JP2016102866A - False recognition correction device and program - Google Patents

Abstract

PROBLEM TO BE SOLVED: To provide a false recognition correction device and program which automatically correct false recognition even when the false recognition has occurred to a result of voice recognition processing.SOLUTION: With a syllable character string obtained by voice recognition as a processing object, among a plurality of key words registered with a data pool, a key word having a syllable character string matching a syllable character string part included in a voice recognition syllable character string or a syllable character string similar to the same is searched for, and by a found key word, a corresponding syllable character string part included in the voice recognition syllable character string is replaced. Key word search is performed by creating a search pattern including a regular expression with respect to each of key words registered with the data pool and searching the voice recognition syllable character strings for a syllable character string part matching the search pattern including the regular expression.SELECTED DRAWING: Figure 13

Description

  The present invention relates to an erroneous recognition correction apparatus and program.

  Processing for recognizing words contained in speech data and converting them into character strings is known. The voice data is converted into a character string by being replaced with words stored in the dictionary data in advance. In order to reduce errors after conversion (increase recognition accuracy), a combination of pronunciation (sound data) and a character string corresponding to the pronunciation is stored (for example, Patent Document 1). ).

Japanese Patent No. 5207642

  However, if noise is mixed in voice data, an error is likely to occur in the converted character string. Misrecognition is likely to occur even when pronunciation is bad. In the speech recognition processing for converting speech data into a character string, no erroneous recognition occurs at all, and the occurrence of erroneous recognition cannot be avoided.

  An object of the present invention is to make it possible to automatically correct a misrecognition occurring as a result of a speech recognition process.

  The erroneous recognition correcting apparatus according to the present invention includes a speech recognition processing means for converting a given speech into a syllable character string, a plurality of words and a word pool storing the syllable character strings in association with each other, and a plurality of words stored in the word pool. Word search means for searching for a word having a syllable character string matching or similar to the syllable character string part included in the speech recognition syllable character string, and a word found by the word search means The word search means includes a regular expression for each of the words stored in the word pool, with word replacement means for replacing the corresponding syllable character string portion included in the speech recognition syllable character string. First search pattern creation means for creating a search pattern, and a search pattern created by the first search pattern creation means Syllable string portion that matches the emission, is characterized in that it includes means for retrieving from the above speech recognition syllable strings. “Syllable” corresponds to a single kana character in Japanese, and “syllable character string” corresponds to a combination of kana characters. The syllable string can also be expressed as “reading”.

  The present invention also provides a program for causing a computer to function as a misrecognition correcting device. This program recognizes the given speech, converts it into a syllable string, and prepares the above-mentioned speech from among a plurality of words stored in a word pool in which a plurality of words and their syllable strings are stored in association with each other. A word having a syllable character string matching the syllable character string part included in the recognized syllable character string or a syllable character string similar thereto is searched, and the corresponding syllable included in the speech recognition syllable character string by the found word. Replace the character string part, where the search for the word creates a search pattern including a regular expression for each word stored in the word pool, and the syllable character string part that matches the created search pattern. , And searching from the speech recognition syllable character string.

  The given voice is converted into a syllable character string by the voice recognition processing means. Correction processing by the erroneous recognition correction device is performed on the converted syllable character string (speech recognition syllable character string).

  Words stored in a word pool prepared in advance are stored in kanji notation, notation including both kanji and kana, kana notation, and the like. The word pool also stores a syllable character string (reading) of the word, which is so-called dictionary data.

  According to the present invention, among words stored in the word pool, a word having a syllable character string that matches a syllable character string part included in the speech recognition syllable character string or a syllable character string similar thereto is searched and searched. The corresponding syllable string part is replaced by the word found by. Here, in the word search process, instead of searching for a plurality of words stored in the word pool, the speech recognition syllable character string is set as a search target. That is, a search pattern including a regular expression is created for each of the words stored in the word pool, and whether or not a syllable character string portion that matches the created search pattern is included in the speech recognition syllable character string. By determining, word search is performed. Since the search pattern includes a regular expression, even if the speech recognition syllable string contains a syllable string part that is not correct as Japanese (for example, a syllable string part not listed in the dictionary), the syllable string The portion can be replaced with a correct Japanese word (words listed in the dictionary) (words stored in the word pool). Of course, a syllable character string portion that is correct as Japanese is likely to be stored in the word pool, and is thus replaced with a correct word. In this way, even if a recognition error occurs by the speech recognition processing means and the speech recognition syllable character string includes a character string portion that does not make sense as Japanese, it is corrected to a character string portion that makes sense as Japanese. be able to.

  In a preferred embodiment, the misrecognition correcting apparatus performs kana-kanji conversion means for converting the speech recognition syllable character string into kana-kanji and whether or not the word stored in the word pool is included in the kana-kanji converted character string. A means for determining, and when it is determined that a word stored in the word pool is included in the character string converted to Kana-Kanji, the speech recognition syllable character string portion excluding the word is Processed by the word search means and the word replacement means. It becomes easy to generate a correct string for Japanese.

  In one embodiment, the search pattern is a search pattern capable of searching for a syllable character string lacking any one of the syllable character strings of the word. Even if speech recognition is performed in a state where one syllable of a correct syllable character string (reading) of a word is missing in the speech recognition processing, it can be corrected (replaced) to the original correct word.

  In another embodiment, the search pattern is a search pattern that can search for a character string in which any one of the syllable character strings of the word is replaced with another character. Even if one syllable of a correct syllable character string (reading) of a word is recognized by an incorrect syllable in the speech recognition process, it can be corrected (replaced) to the original word.

