioned multiple languages is "kopyhyaku" or "kopynyhyya". Similarly, the readings of multiple consecutive languages that match the condition of "Product name B + Product B" are "Fakushiri Nihyakuju", "Fakushiri Nihyaku niju", and "Fakushiri mihyaku goju". , Etc. are generated, but the number at the end is limited to one character at the beginning, so the whole reading of the above-mentioned multiple languages is "Fakushimari ni""Fakushimirigo" .

In the speech recognition apparatus of the present embodiment, the recognition accuracy is improved by preliminarily combining and processing a plurality of languages which are uttered continuously, and the reading of only the last language is restricted to the beginning. It improves the processing speed. At this time, since the reading length of the last language is changed according to the classification, the longer the beginning language, the shorter the ending language can be made, and the processing speed is stably improved. Then, the dictionary creating apparatus of the present embodiment can set the above-mentioned languages in the word notation correspondence table 5 by a simple process, so that a high-performance speech recognition apparatus can be created.

[0105]

According to the first aspect of the present invention, there is provided a speech recognition apparatus.
The recognition candidate language is stored in advance in the word notation corresponding dictionary for each reading of the head portion, and when the voice to be recognized is input to the voice input means, the voice recognition means causes the voice recognition means to read the beginning portion of the voice. By detecting a language that matches the word from the word notation compatible dictionary, the number of readings that are the target of voice recognition processing is limited to the beginning, so the number is reduced, and this voice recognition processing is executed at high speed. Human utterance tends to be clear only at the beginning of the speech, and since only this beginning is processed, the incidence of false recognition can be reduced and multiple recognition results can be obtained. However, since the correct answer is included with a high probability, it is possible to obtain higher practicality than that in which one wrong recognition result is output in a long-time process.

In the dictionary creating apparatus according to the second aspect of the present invention, various languages are stored in advance in the general language dictionary together with the readings, and the reading generation means selects the language extracted from the general language dictionary for each leading part of the readings. By storing in a dictionary for word notation, multiple languages that have the same beginning but different ends and are likely to cause misrecognition can be aggregated into one reading, and a small number of readings are assigned to many languages for recognition. Since the number of objects to be processed can be reduced, it can contribute to the improvement of the recognition accuracy and processing speed of the speech recognition device, and such a language can be created mechanically from a general language dictionary consisting of an existing database. Humans do not have to perform this task because they can.

In the dictionary creating apparatus according to the third aspect of the invention,
By providing a phonetic unit dictionary in which phonetic units for reading the language are stored in advance, and the reading generation means generates a certain phonetic unit portion from the beginning of the language as the beginning part of the reading, thereby performing a simple process. You can make the length of the beginning of the reading constant,
Since this processing is executed for each phonetic unit of the language, it is possible to favorably reflect the characteristics of human utterance in the processing operation of the voice recognition device.

According to the dictionary creating apparatus of the invention described in claim 4,
The reading generation means can make the length of the reading head constant by a simple process by generating a part of a fixed notation unit from the head of the language as the reading head part. Since the processing is executed in the notation unit, it is possible to favorably reflect the characteristics of human utterance in the processing operation of the voice recognition device.

According to the dictionary creating apparatus of the invention of claim 5,
By providing a length setting dictionary in which the length of the reading head is preset for each language classification, and the reading generation means varies the length of the reading head of the language to be generated for each classification, for example, It is possible to extend the reading length of only a specific language and improve the recognition accuracy of a specific voice without increasing the overall processing time of the voice recognition device. It can be adjusted according to the classification.

According to the dictionary creating apparatus of the invention described in claim 6,
A digit-corresponding dictionary in which the reading of each digit of the numeral is stored in advance for each number of digits is provided, and the reading generation means, when generating the reading of the leading part using a plurality of digits of the language as the language, If the general language dictionary consists of a general database, etc., one digit digit is stored if the reading is detected from the general language dictionary and the reading of the predetermined digit at the beginning is combined from the digit correspondence dictionary. There is a high possibility that a multi-digit number will not be stored, but even in such a case, it is possible to generate a multi-digit number reading by a simple process, so set a multi-digit number in the voice recognition device. be able to.

