JP2023146216A - Conversion-into-text support device and conversion-into-text support method - Google Patents

Abstract

To enable keyword matching of a call content to be carried out with high accuracy regardless of characteristics of speech-to-text conversion concerning the call content.SOLUTION: A conversion-into-text support device 100 includes: a storage device 101 that holds master data 126 that defines information on correct phonemes of each vocabulary; and an arithmetic device 104 that applies call recording data to an acoustic model 110 to extract phonemes, calculates a coincidence rate between correct phonemes of a vocabulary assumed to appear concerning the call recording data and the phonemes obtained by the extraction in the vocabulary for which the phonemes have been defined by the master data 126, and as a result of this calculation, specifies a vocabulary in which the phonemes indicate a predetermined coincidence rate as a result of keyword matching.SELECTED DRAWING: Figure 2

Description

The present invention relates to a text conversion support device and a text conversion support method.

There is a need to confirm whether the content of calls made by salespeople, call centers, etc. is appropriate from a compliance perspective. In addition, in recent years, instead of the traditional method of re-listening to and checking the recorded data of such a call, a method has been proposed in which the voice data is converted into text and then checked.
Conventional technology related to such text conversion is based on data such as the content of explanations to customers during business negotiations and sales activities, and determines whether "prohibited expressions" are included and whether "required matters" are included. Compliance check systems and compliance check programs (see Patent Document 1), etc., have been proposed to check both.

This technology is a compliance check system that checks whether compliance is observed with respect to each utterance made by a sales representative to a customer, and the content of each utterance made by the sales representative is converted into text using voice recognition technology. a text analysis unit that performs natural language analysis processing including morphological analysis on the analyzed text data and outputs it as analyzed text data; Judgment for grouping into blocks consisting of one or more consecutive utterances and determining whether or not the contents of first text data predefined as necessary matters to be explained to the customer are explained in each block. and the analyzed text data related to each of the utterances, if there is one that matches second text data that is predefined as content of prohibited expressions that must not be said to the customer, the target utterance is a keyword matching unit that determines that the target prohibited expression is said in the above, and management that identifies the salesperson and/or the customer in data of each of the utterances made by the salesperson to the customer. a data recording unit that records the information in association with the information; When the determination unit determines that the content of the first text data is explained in the block, including past utterances made by the salesperson to the customer, the determination unit In this system, the necessary matters are assigned categories and recorded, and the degree to which each of the necessary matters has been explained is determined based on predetermined evaluation criteria set in advance.

Japanese Patent Application Publication No. 2018-120640

With regard to the above-mentioned text conversion, advances in deep learning technology and the like have led to improvements in accuracy, and its utilization is progressing. For example, one example of the use of call recording data in the financial field is to check whether NG words are said or not, and whether correct customer names and product names are pronounced.
When performing this check, keyword matching is often performed on the text of call recording data. However, it has been found that there are calls for which the accuracy of text conversion is likely to be low (calls with many false positives) due to factors such as recording conditions and the habits of the speaker, and it is difficult to perform accurate keyword matching for these calls. It was difficult.

In other words, even if keyword matching is performed for phone calls, where the accuracy of voice-to-text conversion tends to be low, the accuracy cannot be expected, and this is a factor that ends up being overlooked.

SUMMARY OF THE INVENTION Therefore, an object of the present invention is to provide a technique that enables keyword matching of the contents of a call with high accuracy, regardless of the characteristics of the speech-to-text conversion of the contents of the call.

The text conversion support device of the present invention that solves the above problems includes a storage device that holds master data specifying correct phoneme information for each vocabulary that is expected to appear in each conversation scene or subject, and A process of applying phone call recording data to an acoustic model to extract phonemes, and a process of extracting phonemes from among the vocabulary whose phonemes are defined in the master data, which are expected to appear in the conversation scene or subject of the call recording data. a calculation device that executes a process of calculating a match rate between a correct phoneme and the extracted phoneme, and a process of specifying the vocabulary whose phonemes show a predetermined match rate as a keyword matching result as a result of the calculation; It is characterized by containing.
Further, in the text conversion support method of the present invention, the information processing device stores master data in a storage device that specifies correct phoneme information for each vocabulary that is expected to appear in each conversation scene or subject, and A process of extracting phonemes by applying the phone call recording data obtained from the phone call recording data to an acoustic model, and a process of extracting phonemes by applying the phoneme recording data obtained from the phone call recording data to an acoustic model, and a process in which phonemes are expected to appear in the conversation scene or subject of the phone call recording data from among the vocabulary for which phonemes are defined in the master data. A process of calculating a match rate between the correct phoneme of the vocabulary and the extracted phoneme, and a process of identifying the vocabulary whose phonemes show a predetermined match rate as a keyword matching result as a result of the calculation. It is characterized by

According to the present invention, regardless of the characteristics of voice-to-text conversion regarding the contents of a call, keyword matching of the contents of the call can be performed with good accuracy.