  In still another embodiment, the search pattern is a search pattern capable of searching for a character string in which another character is inserted between characters included in the syllable character string of the word. Even if a wrong syllable is added to the correct syllable string sequence of a word and recognized in the speech recognition process, it can be corrected (replaced) to the original word.

  Preferably, the automatic misrecognition correcting device finds the co-occurrence word pool for each of a plurality of words stored in the word pool and the co-occurrence word pool stored in association with the syllable character string, and the word search means. Keyword selection means for selecting a target word as a target keyword, co-occurrence word specifying means for specifying a co-occurrence word for the target keyword from co-occurrence words stored in the co-occurrence word pool, co-occurrence word specification Co-occurrence word search means for searching for a co-occurrence word having a syllable character string similar to the speech recognition syllable character string portion obtained by removing the target keyword from the voice recognition syllable character string from the co-occurrence words specified by the means , And the co-occurrence words found by the co-occurrence word search means, are included in the speech recognition syllable string. And a co-occurrence word replacing means for replacing the corresponding syllable string portion. If a misrecognized character string portion that could not be found by the word search means still exists in the speech recognition syllable character string, the character string portion can be replaced with a co-occurrence word. A co-occurrence word is a word that is often used together with a word (target keyword) found by the word search means. The misrecognized character string portion that could not be found by the word search means can be corrected (replaced) with a character string that makes sense as Japanese.

  Preferably, the co-occurrence word search unit generates a search pattern including a regular expression for each of the co-occurrence words stored in the co-occurrence word pool, similar to the word search unit described above. Pattern generation means, and means for searching for a syllable character string portion matching the search pattern generated by the second search pattern generation means from the speech recognition syllable character string portion excluding the target keyword.

  In one embodiment, the misrecognition correcting apparatus includes a plurality of field-specific word pools each storing only words frequently used in a specific field, and includes word search processing by the word search means and syllable character strings by the word replacement means. The partial replacement process is executed using each of the plurality of field-specific word pools. It is easy to generate correct processing results for Japanese. Preferably, also for the co-occurrence word pool, a plurality of field co-occurrence word pools corresponding to each of the plurality of field word pools are prepared, and the co-occurrence word specifying process by the co-occurrence word specifying means, The co-occurrence word search process by the word occurrence search means and the syllable character string portion replacement process by the co-occurrence word replacement means are executed using each of the plurality of field-specific co-occurrence word pools. By using each of a plurality of field-specific word pools and a plurality of field-specific co-occurrence word pools, processing results (misrecognition correction results) for each field are generated. All of the generated multiple processing results may be presented to the user, or a more probable processing result among the multiple processing results is automatically selected according to a predetermined rule and presented to the user. May be.

The external appearance of the smart phone which comprises a misrecognition automatic correction system is shown. It is a block diagram which shows the electric constitution of the server computer which comprises a misrecognition automatic correction system. 3 shows a data pool stored in a storage device. An example of a target keyword data pool is shown. An example of a co-occurrence word data pool is shown. It is a flowchart which shows a misrecognition correction process. It is a flowchart which shows a part of misrecognition correction process (kana-kanji data utilization process) in detail. It is a flowchart which shows a part of misrecognition correction process (syllabic character string utilization process) in detail. It is a flowchart which shows a part in detail (the search pattern creation process of the syllable character string utilization processes) of the misrecognition correction process. It is a flowchart which shows a part in detail (the search pattern creation process of the syllable character string utilization processes) of the misrecognition correction process. It is a flowchart which shows a part in detail (character string search process using the search pattern of syllable character string utilization processes) of a misrecognition correction process in more detail. (A) shows the result of Kana-Kanji conversion of a misrecognized syllable character string, (B) shows the result of morphological analysis of the Kana-Kanji converted character string, and (C) shows a part of the Kana-Kanji conversion character string. Each is shown. This shows how the search pattern is created. Indicates the created search pattern. A search pattern is created for each target keyword registered in the target keyword / data pool. (A)-(E) show the state of the character string search process using a search pattern. (A)-(D) show the state of the character string search process using a search pattern. It is a flowchart which shows a part of misrecognition correction process (co-occurrence word utilization process) in detail. A search pattern is created for each co-occurrence word. The flow of a category determination process is shown. It is a flowchart which shows the misrecognition correction process of a modification.

1. Hardware Configuration of Misrecognition Automatic Correction System FIG. 1 shows an external appearance of a smartphone used as a client in a misrecognition automatic correction system. FIG. 2 is a block diagram showing an electrical configuration of a server computer (hereinafter referred to as a server) that communicates with the client in the automatic recognition correction system.

  The erroneous recognition automatic correction system is constituted by a client-server model. Referring to FIG. 1, a client (smart phone) 1 includes a touch panel 2 that functions as an input device and a display device 2, a microphone 3 that converts voice into an electrical signal, a communication device (not shown) that transmits and receives data to and from a server, and the like. ), A processing device (not shown) for comprehensively controlling the operation of the client 1, a storage device (not shown) for storing programs and data, and the like. When the microphone-shaped icon 4 a displayed on the touch panel 2 is clicked, a program for accepting voice input is started in the client 1 and the voice input window 4 is displayed on the touch panel 2. Audio data received by the microphone 3 is input to the client 1.

