JP5148532B2 - Topic determination device and topic determination method - Google Patents

Description

  The present invention relates to a topic determination device and a topic determination method.

  In recent years, in light of the fact that fraud cases using telephones have occurred, for example, in Patent Document 1, for example, by detecting a keyword related to fraud from speech content using voice recognition technology, for example, “ole fraud” is prevented. Technology is disclosed.

JP 2007-139864 A

  In conventional techniques such as Patent Document 1, topic determination is performed by keyword detection. That is, pre-set keywords (such as “money”, “exploited”, “audit”, “embedding”, etc.) to determine the topic of fraud, and count the number of occurrences of the set keyword during utterance, If the number of appearances exceeds a certain number, it is determined that the topic is a fraud.

Such a conventional method has a problem in that when the recognition rate for a keyword is lowered, the accuracy of topic determination is also lowered. For example, assume that the following utterance is made.
Speaking: “We spent a lot of company money. If we do n’t pay, we wo n’t be able to meet the audit.
As a result of performing speech recognition on this utterance, it is assumed that the recognition error is as follows.
Speech recognition result: “You have used your company's money. If you don't pay, you won't be in time for the test.
Note that such misrecognition is not particularly uncommon and can be sufficiently caused depending on the accuracy of speech recognition.

  In the case of such misrecognition as described above, keywords such as “used”, “audit”, “embedding” set in advance cannot be detected, and it is difficult to determine the topic of fraud. Furthermore, since the telephone voice has a narrow band and is greatly affected by noise, the recognition rate is low. In such a situation, it is very difficult to appropriately determine the topic of fraud only with keywords. It is also difficult to set keywords. For example, when “money”, “audit”, etc. are set as keywords, only the detection of such a keyword in the utterance is used as the criterion for fraud topics. .

  Therefore, the present invention has been made in view of the above, and an object thereof is to provide a topic determination device and a topic determination method capable of accurately determining topics such as fraud, for example.

  In order to solve the above-described problems, a topic determination device according to the present invention uses a data collection unit that collects data specialized for a specific topic, and data collected by the data collection unit. A topic specialized language model creating means for creating a specialized topic specialized language model, a language model holding means for holding the topic specialized language model created by the topic specialized language model creating means, and speech recognition Using the acoustic model holding means for holding the acoustic model, the topic-specific language model held in the language model holding means and the acoustic model held in the acoustic model holding means, and performing speech recognition on the input speech; and A voice recognition means for calculating a score for the result of the voice recognition, and a criterion for judging whether or not the content of the input voice corresponds to the specific topic When a threshold value holding means for holding a threshold value and the score calculated by the voice recognition means are input and the score is equal to or higher than the threshold value held by the threshold value holding means, the content of the input voice is the specific topic. And determining means for determining that it corresponds.

  In the topic determination method of the present invention, the data collection unit collects data specialized for a specific topic, and the topic specialized language model creation unit uses the data collected by the data collection unit. A topic-specific language model creating step for creating a topic-specific language model specialized for the specific topic, and a language model holding means, the topic-specific language model created by the topic-specific language model creating means A language model holding step for holding, an acoustic model holding step for holding an acoustic model for performing voice recognition, and a topic-specific language model for which the voice recognition means is held by the language model holding means. And using the acoustic model held in the acoustic model holding means, voice recognition is performed on the input voice, and a score for the voice recognition result is obtained. A voice recognition step for calculating, a threshold value holding means, a threshold value holding step for holding a threshold value serving as a determination criterion for determining whether or not the content of the input voice corresponds to the specific topic, and a determination means, A determination step of inputting a score calculated by the voice recognition means and determining that the content of the input voice corresponds to the specific topic when the score is equal to or greater than a threshold value held by the threshold value holding means; It is characterized by providing.

  According to the topic determination device and the topic determination method of the present invention, speech recognition is performed on input speech using a topic-specific language model created from data specialized for a specific topic, and a score for the result is obtained. The topic is determined by comparing with a threshold value. This makes it possible to perform topic determination with higher accuracy than the topic determination method based on the number of appearances of a keyword. For example, even when a specific keyword is not recognized, appropriate topic determination is performed.

  In the topic determination device of the present invention, the topic specific language model may be a statistical language model in which the appearance probability of a word is modeled.

  By using the appearance probability of a word for topic determination, it is possible to perform topic determination with higher accuracy than the conventional topic determination method based on the mere appearance frequency of a word. For example, a uni-gram is a statistical language model that models the probability of word appearance.

  In the topic determination device of the present invention, the topic specific language model may be a statistical language model in which a connection probability between words is modeled.

  By using the connection probability between words for topic determination, it is possible to perform topic determination with higher accuracy than in the conventional topic determination method based on the mere appearance count of words. An example of a statistical language model that models the probability of connection between words is n-gram.