FIG. 1 is a network configuration diagram including a text conversion support device according to the present embodiment. 1 is a diagram illustrating an example of the hardware configuration of a text conversion support device in this embodiment. FIG. It is a diagram showing an example of the hardware configuration of an operator terminal in this embodiment. It is a diagram showing an example of the hardware configuration of a call center system in this embodiment. It is a diagram showing an example of the hardware configuration of an administrator terminal in this embodiment. It is a figure showing an example of composition of call recording DB of this embodiment. It is a figure showing an example of composition of a phoneme master table of this embodiment. It is a figure showing the example of composition of the utterance similarity table of this embodiment. It is a figure showing example 1 of a flow of a textization support method in this embodiment. It is a figure which shows the flow example 2 of the text conversion support method in this embodiment. It is a figure which shows the flow example 3 of the text conversion support method in this embodiment.

<Network configuration>
Embodiments of the present invention will be described in detail below with reference to the drawings. FIG. 1 is a network configuration diagram including a text conversion support apparatus 100 of this embodiment. The text conversion support device 100 shown in FIG. 1 is a computer that is capable of performing keyword matching of the contents of a call with high accuracy, regardless of the characteristics of converting the contents of the call into voice to text.

As shown in FIG. 1, the text conversion support device 100 of this embodiment communicates with an operator terminal 200, a call center system 300, and an administrator terminal 400 via an appropriate network 1 such as the Internet or a secure line within an organization. , and are communicably connected as necessary. Therefore, these may be collectively referred to as the text conversion system 10.

The text conversion support device 100 of this embodiment can, for example, check whether the content of the conversation between the operator and the customer at the call center was appropriate from the viewpoint of compliance and customer service, the appearance of NG words during the conversation, or the absence of essential words. It can be said that it is a support device that identifies phenomena such as appearance by keyword matching.

Of course, the present invention is not only applicable to situations in which keyword matching is performed regarding conversations between operators and customers in call center operations, but also there is an opportunity to verify the appearance of necessary/contraindicated keywords in voice data. It can be applied to any business, etc.

On the other hand, the operator terminal 200 is a terminal used by a person in charge of responding to inquiries from customers regarding various products and services, or conducting telephone sales to potential customers. Specifically, a telephone terminal integrated with a PC, a smartphone, a tablet terminal, a personal computer, etc. can be assumed. The conversation between the person in charge and the customer at the operator terminal 200 is recorded, and is managed and utilized as recorded call data.

Further, the call center system 300 is a system that manages incoming and outgoing calls between the above-described operator terminal 200 and customer telephones, and manages call recording data that is the content of conversations at the above-described operator terminal 200. Therefore, call center system 300 retains and manages call recording data in a storage device, and distributes it to text conversion support device 100 as appropriate.

Further, the administrator terminal 400 is a terminal operated by the administrator of the above-mentioned call center. This administrator terminal 400 instructs keyword matching processing to check call recording data for a predetermined period of time, such as one day, from a predetermined viewpoint such as the above-mentioned compliance at an appropriate timing such as at the end of work at the call center. This is a terminal that performs the following on the text conversion support device 100 and obtains the processing results.
<Hardware configuration>
Further, the hardware configuration of the text conversion support apparatus 100 of this embodiment is shown in FIG. 2 as follows.

That is, the text conversion support device 100 includes a storage device 101, a memory 103, an arithmetic device 104, and a communication device 105.

Among these, the storage device 101 is configured with an appropriate nonvolatile storage element such as an SSD (Solid State Drive) or a hard disk drive.

Furthermore, the memory 103 is composed of a volatile storage element such as a RAM.

Further, the arithmetic device 104 is a CPU that reads the program 102 held in the storage device 101 into the memory 103 and executes it, performs overall control of the device itself, and performs various judgments, calculations, and control processing.

Further, the communication device 105 is assumed to be a network interface card or the like that connects to the network 1 and handles communication processing with at least the call center system 300.

Note that when the text conversion support device 100 is a stand-alone machine, it is preferable to further include an input device that accepts key inputs and voice inputs from the user, and an output device such as a display that displays processed data.

In addition to the program 102 for implementing the functions necessary for the text conversion support device of this embodiment, the storage device 101 stores at least a call recording DB 125, a phoneme master table 126, and an utterance similarity table 1276. has been done. However, details regarding these databases and the like will be described later.

It is also assumed that the program 102 includes an acoustic model 110 and a language model 111. The acoustic model 110 has a function of extracting phonemes constituting the voice of a conversation from call recording data regarding a conversation between an operator and a customer.

Therefore, the text conversion support device 100 analyzes the voice features (frequency and sound strength) shown by the call recording data, performs acoustic analysis in advance to convert it into data that is easy to handle, and the acoustic analysis results indicate The feature amounts will be given to the acoustic model 110.

The acoustic model 110 is a model that defines the correspondence between the above-mentioned features and phonemes through appropriate deep learning, and extracts phonemes, which are the smallest units of sound waves, by giving the above-mentioned voice features. do.

Note that a phoneme is the smallest component of a sound wave that can be observed when a voice is uttered. Phonemes in Japanese are made up of three types: vowels (aiueo), onomatopoeias (n), and consonants (23 types). For example, in the case of "Tanaka-san," the phoneme is "t-a-n-a-k-a-s-a-n."