  The input voice data is transmitted from the client 1 to the server 10 through a network (such as the Internet) and received by the communication device 9 of the server 10. Referring to FIG. 2, a server 10 includes a central processing unit (CPU) 5 that controls the operation of the server 10 in an integrated manner, a touch panel that functions as an input device and a display device 6, and a memory 7 that temporarily stores data. , A program for executing an erroneous recognition automatic correction process to be described later, a storage device 8 for storing a data pool and the like to be described later, and a communication device 9 for performing data communication with the client 1. As will be described later, the server 10 performs automatic recognition error correction processing. Even if the syllable character string (hereinafter referred to as reading data) recognized based on the voice data transmitted from the client 1 includes a grammatical or linguistic error, it is corrected (corrected) and corrected. The character string data thus transmitted is transmitted to the client 1.

  3 to 5 show data (data pool) used for the erroneous recognition automatic correction process executed in the server 10. The data shown in FIGS. 3 to 5 are stored in advance in the storage device 8 provided in the server 10.

  Referring to FIG. 3, the storage device 8 stores a set of target keyword data pools 11, 13, 15 and co-occurrence word data pools 12, 14, 16 for each category. An example will be described. 4 shows the data contents of the target keyword / data pool 11 of category 1 related to gourmet, and FIG. 5 shows the data contents of the co-occurrence word data pool 12 of category 1 related to gourmet, respectively.

  Referring to FIG. 4, a large number of target keywords (words) related to gourmet are stored (registered) in the target keyword data pool 11. Each of the target keywords includes identification code (number), prototype, reading data (syllable character string), and part of speech data. “Prototype” uses kanji if it is generally expressed by or including kanji, and katakana if it is generally expressed by or including katakana. , Represents a word about gourmet. “Reading data” is data representing a syllable character string. In this specification, the reading data (syllable character string) is expressed in katakana. “Part of speech” stores either a noun, an adjective, an adjective verb, a verb, or an adverb according to the word represented by the prototype.

  Referring to FIG. 5, the co-occurrence word data pool 12 paired with the gourmet target keyword data pool 11 is the same for each of the many target keywords stored in the gourmet target keyword data pool 11. Many words (co-occurrence words) that are frequently used together in a document (web page) are stored (registered). The identification code (number), the target keyword (prototype), the co-occurrence word for the target keyword and the co-occurrence word reading data are stored as a set of data. In general, a plurality of co-occurrence words are stored for one target keyword (prototype).

  By using the target keyword / data pool 11 and the co-occurrence word data pool 12 related to gourmet, one processing result (character string after erroneous recognition correction processing) is generated as described in detail below. Is output. Similarly, a set of target keyword data pool 13 and co-occurrence word data pool 14 relating to tourism (category 2) (see FIG. 3), target keyword data pool 15 relating to accommodation facility (category N) and co-occurrence word data pool By using 16 (see FIG. 3), the processing results are also generated and output. That is, the erroneous recognition automatic correction system generates a processing result for each category. Since a plurality of processing results for each category are all generated by the same algorithm, in the following embodiment, a set of a target keyword data pool 11 (FIG. 4) and a co-occurrence word data pool 12 (FIG. 5) relating to gourmet is created. Processing of the server 10 when used (error recognition automatic correction processing) will be described.

2. Overall Flow of False Recognition Automatic Correction Processing FIG. 6 is a flowchart showing the overall flow of false recognition automatic correction processing executed in the server 10. First, the erroneous recognition automatic correction processing executed in the server 10 will be schematically described with reference to FIG.

  As described above, the audio data input from the client 1 is transmitted to the server 10 via the network and input (received) to the server 10 (step 21).

  Speech recognition processing, that is, processing for converting speech data into syllable character string data (reading data) is executed (step 22). For example, commercially available voice recognition software can be used for the voice recognition processing.

  The syllable character string data generated by converting the sound data by the speech recognition processing may have a grammatical or linguistic error as Japanese. In this embodiment, it is assumed that the syllable character string that is the result of the speech recognition processing is “cocaraune wotabaitai”, that is, that it has been misrecognized, although “I want to eat eel from now on” is spoken. . The above-described misrecognition may be caused by poor pronunciation of a person who speaks, or may be caused by a lot of noise included in voice data. In any case, the misrecognition correction process in the server 10 will be described by taking as an example a state in which a syllable character string that is not correct in Japanese (not meaningful) is generated by the speech recognition process.

  The result of the speech recognition process (the syllable character string of “Coca Raunegi Otabaitai”) (hereinafter referred to as “misrecognized syllable character string”) is the target data for the erroneous recognition correction process. The following two processes are performed on the misrecognized syllable character string.

  One of them is a kana-kanji conversion process on a misrecognized syllable string, and the target keyword data pool 11 described above is included in the kanji-mixed string generated by the kana-kanji conversion process. Is a process of searching whether or not the target keyword registered in is included and marking the found character string portion (steps 23 to 25). Hereinafter, the processing of steps 23 to 25 using the kana-kanji conversion data is referred to as “kana-kanji conversion data use processing”. For example, a character string of “I want to eat a request from an individual” is generated from a syllable character string of “Coca Raunegiota Tabaitai”, and “Eat” (eat) is registered in the target keyword / datapool 11 If it is, the character string portion of “eat” is marked as a target keyword found by the kana-kanji conversion data use processing. This process is performed in order to improve the accuracy of the process result finally determined (details will be described later).