According to the dictionary creating apparatus of the invention described in claim 7,
A reading change dictionary is provided in which the readings of each digit that changes depending on the combined number are stored in advance, and the reading generation unit refers to the reading change dictionary when generating the reading of the head part using a plurality of digits as a language. By correcting the reading of the corresponding numeral, it is possible to correct the reading that is unnatural with a simple combination into a natural one. Therefore, set multiple digits with appropriate readings in the voice recognition device. You can

According to the dictionary creating apparatus of the invention described in claim 8,
A language classification dictionary having preset language classifications is provided, and the reading generation unit classifies the languages corresponding to the generated readings according to the setting of the language classification dictionary, and the number of the classified languages is a preset reference value. If it does not exceed, the reading is not restricted to the beginning, so that, for example, the reading of a language with a high appearance frequency of voice recognition is limited to the beginning, and the reading of a language with a low appearance frequency is not restricted to the beginning. Thus, it is possible to shorten the processing time of the voice recognition device and improve the recognition accuracy.

According to the dictionary creating apparatus of the invention described in claim 9,
A language classification dictionary in which the classification of languages is set in advance is provided, and the reading generation means classifies the language corresponding to the generated reading according to the setting of the language classification dictionary, and sets the reference value in which the number of readings in this classification is set in advance. By changing the length of the leading part of the reading of the generated language so that it does not exceed, the number of readings is constant even if there are many languages of one classification,
The processing time of the voice recognition device can be shortened, and when the number of languages in one classification is small, the reading is not limited to the beginning part, so that the recognition accuracy of the voice recognition device can be improved.

In the dictionary creating apparatus according to the tenth aspect of the present invention, a condition setting dictionary in which continuous conditions of a plurality of languages are preset is provided, and the reading generation means refers to the condition setting dictionary and a plurality of continuous languages. By limiting the reading to only the language located at the end of the reading, and by combining a plurality of languages expected to be continuous in advance, the recognition accuracy of the voice recognition device can be improved,
By limiting the reading only to the last language to the first part, the processing speed of the voice recognition device can be improved.

In the dictionary creating apparatus according to the eleventh aspect of the present invention, the reading generation means reads the reading of the language located at the end so that the total reading of a plurality of consecutive languages does not exceed a preset reference value. By limiting the length of the head portion, the entire length is constant even if the head language of the combined plural languages is long, so that the processing speed of the voice recognition device can be stably improved.

In the dictionary creating apparatus according to the twelfth aspect of the present invention, the condition setting dictionary is set in advance with the continuous condition of a plurality of languages and the reading length of the language located at the end. , By limiting the reading of the language located at the end to the set length, the reading length of the ending language is changed according to the classification, so that the longer the starting language is, the shorter the ending language is. Therefore, the processing speed of the voice recognition device can be stably improved.

The speech recognition method according to the invention of claim 13 is
The language of the recognition candidate is stored in advance in the word notation corresponding dictionary for each reading of the beginning part, and the language in which the beginning reading matches the beginning part of the speech to be recognized is detected from the word notation corresponding dictionary. As a result, the number of readings to be processed by voice recognition is limited to the beginning, so the number is reduced,
This voice recognition process can be executed at high speed, and human speech tends to be clear only at the beginning of the voice, and since only this beginning is processed, it is possible to reduce the false recognition rate. Even if there are a plurality of recognition results, the correct answer is included with a high probability, so that it is possible to obtain higher practicality than that in which one wrong recognition result is output in a long-time process.

The dictionary creating method according to the fourteenth aspect of the invention is
By extracting a language from a general language dictionary in which various languages are stored in advance with reading, and storing the extracted language in the word notation corresponding dictionary for each head of reading,
Multiple languages that have the same beginning but different ends and are likely to cause misrecognition can be aggregated into one reading, and a small number of readings can be assigned to many languages to reduce the number of recognition processing targets. Therefore, it is possible to contribute to the improvement of the recognition accuracy and the processing speed of the speech recognition apparatus, and since such a language can be mechanically created from a general language dictionary including an existing database, this work is performed by a human. Does not need to run.