  In the topic determination device of the present invention, the specific topic may be a fraud topic.

  The present invention is particularly useful when determining the topic of fraud.

  In the topic determination device of the present invention, the input voice may be utterance using a telephone.

  According to the present invention, the accuracy is improved as compared with the conventional topic determination method based on the mere appearance number of words. For this reason, the present invention is particularly useful in the case of a telephone voice with a low recognition rate because the band is narrow and the influence of noise is large.

  Further, the topic determination device of the present invention includes a data collection means for collecting data specialized for a specific topic, and a topic specialization language specialized for the specific topic using the data collected by the data collection means. Language model holding means for holding a topic-specific language model creating means for creating a model, a topic-specific language model created by the topic-specific language model creating means, and a general topic language model not specialized for the specific topic Input using means, an acoustic model holding means for holding an acoustic model for speech recognition, a topic-specific language model held in the language model holding means, and an acoustic model held in the acoustic model holding means A first speech recognition for speech, and a general topic language model held in the language model holding unit and an acoustic model held in the acoustic model holding unit. Speech recognition means for performing second speech recognition on the input speech using a voice, further calculating a first score for the result of the first speech recognition and a second score for the result of the second speech recognition, and the speech recognition Determining means for inputting the first score and the second score calculated by the means and determining that the content of the input speech corresponds to the specific topic when the first score is equal to or greater than the second score; It is characterized by providing.

  In the topic determination method of the present invention, the data collection unit collects data specialized for a specific topic, and the topic specialized language model creation unit uses the data collected by the data collection unit. A topic-specific language model creating step for creating a topic-specific language model specialized for the specific topic, and a language model holding unit, the topic-specific language model created by the topic-specific language model creating unit, A language model holding step for holding a general topic language model not specialized for the specific topic, an acoustic model holding step for holding an acoustic model for performing voice recognition, and a voice recognition unit for , Using the topic-specific language model held in the language model holding means and the acoustic model held in the acoustic model holding means. Performing a first speech recognition, performing a second speech recognition on the input speech using a general topic language model held in the language model holding unit and an acoustic model held in the acoustic model holding unit; A voice recognition step of calculating a first score for the result of one voice recognition and a second score for the result of the second voice recognition; and a determination means that inputs the first score and the second score calculated by the voice recognition means. And a determination step of determining that the content of the input voice corresponds to the specific topic when the first score is equal to or higher than the second score.

  According to such a topic determination apparatus and topic determination method of the present invention, first speech recognition is performed on input speech using a topic-specific language model created from data specialized for a specific topic, and a specific topic The second speech recognition for the input speech is performed using a general topic language model that is not specialized for the topic, and the topic is determined by comparing the scores for the results. This makes it possible to perform topic determination with higher accuracy than the topic determination method based on the number of appearances of a keyword. For example, even when a specific keyword is not recognized, appropriate topic determination is performed. Further, there is an advantage that it is not necessary to separately set a threshold value.

  Further, the topic determination device of the present invention further includes keyword holding means for holding a keyword specialized for the specific topic, and the determination means has the first score equal to or higher than the second score, When the keyword is detected a predetermined number of times or more in the result of the first speech recognition or the result of the second speech recognition, it may be determined that the content of the input speech corresponds to the specific topic.

  When the topic comparison has progressed to some extent by the score comparison, further topic determination by keyword is performed at this point. That is, the first topic determination based on the language model and the second topic determination based on the keyword are repeated. Since the second topic determination by keyword is performed when the first topic determination is completed, it is relatively easy to set the keyword for the topic. Moreover, the accuracy of topic determination can be further increased by performing topic determination multiple times.

  ADVANTAGE OF THE INVENTION According to this invention, the topic determination apparatus and topic determination method which can determine topics, such as fraud, for example can be provided.

It is a composition outline figure of fraud determination device 100 concerning a 1st embodiment. 2 is a hardware configuration diagram of a fraud determination apparatus 100. FIG. 4 is a flowchart illustrating an operation performed by the fraud determination apparatus 100. It is a structure schematic diagram of the fraud determination apparatus 200 concerning 2nd Embodiment. 5 is a flowchart showing an operation performed by fraud determination apparatus 200. It is a structure schematic diagram of the fraud determination apparatus 300 concerning 3rd Embodiment. 4 is a flowchart showing an operation performed by fraud determination apparatus 300.

  DESCRIPTION OF EMBODIMENTS Hereinafter, preferred embodiments of a topic determination device and a topic determination method according to the present invention will be described in detail with reference to the accompanying drawings. In the description of the drawings, the same elements are denoted by the same reference numerals, and redundant description is omitted.