  The other one matches the partial syllable character string part included in the misrecognized syllable character string ("Coca Raunegiwo Tabaitai") among the target keywords registered in the target keyword data pool 11. Alternatively, a target keyword having a syllable character string portion (reading) similar to this is searched, and the corresponding syllable character string portion in the misrecognized syllable character string is replaced with the found target keyword (prototype) (step) 26, 27). Hereinafter, the processing of steps 26 and 27 using the syllable character string is referred to as “syllabic character string use processing”. Through the syllable character string use processing, a character string of “Kokara-wo-eaten-tai” is generated from the misrecognized syllable character string “Koca Raunegi wota-tai”. In addition, the character string portion replaced by the syllable character string use processing, here “鰻” and “eat” is marked as the character string portion (target keyword) replaced by the syllable character string use processing. . Details of the syllable character string use processing will also be described later.

  The character string part included in the character string generated by the syllabic character string use processing (Kokora 鰻 eat tie) is replaced by the syllable character string use processing, and in the kana-kanji conversion data use processing It is determined whether or not there is a marked character string part, and it is labeled that such a character string part is found in the kana-kanji conversion data utilization processing (step 28). For example, in the character string “Kokhara-wo-eaten-tai” generated by syllable character string use processing, the character string portion of “eat” is replaced by syllable character string use processing, and kana-kanji conversion data is used. If the character string portion is marked in the process, the character string portion of “eat” is labeled as found in the kana-kanji conversion data use processing. This process is also performed in order to improve the accuracy of the finally determined process result (details will be described later).

  Of the character strings generated so far, the syllable character string part other than the target keyword matches the syllable character string part from the co-occurrence words registered in the co-occurrence word data pool 12. Alternatively, a co-occurrence word having a syllable character string (reading) similar to this is searched, and the corresponding syllable character string part is replaced with the found co-occurrence word (prototype) (steps 29 and 30). The data pool used for data retrieval is different, but the algorithm is the same as the above-described syllable character string use processing (steps 26 and 27). By this processing, the syllable character string portion of “Kokhara” is replaced with the co-occurrence word “from now on”, and a character string “from now on” is generated.

  When the last remaining syllable character string portion is confirmed as it is (using the expression as it is), the erroneous recognition automatic correction processing is ended (step 31). The character string “I want to eat rice cake from now on” is generated.

  Through the above processing, the processing result (misrecognition correction character string) of “I want to eat potatoes from now” is generated from the misrecognized syllable character string “Coca Raunegiwo Tabaitai”. The generated misrecognition correction sentence is a sentence correct in Japanese that cannot be obtained by simply kana-kanji conversion of the misrecognized syllable character string. Character string data representing the generated erroneous recognition correction sentence is transmitted from the server 10 to the client 1. In the client 1, the character string transmitted from the server 10 is used, and the process proceeds to subsequent processing (for example, web page search processing using a search engine).

  Hereinafter, the above processing will be described in detail with reference to the flowcharts of FIGS. 7 to 11 and 18 and the specific examples shown in FIGS. 12 to 17 and FIG.

3. Kana-Kanji Conversion Data Utilization Processing FIG. 7 is a flowchart showing in detail the processing of step 24 in FIG. Here, it is assumed that a kana-kanji conversion process is performed on the misrecognized syllable character string of “Kokaraunegiwo Tabaitai”, and a character string mixed with kanji “I want to eat debate from an individual” is generated. The process shown in FIG. 7 will be described with reference to FIGS.

  First, a morphological analysis process is performed on a character string mixed with kanji after the kana-kanji conversion process (step 41). In the case of a character string mixed with kanji “I want to eat an applause from an individual” (FIG. 12A), “Individual”, “From”, “畝”, “Meeting”, “O”, “Eat” and “ Is divided into morphemes (minimum units), and each reading, prototype, and part of speech are determined (FIG. 12B). For the morphological analysis processing, a morphological analysis system “Chasen” or the like can be used.

  Of the divided morphemes, morphemes having nouns, adjectives, adjective verbs, verbs and adverbs are selected (step 42). In the above example, four morphemes of “individual”, “mochi”, “discussion” and “eat” are selected. For each selected morpheme, it is determined whether the morpheme (word) is registered as a target keyword in the gourmet target keyword data pool 11 (see FIG. 4) (step 43).

  If the target keyword is included in the selected morpheme, the target keyword for which the morpheme (target keyword) was found using the Kana-Kanji conversion data utilization process as described above Marking is performed (FIG. 6, step 25). For example, among four morphemes of “individual”, “mochi”, “discussion”, and “eat”, “eat” (prototype “eat”) is registered in the target keyword / data pool 11 as a target keyword. The character string portion “eat” is marked to indicate that it is a target keyword found using the kana-kanji conversion data utilization processing. This is schematically shown in FIG.

  If not all of the selected morphemes are registered in the target keyword / data pool 11, no special processing is performed.

4). Syllable string usage processing (steps 26 and 27)
FIG. 8 is a flowchart showing in detail the processing of step 26 in FIG. 9 to 11 are flowcharts showing the process shown in FIG. 8 in more detail. The process shown in FIG. 8 will be described with reference to specific examples shown in FIGS.

  First, processing for creating a search pattern (search formula) is performed for each of a plurality of target keywords stored in the target keyword / data pool 11 related to gourmet (FIG. 4) (step 51). This search pattern creation process not only determines whether the character string part that completely matches the target keyword is included in the misrecognized syllable character string (cocaraunegiotatabai), but also similar characters. This is a process performed to determine whether or not a column part is included. Through this process, as shown in FIG. 15, one or a plurality of search patterns are created for each target keyword registered in the target keyword / data pool 11.

  With reference to FIGS. 9 and 10, the search pattern creation processing in step 51 of FIG. 8 will be described in detail.