An information storage medium according to the fifteenth aspect of the present invention is
A word notation corresponding dictionary in which the language of the recognition candidate is stored in advance for each reading of the head, and a program for causing a computer to detect a language whose head reading coincides with the head of the speech to be recognized from the word notation corresponding dictionary, Since this is written, if the software of this information storage medium is read and operated by a computer, this computer can function as a voice recognition device. Since the number of readings is limited to the beginning, the number is reduced, and this speech recognition processing can be executed at high speed, and human speech tends to be clear only at the beginning of the speech. Since only the beginning part is processed, the occurrence rate of false recognition can be reduced, and even if there are multiple recognition results, the correct answer is included with a high probability, so long-term processing is possible. In it can be obtained one wrong recognition result is higher than what is output practicality.

An information storage medium according to the invention of claim 16 is
A program is written to cause a computer to extract a language from a general language dictionary that is stored in advance with various languages and store the extracted language in a word notation corresponding dictionary for each head of the reading. Thus, if a computer reads the program of this information storage medium and operates it, this computer can function as a dictionary creating device. Multiple languages that are likely to be recognized can be aggregated into one reading, and a small number of readings can be assigned to a large number of languages to reduce the number of recognition processing targets. It is possible to contribute to the improvement of speed and to mechanically create such a language from a general language dictionary such as an existing database. Since it is, there is no need to perform human beings this task.

[Brief description of the drawings]

FIG. 1 is a schematic block diagram showing a voice recognition device and a dictionary creation device according to a first embodiment of the present invention.

FIG. 2 is a block diagram showing hardware of a computer system that realizes a voice recognition device and a dictionary creation device.

FIG. 3 is a perspective view illustrating an appearance of a computer system.

FIG. 4 is a schematic diagram showing stored contents of a word dictionary which is a general language dictionary.

FIG. 5 is a schematic diagram showing stored contents of a word notation correspondence table which is a word notation correspondence dictionary.

FIG. 6 is a flowchart showing a voice recognition method by the voice recognition device.

FIG. 7 is a block diagram showing a voice recognition device and a dictionary creation device according to a second embodiment of the present invention.

FIG. 8 is a schematic diagram showing stored contents of a syllable table which is a phonetic unit dictionary.

FIG. 9 is a schematic diagram showing stored contents of a word notation correspondence table.

FIG. 10 is a block diagram showing a voice recognition device and a dictionary creation device according to a third embodiment of the present invention.

FIG. 11 is a schematic diagram showing stored contents of a recognition word table which is a length setting dictionary.

FIG. 12 is a schematic diagram showing stored contents of a word dictionary.

FIG. 13 is a schematic diagram showing stored contents of a word notation correspondence table.

FIG. 14 is a block diagram showing a voice recognition device and a dictionary creation device according to a fourth embodiment of the present invention.

FIG. 15 is a schematic diagram showing stored contents of a recognition word table.

FIG. 16 is a schematic diagram showing stored contents of a word dictionary.

FIG. 17 is a schematic diagram showing stored contents of a digit correspondence table which is a digit correspondence dictionary.

FIG. 18 is a schematic diagram showing stored contents of a word notation correspondence table.

FIG. 19 is a block diagram showing a voice recognition device and a dictionary creation device according to a fifth embodiment of the present invention.

FIG. 20 is a schematic diagram showing stored contents of a recognition word table.

FIG. 21 is a schematic diagram showing stored contents of a word dictionary.

FIG. 22 is a schematic diagram showing stored contents of a digit correspondence table.

FIG. 23 is a schematic diagram showing stored contents of a reading change table which is a reading change dictionary.

FIG. 24 is a schematic diagram showing stored contents of a word notation correspondence table.

FIG. 25 is a schematic diagram showing stored contents of a recognition word table which is a language classification dictionary of a dictionary creating device according to a sixth embodiment of the present invention.

FIG. 26 is a schematic diagram showing stored contents of a word dictionary.