<First Embodiment>
(Configuration of fraud determination apparatus 100)
First, the configuration of the fraud determination apparatus 100 according to the first embodiment of the present invention will be described with reference to FIG. 1 and FIG. FIG. 1 is a schematic configuration diagram of the fraud determination apparatus 100, and FIG. 2 is a hardware configuration diagram thereof. The fraud determination device 100 is a device that performs voice recognition on input speech that is utterance using, for example, a telephone (not shown), and determines whether or not the content is content related to a specific topic. . The “specific topic” in the present embodiment is a topic of fraud using a telephone such as “ole fraud”, for example.

  As shown in FIG. 2, the fraud determination device 100 physically includes a CPU 11, a main storage device such as a ROM 12 and a RAM 13, an input device 14 such as a keyboard and a mouse, an output device 15 such as a display, such as a telephone. It is comprised as a normal computer system containing the communication module 16 which transmits / receives data between these apparatuses (not shown), auxiliary storage devices 17, such as a hard disk. Each function of the fraud determination apparatus 100 described later allows the input device 14, the output device 15, and the communication module 16 to be controlled under the control of the CPU 11 by reading predetermined computer software on hardware such as the CPU 11, the ROM 12, and the RAM 13. This is realized by operating and reading and writing data in the main storage devices 12 and 13 and the auxiliary storage device 17.

  As shown in FIG. 1, the fraud determination device 100 functionally includes a voice input unit 110, a fraud data collection unit 120 (corresponding to “data collection means” in the claims), and a fraud language model creation unit 130 ( (Corresponding to “topic specialized language model creation means” in claims), fraud language model holding unit 140 (corresponding to “language model holding means” in claims), acoustic model holding unit 150 (claims) ), A voice recognition unit 160 (corresponding to “voice recognition unit” in claims), a threshold holding unit 170 (corresponding to “threshold holding unit” in claims), and A determination unit 180 (corresponding to “determination means” in the claims) is provided.

  The voice input unit 110 receives input voice, and can be physically configured as the communication module 16 shown in FIG. When the fraud determination apparatus 100 is configured separately from the telephone, the voice input unit 110 receives input voice data from the telephone through a wired or wireless network (not shown). When the fraud determination apparatus 100 is incorporated as one module in the telephone, the voice input unit 110 receives input voice data through a predetermined communication path (not shown) in the telephone. The voice input unit 110 outputs the input voice data to the voice recognition unit 160.

The fraud data collection unit 120 collects language data specialized for a specific topic, that is, a fraud topic. Examples of language data collected by the fraud data collection unit 120 are shown below. The fraud data collecting unit 120 may update the fraud data at any time according to the fraud tricks.
Examples of language data:
“I was asked by my friend when I was a student and became a guarantor of debt.”
“My friend couldn't repay, and the guarantor servant borrowed from Sarah and returned it.”
“The husband is getting caught in the train and doing obscenity. If the talk is not successful, it will be tried.”

The fraud language model creation unit 130 uses the data collected by the fraud data collection unit 120 to create a fraud topic-specific language model that is a language model specialized for the topic of fraud (the “topic-specific language model in the claims”). Is equivalent). Hereinafter, “fraud topic specific language model” is abbreviated as “fraud LM”, and “language model” is abbreviated as “LM”. As an example of the fraud LM created by the fraud LM creation unit 130, for example, a statistical language model in which the appearance probability of a word such as uni-gram is modeled. Another example of the fraud LM created by the fraud LM creation unit 130 is a statistical language model that models the connection probability between words such as n-grams. Since the technique itself for creating the statistical language model is a known technique as disclosed in Reference Document 1 below, for example, description thereof is omitted here.
Reference 1: Speech recognition system, ISBN / ASIN: 4-274-13228-5, Kiyohiro Shikano et al., Ohm

  Since the fraud LM created in this way is specialized in fraud topics, the appearance probability of words related to fraud and the connection probability between words are set high. For this reason, if the input voice is fraudulent content, the result of voice recognition for the input voice has a high score in the fraud LM. On the other hand, the appearance probability of words not related to fraud and the connection probability between words are set low in the fraud LM. For this reason, if the input voice is not fraudulent content, the result of voice recognition for the input voice has a low score in the fraud LM.

  The fraud LM holding unit 140 holds the fraud LM created by the fraud LM creation unit 130.

  The acoustic model holding unit 150 holds an acoustic model for performing voice recognition.