  First, a counter i for specifying each target keyword is incremented (step 60), and one target keyword is read out to the memory 7 among a plurality of target keywords stored in the target keyword / data pool 11 ( Step 61). For example, the first (No. 1) target keyword (prototype “鰻”) among a plurality of target keywords stored in the target keyword / data pool 11 shown in FIG. Is read out.

  Half the number of characters T of the read data of the read target keyword is calculated, and the number is set as the effective bit number B (step 62). Details of the effective bit number B will be described later. If the number of characters T in the reading data of the target keyword is an odd number, the rounded integer is set as the effective bit number B. For example, when the read data of the read target keyword is “eel”, since the number of characters is “3”, “2” is calculated as the effective bit number B. When the number of characters T of the read data of the read target keyword is 2 characters, the effective bit number B is exceptionally 2 characters. If the number T of characters read from the read target keyword is 1, the target keyword may be ignored (the target keyword is not read).

The maximum processing number P is calculated (step 63). As will be described later, the process of trying to create a search pattern is repeatedly performed by the number depending on the number of characters T of the read data. The maximum number of processing times P is calculated by 2 ^T −1. For example, if the read data of the read target keyword is “eel”, since the number of characters T of the read data is 3, “7” (= 2 ³ −1) is calculated as the maximum processing count P.

  An area for storing the search patterns for the maximum number of processing times described above is secured in the memory 7 and is initialized (step 64).

  The processing counter C is incremented (step 65), and it is determined whether the counter C is smaller than the maximum processing count P (step 66).

  If the counter C is smaller than the maximum processing count P (YES in step 66), the value of the counter C is converted to a binary number (FIG. 10, step 70). For example, in the case of the first processing, the counter C = 1, and this binary number is “001” (see FIG. 13). The number of bits of the binary digitizing counter C is matched with the number of characters T of the read data of the target keyword to be processed.

  The number N of bits “1” in the binary digitizing counter C is acquired (step 71), and it is determined whether the number N of bits “1” is equal to or greater than the effective bit number B described above (step 72). If the number N of bits “1” is equal to or greater than the number of valid bits B, the process proceeds to search pattern creation processing (YES in step 72). When the number N of bits “1” is less than the effective bit number B, the process does not proceed to the search pattern creation process, and the counter C is incremented (NO in step 72, step 76). This is a process for avoiding creation of a search pattern that will be hit if only one character matches in a search pattern (FIG. 13) created as described later. For example, if the binary digitizing counter C is “001”, the number N of bits “1” included therein is “1”. If the number of effective bits B = 2 calculated using the number of characters T of the read data is N (= 1) <the number of effective bits B (= 2), the number of bits “1” is smaller than the number of effective bits B (= 2). The counter C is incremented without progressing.

  After the counter C is incremented, the processing from step 66 is repeated. When the counter C = 2, the binarization counter C is “010”. Also in this case, since the number N (= 1) of bits “1” <the number of effective bits B (= 2), the counter C is further incremented without proceeding to the search pattern creation process (NO in step 72, step 76, step 66, FIG. 13).

  When the counter C = 3, the binary counter C is “011”. Since the number of bits “1” N = 2 and the number of bits “1” N (= 2) ≧ the number of effective bits B (= 2), the process proceeds to search pattern creation processing (YES in step 72).

  A character pattern is created using the binary digitizing counter C (step 73). This will be specifically described with reference to FIG. It is assumed that the reading data of the target keyword to be processed is “eel” and the binary counter C is “011”. In this case, “○ Nagi” is created as a character pattern. That is, for the bit 0 included in the binary digitizing counter C “011”, “0” is associated with the position, and the read data (syllable) at that position is directly associated with the bit 1 as the character to be created. It is a pattern. As described above, the binary digitizing counter C has the same number of bits as the number of characters T of the read data of the target keyword. Therefore, depending on the value of the binary digitizing counter C (which may be referred to as flag ON / OFF), A character pattern will be created.

  Next, a search pattern is created using the created character pattern (step 74). The search pattern is performed as follows.

(1) Ignore ○ included in the character pattern.
(2) If characters other than ○ are consecutive, add a regular expression (. *?) Between the characters.
(3) If regular expressions (. *?) Continue in the process of (2), one regular expression (. *?) Is assumed.

  The meaning of regular expressions (. *?) Will be described later.

  Referring to FIG. 13, for example, if the character pattern is “○ Nagi”, “Na (. *?) G” is created. The created search pattern is stored in the memory 7 (step 75), then the counter C is incremented (step 76), and the processing from step 66 is repeated.

  Referring to FIG. 13, when the above processing is performed with the target keyword whose reading data is “eel” as the processing target, as described above, in the processing when counter C = 1 and counter C = 2, the search pattern Is not created, and the search pattern “NA (. *?) GI” is created in the process when the counter C = 3. In the processing when the counter C = 4, the search pattern is not created. In the processing when the counter C = 5, 6 and 7, the search patterns “U (. *?)”, “U (. *?) "And" U (. *?) Na (. *?) Gi "are created respectively.

  When the counter C is further incremented to C = 8, the counter C exceeds the maximum number of processing times P (= 7) calculated according to the number of characters T of the read data (NO in step 66). The search pattern created in the process so far (Fig. 13) is rearranged in the descending order of the number of characters other than the regular expression (. *?) (Step 67), thereby completing the process for the target keyword "鰻". To do. The search pattern after the rearrangement is stored in the memory 7 together with the target keyword prototype, reading data, the number of characters of the reading data, and the number of characters included in the search pattern (FIG. 14).