FIG. 27 is a schematic diagram showing stored contents of a word notation correspondence table.

FIG. 28 is a schematic diagram showing stored contents of a recognition word table of the dictionary creating device according to the seventh embodiment of the present invention.

FIG. 29 is a schematic diagram showing stored contents of a word dictionary.

FIG. 30 is a schematic diagram showing stored contents of a word notation correspondence table.

FIG. 31 is a schematic diagram showing stored contents of a recognition word table which is a condition setting dictionary of a dictionary creating device according to an eighth embodiment of the present invention.

FIG. 32 is a schematic diagram showing stored contents of a word dictionary.

FIG. 33 is a schematic diagram showing stored contents of a word notation correspondence table.

FIG. 34 is a schematic diagram showing stored contents of a recognized word table of a dictionary creation device according to a ninth embodiment of the invention.

FIG. 35 is a schematic diagram showing storage contents of a word dictionary.

FIG. 36 is a schematic diagram showing stored contents of a word notation correspondence table.

FIG. 37 is a schematic diagram showing stored contents of a recognized word table of the dictionary creation device according to the tenth embodiment of the present invention.

FIG. 38 is a schematic diagram showing stored contents of a word dictionary.

FIG. 39 is a schematic diagram showing stored contents of a word notation correspondence table.

[Explanation of symbols]

 1,11,21,31,41 Speech recognition device 2,12,22,32,42 Dictionary creation device 3 Speech input means 4 Speech recognition means 5 Word notation compatible dictionary 8 General language dictionary 9 Reading generation means 13 Phonetic unit dictionary 23 length setting dictionary, language classification dictionary, condition setting dictionary 34 digit correspondence dictionary 43 reading change dictionary 101 computer 103 to 106, 108 information storage medium

────────────────────────────────────────────────── ───

[Procedure amendment]

[Submission date] August 1, 1996

[Procedure amendment 1]

[Document name to be amended] Statement

[Correction target item name] 0036

[Correction method] Change

[Correction contents]

In such a configuration, the voice recognition device 1
Recognizes human voices. More specifically, as shown in FIG. 6, when a voice uttered by a human is input to the voice input unit 3, the voice recognition unit 4 extracts the three characters at the beginning of the voice and sets the start point to the beginning. By fixed spotting, the beginning of the voice and a plurality of words stored in the word notation correspondence table 5
The score is calculated by matching with the reading of
Detect high readings . Since the number output detected in this way is output from the result output unit 6, if there are a plurality of detected number output, the result selection unit 7 manually selects one number.

[Procedure amendment 2]

[Document name to be amended] Statement

[Correction target item name] 0052

[Correction method] Change

[Correction contents]

With such a configuration, the voice recognition device 11 of the present embodiment also recognizes a voice uttered by a human.
At this time, the voice recognition unit 4 and the word notation pair with the beginning of this voice.
Match a plurality of readings stored in table 5 with the score
Calculate and detect the reading with the highest score . Since this voice recognition device 11 uses a syllable , which is a phonetic unit, as a unit for collating voices, it is possible to favorably reflect the characteristics of human speech in the processing operation of voice recognition.

[Procedure 3]

[Document name to be amended] Statement

[Correction target item name] 0058

[Correction method] Change

[Correction contents]

With such a configuration, the voice recognition device 21 of the present embodiment also recognizes a voice uttered by a human.
At this time, the voice recognition unit 4 and the word notation pair with the beginning of this voice.
Match a plurality of readings stored in table 5 with the score
Calculate and detect the reading with the highest score . For this reason, the number "hyaku" classified as the product A is recognized at the second syllable, while the number "nyak" classified as the product B is recognized at the third syllable.