The text conversion support device 100 of this embodiment performs keyword matching based on the phonemes obtained by the acoustic model 110. In the above case, the phoneme "tanakasan"
This corresponds to the process of specifying ``Tanaka-san'' as a Japanese vocabulary word. More specifically, phonemes are replaced with vocabulary by applying the text conversion support method of the present invention while appropriately using the phoneme master table 126 to determine which word each phoneme corresponds to.

On the other hand, the language model 111 is responsible for appropriately converting a group of vocabulary obtained through keyword matching into sentences. For example, for a cluster of vocabulary such as "Mr. Tanaka," "in Shinshu," "snow," and "It piled up," statistical data about the relationship between the cluster of vocabulary and a sentence that is correct (or appears frequently), etc. Construct meaningful sentences based on the most likely combinations.

Further, the hardware configuration of the operator terminal 200 of this embodiment is shown in FIG. 3 as follows.

That is, the operator terminal 200 includes a storage device 201, a memory 203, a calculation device 204, an input device 205, an output device 206, and a communication device 207.

Among these, the storage device 201 is configured with an appropriate nonvolatile storage element such as an SSD (Solid State Drive) or a hard disk drive.

Furthermore, the memory 203 is composed of a volatile storage element such as a RAM.

Further, the arithmetic device 204 is a CPU that reads out and executes the program 202 held in the storage device 201 into the memory 203, performs overall control of the device itself, and performs various judgments, calculations, and control processing.

Further, the input device 205 is configured of a device such as a keyboard and a mouse that accept key input and voice input from an operator who is a user.

Further, the output device 206 is configured with a device such as a display or a speaker that displays the processing results of the arithmetic device 204.

Furthermore, the communication device 207 is assumed to be a network interface card or the like that connects to the network 1 and handles communication processing with the call center system 300 and the administrator terminal 400 (or the text conversion support device 100).

Further, the hardware configuration of the call center system 300 of this embodiment is shown in FIG. 4 as follows.

That is, the call center system 300 includes a storage device 301, a memory 303, a calculation device 304, and a communication device 305.

Among these, the storage device 301 is composed of an appropriate nonvolatile storage element such as an SSD (Solid State Drive) or a hard disk drive.

Furthermore, the memory 303 is composed of a volatile storage element such as a RAM.

Further, the arithmetic device 304 is a CPU that reads a program 302 held in the storage device 301 to the memory 303 and executes it, performs overall control of the device itself, and performs various judgments, calculations, and control processing.

Furthermore, the communication device 305 is assumed to be a network interface card or the like that connects to the network 1 and handles communication processing with at least the text conversion support device 100 and the operator terminal 200.

Note that when the call center system 300 is a stand-alone machine, it is preferable to further include an input device that accepts key input and voice input from the user, and an output device such as a display that displays processed data.

Furthermore, in the storage device 301, in addition to a program 302 for implementing functions necessary for the call center system 300 of this embodiment, at least call recording data 325 is stored. This call recording data 325 is data that becomes a record of the call recording DB 125 in the text conversion support device 100.

Further, the hardware configuration of the administrator terminal 400 of this embodiment is shown in FIG. 5 as follows.

That is, the administrator terminal 400 includes a storage device 401, a memory 403, a calculation device 404, an input device 405, an output device 406, and a communication device 407.

Among these, the storage device 401 is configured with an appropriate nonvolatile storage element such as an SSD (Solid State Drive) or a hard disk drive.

Further, the memory 403 is composed of a volatile storage element such as a RAM.

Further, the arithmetic device 404 is a CPU that reads out and executes the program 402 held in the storage device 401 into the memory 403, performs overall control of the device itself, and performs various judgments, calculations, and control processing.

Further, the input device 405 includes a device such as a keyboard and a mouse that accept key input and voice input from an operator who is a user.

Further, the output device 406 is configured with a device such as a display or a speaker that displays the processing results of the arithmetic device 404.

Furthermore, the communication device 407 is assumed to be a network interface card or the like that connects to the network 1 and handles communication processing with the text conversion support device 100 and the call center system 300.
<Data structure example>
Next, various types of information used by the text conversion support device 100 of this embodiment will be explained. FIG. 6 shows an example of the call recording DB 125 in this embodiment. The call recording DB 125 of this embodiment is, for example, a database that stores call recording data between operators and customers acquired from the call center system 300 (or from the operator terminal 200).

This call recording DB 125 stores, for example, data such as the name of the customer, the product/service name specified by the customer, the ID of the corresponding operator, and the recorded data file, using the date and time of the call and the customer ID indicating the customer to be called as keys. It is a collection of linked records.

Further, FIG. 7 shows an example of the configuration of the phoneme master table 126 in this embodiment. The phoneme master table 126 of this embodiment is a table that defines correct phonemes for each vocabulary.

This phoneme master table 126 has a configuration in which, for example, conversation scenes and objects are used as keys, and information on correct phonemes of vocabulary that is expected to appear during conversations regarding these scenes and objects is defined.

Further, FIG. 8 shows a configuration example of the utterance similarity table 127 in this embodiment. The utterance similarity table 127 of this embodiment is a table that defines the similarity between utterances of Japanese vowels.