  Next, the counter i is incremented, and the process proceeds to the process for the next target keyword registered in the target keyword / data pool 11 (NO in step 68 and step 69). When the above process is completed for all target keywords registered in the target keyword / data pool 11, the search pattern creation process ends (YES in step 69).

  FIG. 15 shows search patterns created for the target keywords registered in the target keyword / data pool 11 shown in FIG.

  Returning to FIG. 8, the created search pattern is used, and a process of searching for a character string matching the search pattern from the erroneously recognized syllable character strings is performed (step 52).

  FIG. 11 is a flowchart showing details of character string search processing using a search pattern.

  First, a search pattern including the regular expression (. *?) Described above will be described.

  A regular expression (. *?) Represents a pattern match character string. There can be no characters in the (. *?) Part, and the (. *?) Part can be any number of characters (however, the character length Represents the shortest part).

For example, if the search pattern “u (. *?) Na (. *?) Gi” includes “eel” in the search target syllable character string, that portion is hit. In addition, if a character string that is the target of a search is "a Sono Umano Naki Goe two-Gyo Ten Sul" (astounded to cries of Aso of the horse), part of the "c Mano Na Kigoeni formate" is a hit. Similarly, the search pattern “NA (. *?) GI” is “Nagi”, “Nice gift” (hit part is “Nice cedar”), “Naminorigal” (hit part is “Namino rigi”). )) Etc., it hits.

  Referring to FIG. 11, the character string search process using the search pattern is performed in parallel using the search pattern created for each of the plurality of target keywords registered in the target keyword / data pool 11.

  Referring to the leftmost process in the flowchart of FIG. 11, this process uses a search pattern created from the target keyword “鰻” that is first registered in the target keyword data pool 11 (FIG. 4). The search process is performed, and then the search process is performed using the search pattern created from the second target keyword “鯛” registered in the target keyword / data pool 11, and finally the last (Mth) The search processing is performed using the search pattern created from the target keywords registered in (in the case of the target keyword / data pool 11 shown in FIG. 4, the 12th “decrease”). In the second process from the left in FIG. 11, first, a search process is performed using a search pattern created from the target keyword “鯛” registered second in the target keyword / data pool 11, and then the third process. The search process is performed using the search pattern created from the target keyword “Negiyaki” registered in, and finally the search pattern created from the target keyword “キ ー ワ ー ド” registered first is used. Search processing is performed.

  One of the parallel processes (the leftmost process in FIG. 11) will be described. The counter j is incremented (step 70), and each of the search patterns (see FIGS. 14 and 15) created from the target keyword “鰻” registered first in the target keyword / data pool 11 is used. A character string search is performed on the erroneously recognized syllable character string “Coca Raunegi wotabetai”. From the target keyword “鰻”, “U (. *?) Na (. *?) Gi”, “Na (. *?) Gi”, “U (. *?) Gi” and “U (. *? Four search patterns using the regular expression “)” have been created. For all four search patterns, it is determined whether or not the character string portion matching the search pattern is included in the misrecognized syllable character string (steps 71a and 72a).

  A specific description will be given with reference to FIGS. For the misrecognized syllable character string "Kokalauunegiwotabaitai", "U (. *?) Na (. *?) Gi", "Na (. *?) Gi" and "U (. *?) Na" The search pattern “U (. *?) GI” hits the “UNIGI” part (FIG. 16A).

  It is determined whether or not the hit character string portion follows the verb (step 73a). Whether the character string portion in front of the hit character string portion is a verb is determined using, for example, dictionary data. If the hit character string portion is not the character string portion following the verb, the target keyword (“鰻” in the above case) corresponding to the hit character string portion is stored in the memory 7 and the hit character string portion Is removed from the misrecognized syllable character string (step 74a). The counter j is incremented, and the process proceeds to the process using the search pattern created from the next target keyword “鯛” (NO in steps 75a and 76a). Even when the character string portion matching the search pattern is not found, the counter j is incremented and the process proceeds to the process using the search pattern created from the next target keyword (NO in step 72a, and in steps 75a and 76a). NO, step 71a).

  Each of the character string portions “kokara” and “wotabeitai” obtained by removing the hit character string portion “unigi” from the misrecognized syllable character string “cocaraunegiwotabaitai” is processed (FIG. 16B). The part of “Thailand” included in “Otabetai” is hit by the search pattern “Ta (. *?) I” created from the target keyword “鯛”. However, since the hit character string portion “Thai” is a character string portion following the verb “Tab”, the target keyword “鯛” is not saved in the memory 7 and is not removed (YES in Step 73a). . A similar process is performed using a search pattern created from the next target keyword “green onion” and the next target keyword “ramen” (step 75a, NO at step 76a, step 71a).

  The character string portion of “Tabe” included in “Otabetai” is hit by the search pattern “Tabe (. *?) Be” created from the target keyword “Eat” (FIG. 16C). The target keyword (“eat” in the above case) corresponding to the hit character string portion is stored in the memory 7. In addition, the hit character string portion is removed from the misrecognized syllable character string (step 74a).

  The character string portion of “tab” is further removed, and the rest becomes “kokara”, “wo”, and “tie” (FIG. 16D). A search process using a search pattern created from the remaining target keywords is performed for each of these character string portions (NO in steps 75a and 76a, 71a). If no character string is hit, the search process using the search pattern ends there (YES in step 76a). Eventually, the two target keywords “鰻” and “eat” are stored in the memory 7 (FIG. 16E).