[Procedure amendment 4]

[Document name to be amended] Drawing

[Correction target item name] Fig. 6

[Correction method] Change

[Correction contents]

FIG. 6

Claims (16)

[Claims] 1. A voice input means for inputting a voice to be recognized, a word notation corresponding dictionary in which a recognition candidate language is stored in advance for each reading of a head portion, and a head of a voice input by the voice input means. A voice recognition device, comprising: a voice recognition unit that detects a language in which the reading of the head portion matches the part from the word notation corresponding dictionary. 2. A general language dictionary in which various languages are stored together with readings, and a reading generation means for storing the languages extracted from this general language dictionary in a word notation corresponding dictionary for each head of reading. A dictionary creation device characterized by: 3. A phonetic unit dictionary in which phonetic units for reading a language are stored in advance, and the phonetic generation unit generates a part of a fixed phonetic unit from the head of the language as the head part of the phonetic reading. The dictionary creation device according to claim 2, characterized in that 4. The dictionary creating apparatus according to claim 2, wherein the reading generation means generates a portion of a fixed notation unit from the beginning of the language as the beginning of the reading. 5. A length setting dictionary in which the length of the reading head is preset for each language classification is provided, and the reading generation means varies the length of the reading head of the generated language for each classification. The dictionary creating apparatus according to claim 2, wherein 6. A digit-corresponding dictionary in which the readings of each digit of the numeral are stored in advance for each number of digits, and the reading generation means, when the reading of the leading portion is generated by using the plural-digit numeral as a language, 3. The dictionary creating apparatus according to claim 2, wherein the reading of the predetermined digit of the numeral is detected from the general language dictionary, and the reading of the leading predetermined digit from the digit corresponding dictionary is detected and combined. 7. A reading change dictionary is provided in which readings of each digit that change depending on the combined numeral are stored in advance, and when the reading generation means generates the reading of the head part by using a plurality of digits as a language, the reading is formed. 7. The dictionary creating apparatus according to claim 6, wherein the reading of the corresponding numeral is corrected by referring to the change dictionary. 8. A language classification dictionary in which the classification of languages is preset is provided, and the reading generation means classifies the languages corresponding to the readings to be generated according to the setting of the language classification dictionary, and the number of the classified languages is The dictionary creating apparatus according to claim 2, wherein the reading is not limited to the beginning unless the reference value set in advance is exceeded. 9. A language classification dictionary having a preset language classification is provided, and the reading generation means classifies the language corresponding to the reading to be generated according to the setting of the language classification dictionary, and the number of readings in this classification is preset. 3. The dictionary creation device according to claim 2, wherein the length of the leading part of the reading of the language to be generated is changed so as not to exceed the set reference value. 10. A condition setting dictionary in which consecutive conditions of a plurality of languages are preset is provided, and the reading generation means refers to the condition setting dictionary to generate readings of a plurality of consecutive languages and positions at the end. 3. The dictionary creating apparatus according to claim 2, wherein reading of only the languages to be read is restricted to the head part. 11. The reading generation means limits the length of the leading part of the reading of the language located at the end so that the total reading of a plurality of consecutive languages does not exceed a preset reference value. The dictionary creation device according to claim 10, which is characterized in that. 12. The condition setting dictionary is preset with the reading length of the language located at the end together with the continuous condition of a plurality of languages, and the reading generation means sets the reading of the language located at the end. The dictionary creating apparatus according to claim 10, wherein the dictionary creating apparatus limits the length to a specified length. 13. A recognition candidate language is stored in advance in a word notation corresponding dictionary for each head reading, and a language whose head reading matches the head of the speech to be recognized is extracted from the word notation corresponding dictionary. A voice recognition method characterized by being detected. 14. A dictionary characterized in that various languages are taken out from a general language dictionary stored in advance together with reading, and the taken out languages are stored in a word notation corresponding dictionary for each leading part of reading. How to make. 15. An information storage medium in which computer-readable software is pre-written, wherein a recognition candidate language is pre-stored for each reading of the head part and a word notation corresponding dictionary is stored in the head part of the voice to be recognized. An information storage medium, in which a program that causes the computer to detect a language in which the reading of the head portion matches is detected from the word notation corresponding dictionary. 16. A language is extracted from a general language dictionary in which various languages are stored in advance in an information storage medium in which a program which a computer reads and executes a corresponding operation is written in advance. Store in the dictionary for word notation for each beginning of reading,
An information storage medium, wherein a program for causing a computer to execute the above is written.