This utterance similarity table 127 lists vowels vertically and horizontally, and the similarity between each vowel is expressed as a discontinuous numerical value between a maximum value of 1 (perfect match) and a minimum value of 0 (zero similarity). It constitutes a specified matrix.
<Flow example 1>
Hereinafter, the actual procedure of the text conversion support method in this embodiment will be explained based on the drawings. Various operations corresponding to the text conversion support method described below are realized by a program that the text conversion support apparatus 100 reads into a memory or the like and executes. This program is composed of codes for performing various operations described below.

FIG. 9 is a diagram showing a flow example 1 of the text conversion support method in this embodiment. In this case, the text conversion support device 100 obtains call recording data 325 from, for example, the call center system 300 (or operator terminal 200), and stores it in the call recording DB 125 (s1).

In addition, the text conversion support device 100 detects the arrival of a predetermined timing or receives an instruction from the administrator terminal 400, and converts, for example, data related to a predetermined period out of the call recording data held in the call recording DB 125. is extracted and applied to the acoustic model 110 to extract phonemes (s2).

For example, regarding call recording data in which a call center operator uttered ``Mr. Saeki'' to a customer named ``Saeki'' after a standard greeting, a call center operator said ``T-A-I-K-I-S-A''. -N'' is extracted. Here, the customer name is the processing target, but this is just one example, and it is also suitable to use, for example, the financial product name as the processing target.

Next, the text conversion support device 100 identifies that the customer to whom the call is directed is "Mr. Saeki" from the customer ID linked to the above-mentioned call recording data, and uses this "Mr. Saeki" in the vocabulary for keyword matching. , the phoneme is extracted from the phoneme master table 126 (s3). In this case, the phoneme array “S-A-E-K-I-S-AN” is extracted from the record of customer ID “C018122: Saeki***” in the phoneme master table 126.

Next, the text conversion support device 100 compares the phoneme sequences obtained in s2 and s3, and calculates the matching rate (s4). In the above case, when comparing the phoneme sequence "TA-I-K-I-S-AN" with the phoneme sequence "S-A-E-K-I-S-AN", Of the total 8 phonemes, 6 phonemes match, and the match rate is 6/8=0.75.

If we applied the phoneme sequence "TA-I-K-I-S-AN" obtained from the call recording data to the language model 111 as usual, and obtained the text "Daiki-san", then this When this is compared with the text "Saeki-san" in the specified vocabulary in the phoneme master table 126, the match rate is calculated as 2/4 = 0.5, with 2 out of 4 characters matching. It turns out. If the pass/fail criterion for keyword matching is, for example, a match rate of 0.6, even though the operator has indeed uttered the customer's name as "Mr. Saeki," due to the influence of the conversion accuracy in the language model 111, These will be judged as not matching.

On the other hand, according to the text conversion support device 100 of the present invention, it is possible to solve the problem of conversion accuracy in the language model 111 and perform keyword matching based on the matching rate between phoneme sequences, which is more accurate than before. Good keyword matching is possible.
<Flow example 2>
FIG. 10 is a diagram showing a flow example 2 of the text conversion support method in this embodiment. Here, in order to further enhance the effect of the above-described flow example 1, a method will be described in which the degree of coincidence of phoneme sequences is calculated by adding the viewpoint of vowels. Note that in this flow, steps s1 and s2 are the same as in the flow example 1 described above, so the subsequent processing will be described.

The text conversion support device 100 extracts only vowels (a, i, u, e, o) from the phoneme array extracted as in the flow example 1 described above (s10). In the above example, the vowel array "A, I, I, A" will be extracted.

Further, the text conversion support device 100 extracts phonemes and vowels regarding the above-mentioned customer "Saeki" to be called, in the same manner as in s3 and s10 (s11). In this case, the vowel array "A, E, I, A" is extracted from the phoneme array "SAE-K-I-S-AN".

Next, the text conversion support device 100 compares the vowels in the vowel arrays obtained in s10 and s11 with the utterance similarity table 127 in order from the beginning of the array, and calculates the similarity of vowels between corresponding positions in the vowel arrays. Specify (s12).

For example, the vowel "A" and the vowel "A" have a similarity of "1" according to the utterance similarity table 127, and the vowel "A" and the vowel "I" have a similarity of "0" according to the utterance similarity table 127. ", the vowel "A" and the vowel "U" have a similarity of "0" according to the utterance similarity table 127, and the vowel "A" and the vowel "E" have a similarity of "0" according to the utterance similarity table 127. According to the utterance similarity table 127, the vowel "A" and the vowel "O" have a similarity of "0.5".

As a result, in the above example, "A, I, I, A" and "A, E, I, A" are compared, and the similarity is "1" and "I" for "A" and "A". and "E" have a similarity of "0.5," "I" and "I" have a similarity of "1," and "A" and "A" have a similarity of "1."

Therefore, the text conversion support device 100 calculates the vowel similarity in the above-mentioned phoneme array as (1+0.5+1+1)/4=0.875 based on the similarity for each vowel obtained in s12 (s13).