  Each parallel processing may not produce the same processing result. FIGS. 17A to 17D show the flow of processing when a character string search is performed using a search pattern created from the target keyword “green onion” prior to the target keyword “鰻”. Yes. When a search process using a search pattern created from the target keyword “green onion” is performed, the character string portion of “green onion” in the misrecognized syllable character string “Kokalau Negiwo Tabai” is hit. 17B and 17C showing the subsequent processing are the same as FIGS. 16C and 16B described above. Eventually, two target keywords of “green onion” and “eat” are stored in the memory 7 (FIG. 17D).

  When two or more search results are generated by the parallel processing as described above, the target keyword to be used for replacing the character string in the misrecognized syllable character string is selected according to the following rule (see FIG. 8, Step 53).

(1) Use search results with a large number of target keywords.
If the number of target keywords in two or more search results is the same,
(2) If the same target keyword is included, the target keyword is determined as a target keyword used for replacement. The search result shown in FIG. 16 (E) and the search result shown in FIG. 17 (D) include the same target keyword “eat”. The target keyword “eat” is determined as the target keyword to be replaced.
(3) When different target keywords are included, a target keyword having a large number of characters included in the hit search pattern is determined as a target keyword to be replaced.
(4) If the number of characters included in the searched search pattern is the same, the target keyword to be replaced is determined to have the smaller number of target keyword readings. In the search result shown in FIG. 16E and the search result shown in FIG. 17D, the target keyword “タ ー ゲ ッ ト” is determined as the target keyword to be replaced instead of the target keyword “green onion”.

  Returning to FIG. 6, the character string portion corresponding to the determined target keyword in the erroneously recognized syllable character string “Coca Raunegiwo Tabaitai” is replaced (step 27). In the above example, a character string of “Kokhara 鰻 食 べ eat Thai” is created. As described above, the replaced target keyword is marked as a character string portion replaced by the syllable character string use processing.

  After the above-described labeling process (step 28 in FIG. 6), the process proceeds to the process for the syllable character string part other than the target keyword in the character string generated by the process so far.

5. Co-occurrence word processing (steps 29 and 30)
FIG. 18 is a flowchart showing in detail the processing of step 29 in FIG.

  First, among the plurality of co-occurrence words registered in the co-occurrence word data pool 12 (FIGS. 3 and 5), the co-occurrence words for the target keywords “鰻” and “eat” determined as described above are read out. (Step 81). Note that the same co-occurrence words and duplicate co-occurrence words with the determined target keyword are deleted without being read (step 82).

  A search pattern is created for each of the read co-occurrence words (step 83). The search pattern creation processing algorithm is the same as the search pattern creation processing algorithm for each of the target keywords described above (see FIGS. 9 and 10). As shown in FIG. 19, one or a plurality of search patterns are generated for each co-occurrence word.

  Next, a character string search process is performed on the remaining character string portion obtained by removing the target keyword from the misrecognized syllable character string using a search pattern created from the co-occurrence words (step 84). This process is also performed by the same algorithm as that of the character string search process using the search pattern created from the target keyword described above (see FIG. 11). Search processing using a search pattern created from each of the read co-occurrence words is performed in parallel.

  The character string portion of the “Kokhara 鰻 wo eat Thai” excluding the target keywords “鰻” and “eat” is “Kokhara”, “wo” and “Thai”. Co-occurrence words that match or are similar to these character string portions are searched.

  For example, among the plurality of search patterns created from the co-occurrence word “from now on”, the character string portion of “Kokhara” is hit by the search pattern of “ko (. *?) Ka (. *?) La”. The hit character string portion “Kokhara” is replaced by the co-occurrence word “from now on” (FIG. 6, step 30).

  The character string portion that has not been replaced through the processing so far is left as it is (FIG. 6, step 31). That is, the non-replaced character string parts “wo” and “tie” are used as they are. As a result, the character string “I want to eat rice cake” is completed.

  In addition to the misrecognition correction process using the target keyword data pool 11 and the co-occurrence word data pool 12 in which the target keyword related to the category “gourmet” is stored, the target keyword related to the category “tourism” is stored as described above. Misrecognition correction processing using the target keyword / data pool 13 and the co-occurrence word data pool 14, and the target keyword / data pool 15 and the co-occurrence word data pool 16 storing the target keywords related to the category “accommodation facility” were used. The erroneous recognition correction processing is also performed in parallel (see FIG. 3).

  FIG. 20 shows a character string (category) to be selected as a final result among a plurality of post-processing character strings (analysis result, erroneous recognition correction result) created by performing the erroneous recognition correction process for each category. The process to determine is shown.

  In the category determination process (step 91), the character string (category) to be selected as the final result is determined by the following algorithm.

(1) A character string portion marked in the kanji conversion data utilization process is selected as the final result.
(2) If the number of character string parts marked in the Kanji conversion data use processing is the same, the one with the large number of target keywords replaced in the syllable character string use processing is selected as the final result.
(3) If the number of target keywords replaced in the syllable character string use processing is also the same, the one having a large number of co-occurrence words replaced in the co-occurrence word use processing is selected as the final result.

  In general, in the case of audio data including words relating to gourmet, the character strings obtained using the data pools 11 and 12 relating to gourmet are easily selected as the final result. However, all of the processed character strings (misrecognition correction results) for each category may be presented to the user.

  In the embodiment described above, a search pattern (search pattern for target keywords and search pattern for co-occurrence words) is created in the erroneous recognition correction process. , 12 etc. may be stored.