The text conversion support device 100 also calculates the matching rate for consonants based on the phoneme arrays obtained in s2 and s3 (s14). In the above example, the consonants "T, K, S, N" in the phoneme sequence "TA-I-K-I-S-A-N" and "S-A-E-K-I- When comparing the consonants "S, K, S, N" in the phoneme array "S-AN", three phonemes out of the total four phonemes match, and the match rate is 3/4 = 0.75. Become.

Subsequently, the text conversion support device 100 weights the vowel similarity obtained in s13, performs a weighted average with the consonant matching rate, and calculates the matching rate between phoneme arrays (s15).

For example, if the above-mentioned weighting is set to "2", that is, twice the weight of the consonant matching rate, and weighted averaging is performed, (consonant matching rate 0.75 + vowel similarity 0.875 x weight 2)/3 The matching rate can be calculated as =0.83.
<Flow example 3>
FIG. 11 is a diagram showing a flow example 3 of the text conversion support method in this embodiment. Here, in order to further enhance the effects of the above-described flow examples 1 and 2, a method of calculating the degree of coincidence of phoneme sequences will be described with an added perspective of dealing with omissions and misspellings. Note that in this flow, steps s1 and s2 in the above-described flow example 1 and up to s10 and s11 in the flow example are the same, so the subsequent processing will be described.

The text conversion support device 100 converts the vowel array in the phoneme array in the phoneme array extracted as described above and the vowel array in the phoneme array read from the corresponding record of the phoneme master table 126 into the vowel array. The degree of similarity between a set of two consecutive vowels is specified based on the utterance similarity table 127 (s20).

For example, the phoneme sequence “O-H-A-Y-O-U-G-O-Z-A-I-M-A-S-U-S-A-K-I-S” obtained from call recording data In the vowel arrangement "O, A, O, U, O, A, I, A, U, A, I, A" in "-AN", select two vowels from the beginning and set them to group (1). "O, A", set (2) "O, U", set (3) "O, A", set (4) "I, A", set (5) "U, A", set (6) When a total of six sets such as "I, A" are formed, based on the utterance similarity table 127, set (1) has a similarity of "0.5", set (2) has a similarity of "0.5", and set (2) has a similarity of "0.5". It can be specified that (3) has a similarity of "0.5", set (4) has a similarity of "0", set (5) has a similarity of "0", and set (6) has a similarity of "0".

Next, the text conversion support device 100 convolves the combinations identified in s20 whose similarity exceeds a standard, such as 0.5, into one predetermined vowel (e.g. A, I, U), For pairs whose degree of similarity is lower than the standard, the first vowel is adopted, and the process of forming the next pair with adjacent vowels in the phoneme array is performed for the trailing vowel to generate a syllable array (s21).

In the above example, set (1) is aggregated to the vowel "A" (that is, convolution; the same applies hereafter), set (2) is aggregated to the vowel "U", set (3) is aggregated to the vowel "A", and the set is aggregated to the vowel "A". (4) adopts the first vowel "I" and combines the last vowel "A" with the first vowel "U" of the original set (5) to form a new set (5)'; The sets are also reconstructed for the vowel arrangement, and the aggregation based on the above-mentioned similarity is performed.

As a result, the syllable arrangement remaining after aggregating each set is "A, U, A, I, A, U, A, I, A."

The text conversion support device 100 generates such a syllable arrangement such as the vowel arrangement "A" in the phoneme arrangement "S-A-S-A-K-I-S-AN" of the corresponding record in the phoneme master table 126. The same process is performed for "A, I, A" to obtain "A, I, A".

Next, in s21, the text conversion support device 100 selects the syllable arrangement derived from the phoneme master table 126 and the number of vowels for the portion that matches the syllable arrangement generated from the phoneme master table 126 in the syllable arrangement derived from the recorded call data. If the number of vowels is the same (s22: same number), the match rate of vowels in the corresponding vowel arrangement is calculated between the above-mentioned part and the syllable arrangement derived from the phoneme master table 126 based on the utterance similarity table 127. (s23).

For example, the phoneme arrangement of recorded phone calls “O-H-A-Y-O-U-G-O-Z-A-I-M-A-S-U-S-A-K-I-S-A -N” in the vowel array “O, A, O, U, O, A, I, A, U, A, I, A”, the phoneme array is the syllable array “A, "I, A" (which is based on the vowel arrangement "A, A, I, A") matches "O, A, I, A".

Therefore, the text conversion support device 100 converts each of the vowel sequences "O, A, I, A" derived from the call recording data and the vowel array "A, A, I, A" derived from the phoneme master table 126. The degree of similarity between vowels is specified based on the utterance similarity table 127, and calculated as (0.5+1+1+1)/4=0.875, for example.

On the other hand, as a result of the comparison of the number of vowels in s22 described above, if the number of vowels derived from the master table 126 is greater than the number of vowels derived from the call recording data (s22: many), the text conversion support device 100 determines that an omission has occurred. It is estimated that the syllable arrangement derived from the master table 126 is correct, and the portion where a vowel is missing in the syllable arrangement derived from the call recording data is corrected by supplementing it with the corresponding phoneme derived from the master table 126 (s24); The concordance rate of vowels between the corrected vowel array and the vowel array derived from the master table 126 is calculated based on the utterance similarity table 127 (s25).

For example, the vowel array "A, I, A" in the phoneme array "S-A-K-I-S-A" of the telephone recording data is the syllable array "A, I, A" derived from the phoneme master table 126. , A" (which is based on the vowel sequence "A, A, I, A"). However, the number of vowels in the corresponding vowel array is one more in the vowel array derived from the master table 126.

Therefore, the text conversion support device 100 uses the vowel array "A, I, A" derived from the recorded call data.
Among them, "A" is the second vowel from the beginning that is missing compared to the vowel array "A, A, I, A" derived from the master table 126 described above. Therefore, the text conversion support device 100 corrects the vowel array "A, I, A" derived from the recorded call data by adding "A" between the first "A" and the second "I".

Furthermore, the text conversion support device 100 calculates the degree of similarity between the vowel arrangement subjected to the above correction and the vowel arrangement derived from the master table 126, such as (1+1+1+1)/4=1, based on the utterance similarity table 127. It will be calculated as follows.

On the other hand, as a result of the comparison of the number of vowels in s22 described above, if the number of vowels derived from the master table 126 is smaller than the number of vowels derived from the call recording data (s22: small), the text conversion support device 100 determines that the number of vowels is It is assumed that the syllable arrangement derived from the master table 126 is correct, and the part where there are excessive vowels in the syllable arrangement derived from the call recording data is corrected by deleting it (s26), and the vowel after this correction is corrected. The vowel matching rate between the array and the vowel array derived from the master table 126 is calculated based on the utterance similarity table 127 (s27).

For example, the vowel array "A, A, A, I, A" in the phoneme array "A-K-A-S-A-K-I-S-A" in the phone call recording data is the phoneme master. It matches the syllable arrangement "A, I, A" from table 126 (which is based on the vowel arrangement "A, A, I, A"). However, the number of vowels in the corresponding vowel array is one less in the vowel array derived from the master table 126.

Therefore, the text conversion support device 100 compares the vowel array "A, A, A, I, A" derived from the recorded call data with the vowel array "A, A, I, A" derived from the master table 126 described above. The one that is excessive is the first "A". Therefore, the text conversion support device 100 corrects the vowel array "A, A, A, I, A" derived from the call recording data by deleting the leading "A".

Furthermore, the text conversion support device 100 calculates the degree of similarity between the vowel arrangement subjected to the above correction and the vowel arrangement derived from the master table 126, such as (1+1+1+1)/4=1, based on the utterance similarity table 127. It will be calculated as follows.

Note that, since it has already been explained in flow example 2, a description of the concept of calculating the matching rate by taking into consideration the matching degree of the consonant arrangement as well as the similarity of the vowel arrangement will be omitted.

Although the best mode for carrying out the present invention has been specifically described above, the present invention is not limited thereto and can be modified in various ways without departing from the gist thereof.

According to this embodiment, regardless of the characteristics of voice-to-text conversion regarding the call content, keyword matching of the call content can be performed with good accuracy.

The description of this specification clarifies at least the following. That is, in the text conversion support device of the present embodiment, the storage device further stores information specifying the utterance similarity between vowels, and the calculation device stores the correct phoneme and the extracted phoneme when calculating the matching rate. A process of calculating a match rate between vowels included in each of the correct phonemes based on the utterance similarity information, and a process of calculating a match rate between consonants included in each of the correct phoneme and the extracted phoneme. The matching rate between the vowels is weighted more favorably than the matching rate between the consonants, and the matching rate between the phonemes is calculated based on each matching rate between the vowels and between the consonants. good.

According to this, when matching the phonemes mentioned above, priority is given to vowels (which have a small number of types and are easy to distinguish, i.e., have characteristics that make it difficult to falsely detect) as elements to be matched, and this increases the accuracy of the matching rate. It becomes easier to understand. As a result, regardless of the characteristics of voice-to-text conversion regarding the content of the call, keyword matching of the content of the call can be performed with higher accuracy.

Furthermore, in the text conversion support device of the present embodiment, when calculating the matching rate, the arithmetic device may calculate two consecutive vowels in the arrangement of vowels included in the extracted phoneme and the correct phoneme, respectively, for each of the extracted phoneme and the correct phoneme. The degree of similarity in a set of vowels is specified by the utterance similarity, and for a set in which the degree of similarity is above a standard, it is convolved into one predetermined vowel, and for a set in which the degree of similarity is below the standard, the first vowel is adopted. Then, a syllable array is generated by forming the next set of the trailing vowel with an adjacent vowel in the array, and the syllable array generated for each of the extracted phoneme and the correct phoneme is If the number of vowels is equal, the vowel matching rate between the vowel arrays of the extracted phoneme and the correct phoneme, which are the source of the syllable array, is calculated as the utterance similarity. It may be calculated based on degree.

According to this, in Japanese, when there are consecutive vowels with high vowel similarity, there are cases in which two letters are pronounced as one syllable, cases in which only one letter is pronounced, cases in which the first letter is not pronounced, and cases in which the same letter is unnecessarily pronounced. It is now possible to appropriately deal with problems such as cases of overlapping pronunciations, where phenomena such as omissions and spellings are likely to occur, and furthermore, regardless of the characteristics of the voice-to-text conversion of the content of the call, keyword matching of the content of the call can be performed. It becomes possible to perform the process with better accuracy.