  FIG. 21 is a flowchart showing a modified example of the erroneous recognition automatic correction process executed in the server 10. The same processes as those in the flowchart shown in FIG. 6 are denoted by the same reference numerals to avoid redundant description.

  In the erroneous recognition automatic correction process of the above-described embodiment, the kana-kanji conversion use process is used for marking (step 25 in FIG. 6) and determining a category (step 91 in FIG. 20). 26 and 27) target the entire erroneously recognized syllable character string obtained by the speech recognition process (step 22). In the process shown in FIG. 21, when a target keyword is found in the Kana-Kanji conversion use process, the found target keyword is always included in the final process result, and the remaining syllable character string portion is processed as a syllable character string use process (step 26, 27) is performed.

  The kana-kanji conversion of the misrecognized syllable character string “Koca Raunegiwo Tabaitai” generates “I want to eat a debate from the individual”. When “eat” is stored, the character string portion “eat” is determined to be used as the final processing result at that time. The portion other than the character string portion “eat” is returned to the reading data (step 95). Except for the character string portion “eat”, the reading data of “Coca Raunegiwo” and the reading data of “Thai” are the processing target of the syllable character string use processing (steps 26 and 27). The burden of syllable character string use processing can be reduced.

1 Client (smartphone)
3 Microphone 5 CPU (voice recognition processing means, word search means, word replacement means, first and second search pattern creation means, keyword selection means, co-occurrence word search means, co-occurrence word replacement means)
7 Memory 8 Storage device
10 Server computer
11, 13, 15 Target keyword data pool (word pool)
12, 14, 16 Co-occurrence word data pool (co-occurrence word pool)

Claims (10)

Speech recognition processing means for converting a given speech into a syllable string;
A word pool in which multiple words and their syllable strings are stored in association with each other,
A word search means for searching for a word having a syllable character string that matches a syllable character string portion included in the speech recognition syllable character string or a syllable character string similar to the syllable character string part of the plurality of words stored in the word pool , And word replacement means for replacing the corresponding syllable character string part included in the speech recognition syllable character string by the word found by the word search means,
The word search means
First search pattern creation means for creating a search pattern including a regular expression for each word stored in the word pool, and syllable characters that match the search pattern created by the first search pattern creation means Including means for searching for the sequence part from the speech recognition syllable character string.
False recognition correction device. Kana-kanji conversion means for converting the speech recognition syllable character string into kana-kanji characters, and means for determining whether or not the words stored in the word pool are included in the kana-kanji converted character string,
When it is determined that the word stored in the word pool is included in the character string converted into Kana-Kanji, the speech recognition syllable character string portion excluding the word is the word search means and the word Processed by replacement means,
The erroneous recognition correction apparatus according to claim 1. The above search pattern is
A search pattern that can search for a syllable character string that lacks any one of the syllable character strings of a word.
The erroneous recognition correction apparatus according to claim 1 or 2. The above search pattern is
A search pattern that can search for a syllable string in which any one of the syllable strings of a word is replaced with another character.
The misrecognition correction apparatus according to any one of claims 1 to 3. The above search pattern is
A search pattern that can search for a syllable string in which another character is inserted between characters included in the syllable string of a word.
The misrecognition correction apparatus according to any one of claims 1 to 4. A co-occurrence word pool in which a co-occurrence word and a syllable character string of each of a plurality of words stored in the word pool are stored in association with each other;
Keyword selection means for selecting a word found by the word search means as a target keyword;
A co-occurrence word specifying means for specifying a co-occurrence word for the target keyword from co-occurrence words stored in the co-occurrence word pool;
Co-occurrence searching for a co-occurrence word having a syllable character string similar to the speech recognition syllable character string part obtained by removing the target keyword from the speech recognition syllable character string from the co-occurrence words specified by the co-occurrence word specifying means Word search means, and co-occurrence word replacement means for replacing the corresponding syllable character string portion included in the speech recognition syllable character string by the co-occurrence word found by the co-occurrence word search means,
The erroneous recognition automatic correction device according to any one of claims 1 to 5. The co-occurrence word search means is
For each of the co-occurrence words stored in the co-occurrence word pool, a second search pattern creation means for creating a search pattern including a regular expression, and a search pattern created by the second search pattern creation means Including means for searching for a matching syllable string part from the speech recognition syllable string part excluding the target keyword,
The misrecognition correcting apparatus according to claim 6. A plurality of field word pools each storing only words frequently used in a specific field,
The word search processing by the word search means and the replacement processing of the syllable character string portion by the word replacement means are executed using each of a plurality of field-specific word pools.
The misrecognition correction apparatus according to any one of claims 1 to 7. A plurality of field co-occurrence word pools corresponding to each of the plurality of field word pools are provided.
The co-occurrence word specifying process by the co-occurrence word specifying means, the co-occurrence word searching process by the co-occurrence word searching means, and the replacement process of the syllable character string part by the co-occurrence word replacing means are a plurality of field-specific co-occurrence word pool Are executed using each of the
The misrecognition correcting device according to claim 8. A program for causing a computer to function as a misrecognition correcting device,
Recognizes the given speech and converts it to a syllable string,
A syllable character string that matches a syllable character string part included in the speech recognition syllable character string among words stored in a word pool in which a plurality of words and their syllable character strings are stored in association with each other, or Search for words with similar syllable strings,
The found word replaces the corresponding syllable string part included in the speech recognition syllable string,
Here, for the word search, for each word stored in the word pool, a search pattern including a regular expression is created, and the syllable character string portion matching the created search pattern is defined as the speech recognition syllable character string. Searching from within,
program.

Images