Further, in the text conversion support device of the present embodiment, the calculation device compares the number of vowels between the syllable arrays, and if the number of vowels in the syllable array corresponding to the correct phoneme is greater than the extracted phoneme, the calculation device compares the number of vowels between the syllable arrays. , when the syllable arrangement of the correct phoneme is correct, the portion where a vowel is missing in the syllable arrangement of the extracted phoneme is corrected by supplementing it with the corresponding phoneme of the correct phoneme, and the corrected syllable arrangement of the extracted phoneme is corrected. The vowel matching rate between the vowel arrangement of each of the phoneme and the correct phoneme may be calculated based on the utterance similarity.

According to this, it is possible to appropriately deal with the above-mentioned omitted characters, and to perform keyword matching for the content of the call with better accuracy, regardless of the characteristics of the speech-to-text conversion of the content of the call.

Further, in the text conversion support device of the present embodiment, the calculation device compares the number of vowels between the syllable arrays, and if the number of vowels in the syllable array corresponding to the correct phoneme is smaller than the extracted phoneme, the calculation device compares the number of vowels between the syllable arrays. , if the syllable arrangement of the correct phoneme is correct, correct the syllable arrangement of the extracted phoneme by deleting the portion where the vowel is redundant, and correct the extracted phoneme and the correct phoneme after the correction. The vowel matching rate between each of the vowel arrangements may be calculated based on the utterance similarity.

According to this, it is possible to appropriately deal with the above-mentioned spelling event, and to perform keyword matching for the contents of the call with better accuracy, regardless of the characteristics of the speech-to-text conversion of the contents of the call.

Further, in the text conversion support method of the present embodiment, the information processing device further stores information specifying utterance similarity between vowels in the storage device, and when calculating the matching rate, the information processing device further stores information specifying the utterance similarity between vowels, A process of calculating a match rate between vowels included in each of the extracted phonemes based on the utterance similarity information, and a match rate between consonants included in each of the correct phoneme and the extracted phoneme. In the process, the match rate between the vowels may be weighted more favorably than the match rate between the consonants, and the match rate between the phonemes may be calculated based on each match rate between the vowels and between the consonants.

Further, in the text conversion support method of the present embodiment, when calculating the matching rate, the information processing device may, for each of the extracted phoneme and the correct phoneme, select two consecutive vowels included in the phoneme. The degree of similarity between two pairs of vowels is specified by the utterance similarity, and the pairs whose similarity is above the standard are convolved into one predetermined vowel, and the pairs whose similarity is below the standard are convoluted with the first vowel. the syllable array generated for each of the extracted phoneme and the correct phoneme; If the numbers of vowels are equal, the vowel match rate between the vowel arrays of each of the extracted phonemes and the correct phonemes, which are the source of the syllable array, is calculated from the utterance. It may be calculated based on the degree of similarity.

Further, in the text conversion support method of the present embodiment, the information processing device compares the number of vowels between the syllable arrays, and the number of vowels in the syllable array corresponding to the correct phoneme is greater than the extracted phoneme. In this case, if the syllable arrangement of the correct phoneme is correct, the portion where a vowel is missing in the syllable arrangement of the extracted phoneme is corrected by supplementing it with a corresponding phoneme of the correct phoneme, and the extraction is performed after the correction. The vowel matching rate may be calculated between the vowel arrangement of each of the correct phoneme and the correct phoneme based on the utterance similarity.

Further, in the text conversion support method of the present embodiment, the information processing device compares the number of vowels between the syllable arrays, and the number of vowels in the syllable array corresponding to the correct phoneme is smaller than the extracted phoneme. In this case, if the syllable arrangement of the correct phoneme is correct, the syllable arrangement of the extracted phoneme is corrected by deleting the part where the vowel is redundant, and the extracted phoneme with the correction and the correct syllable arrangement are corrected. A vowel matching rate may be calculated between the vowel arrays of each phoneme based on the utterance similarity.

1 Network 100 Text conversion support device 101 Storage device 102 Program 103 Memory 104 Arithmetic device 105 Communication device 110 Acoustic model 111 Language model 125 Call recording DB
126 Phoneme master table 127 Utterance similarity table 200 Operator terminal 300 Call center system 400 Administrator terminal

Claims (10)

a storage device that holds master data that defines correct phoneme information for each vocabulary that is expected to appear in each conversation situation or subject;
A process of applying phone call recording data obtained from a predetermined device to an acoustic model to extract phonemes, and a process of extracting phonemes from the vocabulary in which phonemes are defined in the master data, which are expected to occur in the conversation scene or subject of the phone call recording data. A process of calculating a match rate between the correct phoneme of the vocabulary and the extracted phoneme, and a process of identifying the vocabulary whose phonemes show a predetermined match rate as a keyword matching result as a result of the calculation. a computing device;
A text conversion support device including The storage device is
Further retains information specifying the utterance similarity between vowels,
The arithmetic device is
When calculating the matching rate, a process of calculating a matching rate between vowels included in each of the correct phoneme and the extracted phoneme based on the utterance similarity information, and a process of calculating the matching rate between the vowels included in each of the correct phoneme and the extracted phoneme, and The process of calculating the matching rate between the consonants included in each, and weighting the matching rate between the vowels more favorably than the matching rate between the consonants, and calculating the phoneme based on the matching rate between the vowels and between the consonants. It calculates the concordance rate between
The text conversion support device according to claim 1, characterized in that: The arithmetic device is
When calculating the above match rate,
For each of the extracted phoneme and the correct phoneme, in the arrangement of vowels included in the phoneme, the degree of similarity in a set of two consecutive vowels is specified by the utterance similarity, and for the set for which the degree of similarity is equal to or higher than a reference value. is convolved into one predetermined vowel, and for pairs whose similarity is below the standard, the first vowel is adopted, and the process of forming the next pair with the vowel adjacent in the array is performed with the last vowel. to generate a syllable array,
The number of vowels is compared between the syllable arrays generated for each of the extracted phoneme and the correct phoneme, and if the numbers of vowels are equal, the extracted phoneme and the correct phoneme are the basis of the syllable array. A vowel matching rate is calculated between each of the vowel arrangements based on the utterance similarity,
3. The text conversion support device according to claim 2. The arithmetic device is
Comparing the number of vowels between the syllable arrays, if the number of vowels in the syllable array corresponding to the correct phoneme is greater than the extracted phoneme, and if the syllable array of the correct phoneme is correct, then the extracted phoneme The part where a vowel is missing in the syllable arrangement is compensated by the corresponding phoneme of the correct phoneme,
A vowel matching rate is calculated between the vowel arrangement of each of the corrected phoneme and the correct phoneme based on the utterance similarity.
4. The text conversion support device according to claim 3. The arithmetic device is
Comparing the number of vowels between the syllable arrays, if the number of vowels in the syllable array corresponding to the correct phoneme is smaller than the extracted phoneme, and if the syllable array of the correct phoneme is correct, then the extracted phoneme Correct by deleting the part where vowels are redundant in the syllable arrangement of
A vowel matching rate is calculated between the vowel arrangement of each of the corrected phoneme and the correct phoneme based on the utterance similarity.
4. The text conversion support device according to claim 3. The information processing device
Master data that defines the correct phoneme information for each vocabulary that is expected to appear in each conversation situation or subject is stored in a storage device,
A process of applying phone call recording data obtained from a predetermined device to an acoustic model to extract phonemes, and a process of extracting phonemes from the vocabulary in which phonemes are defined in the master data, which are expected to occur in the conversation scene or subject of the phone call recording data. a process of calculating a match rate between the correct phoneme of the vocabulary and the extracted phoneme; a process of identifying the vocabulary whose phonemes show a predetermined match rate as a keyword matching result as a result of the calculation;
A text conversion support method that executes. The information processing device
The storage device further retains information defining utterance similarity between vowels,
When calculating the matching rate, a process of calculating a matching rate between vowels included in each of the correct phoneme and the extracted phoneme based on the utterance similarity information, and a process of calculating the matching rate between the vowels included in each of the correct phoneme and the extracted phoneme, and The process of calculating the matching rate between the consonants included in each, and weighting the matching rate between the vowels more favorably than the matching rate between the consonants, and calculating the phoneme based on the matching rate between the vowels and between the consonants. Calculate the match rate between
7. The text conversion support method according to claim 6. The information processing device
When calculating the above match rate,
For each of the extracted phoneme and the correct phoneme, in the arrangement of vowels included in the phoneme, the degree of similarity in a set of two consecutive vowels is specified by the utterance similarity, and for the set for which the degree of similarity is equal to or higher than a reference value. is convolved into one predetermined vowel, and for pairs whose similarity is below the standard, the first vowel is adopted, and the process of forming the next pair with the vowel adjacent in the array is performed with the last vowel. to generate a syllable array,
The number of vowels is compared between the syllable arrays generated for each of the extracted phoneme and the correct phoneme, and if the numbers of vowels are equal, the extracted phoneme and the correct phoneme are the basis of the syllable array. calculating a vowel matching rate between each of the vowel arrays based on the utterance similarity;
8. The text conversion support method according to claim 7. The information processing device
Comparing the number of vowels between the syllable arrays, if the number of vowels in the syllable array corresponding to the correct phoneme is greater than the extracted phoneme, and if the syllable array of the correct phoneme is correct, then the extracted phoneme The part where a vowel is missing in the syllable arrangement is compensated by the corresponding phoneme of the correct phoneme,
calculating a vowel matching rate between the vowel arrays of the corrected extracted phoneme and the correct phoneme based on the utterance similarity;
9. The text conversion support method according to claim 8. The information processing device
Comparing the number of vowels between the syllable arrays, if the number of vowels in the syllable array corresponding to the correct phoneme is smaller than the extracted phoneme, and if the syllable array of the correct phoneme is correct, then the extracted phoneme Correct by deleting the part where vowels are redundant in the syllable arrangement of
calculating a vowel matching rate between the vowel arrays of the corrected extracted phoneme and the correct phoneme based on the utterance similarity;
9. The text conversion support method according to claim 8.

Images