  The voice recognition unit 160 performs voice recognition on the input voice input from the voice input unit 110 using the fraud LM held in the fraud LM holding unit 140 and the acoustic model held in the acoustic model holding unit 150. is there. After performing speech recognition, the speech recognition unit 160 further calculates a score S for the result, and outputs the calculated score S to the determination unit 180. In addition, since the method of performing speech recognition and the method of calculating the score itself are known techniques as disclosed in, for example, Reference Document 1 described above, description thereof is omitted here. The speech recognition unit 160 may ignore the result of speech recognition and output only the score S derived from the word string that is the result of speech recognition to the determination unit 180 as a parameter for topic determination processing. This has the advantage that it is not necessary to hold the result of speech recognition. Further, the voice recognition unit 160 may perform normalization on the score S based on the length of the utterance.

  The threshold value holding unit 170 holds a threshold value X required when the determination unit 180 performs the topic determination process. The threshold value X is a criterion for determining whether or not the content of the input voice corresponds to the topic of fraud.

  The determination unit 180 inputs the score S calculated by the speech recognition unit 160, and determines that the content of the input speech corresponds to the topic of fraud when the score S is equal to or greater than the threshold value X stored in the threshold storage unit 170. To do.

(Operation of fraud determination apparatus 100)
Next, an operation performed by the fraud determination apparatus 100 (“topic determination method” in the claims) will be described with reference to FIG. FIG. 3 is a flowchart showing operations performed by the fraud determination apparatus 100. In the following description, fraud data collection processing and fraud LM creation processing by the fraud data collection unit 120 and fraud LM creation unit 130 (“data collection step” and “topic specialization language model creation step” in the claims) It is assumed that the created fraud LM is already held in the fraud LM holding unit 140 (corresponding to “language model holding step” in the claims). Also, the acoustic model is already held in the acoustic model holding unit 150 (corresponding to “acoustic model holding step” in the claims). In addition, it is assumed that the threshold value holding unit 170 already holds a threshold value X that is a criterion for determining whether or not the content of the input voice corresponds to the topic of fraud (see “ Equivalent to “threshold holding step”).

  First, the voice input unit 110 receives the input voice and outputs it to the voice recognition unit 160. (Step S11).

  Next, the voice recognition unit 160 performs voice recognition on the input voice input from the voice input unit 110 using the fraud LM held in the fraud LM holding unit 140 and the acoustic model held in the acoustic model holding unit 150. Is performed (step S12, corresponding to “voice recognition step” in the claims).

  Next, the speech recognition unit 160 calculates a score S for the result of speech recognition in step S12, and outputs the calculated score S to the determination unit 180 (step S13, “voice recognition step” in the claims). Equivalent).

  Next, when the determination unit 180 inputs the score S calculated by the speech recognition unit 160 and the score S is equal to or greater than the threshold value X held in the threshold holding unit 170 (step S14: YES), It is determined that the content corresponds to the topic of fraud (step S15, corresponding to the “determination step” in the claims). On the other hand, when the score S is not equal to or higher than the threshold value X held in the threshold value holding unit 170 (step S14: NO), the process ends as it is. In other words, it is not judged as a topic of fraud.

In the above flow, it is assumed that, for example, the following input voice is received in step S11.
Example of input speech: “We spent company money. If we don't pay, we wo n’t be able to meet the audit.
It is assumed that the speech recognition process in step S12 is performed on such input speech and the following result is obtained.
An example of speech recognition results: “I used my company's money. If I don't pay it, I can't make it in time for the test.
The score calculation process of step S13 is performed on the result of such voice recognition process, and when the score S is, for example, 0.8 and the threshold value X is, for example, 0.7, the input voice is a topic of fraud in step S15. It is determined that In other words, according to the present embodiment, keywords such as “used”, “audit”, and “embedding” are not properly recognized in the case of telephone voice with a particularly low recognition rate due to the narrow band and the influence of noise. Nevertheless, it is appropriately determined whether or not the input voice is a topic of fraud.

  According to the first embodiment of the present invention described above, speech recognition for input speech is performed using a fraud LM created from data specialized for a specific topic of fraud, and the score S for the result is set as a threshold value X. The topic is determined by comparing with. This makes it possible to perform topic determination with higher accuracy than the topic determination method based on the number of appearances of a keyword. For example, even when a specific keyword is not recognized, appropriate topic determination is performed.

  In addition, by using the word appearance probability in the topic determination, it is possible to perform topic determination with higher accuracy than in the conventional topic determination method based on the mere appearance count of the word. For example, a uni-gram is a statistical language model that models the probability of word appearance.

  In addition, by using the connection probability between words for topic determination, it is possible to perform topic determination with higher accuracy than the conventional topic determination method based on the mere appearance frequency of words. An example of a statistical language model that models the probability of connection between words is n-gram.

  As described above, according to the present invention, the accuracy is improved as compared with the conventional topic determination method based on the mere appearance number of words. For this reason, the present invention is particularly useful in the case of a telephone voice with a low recognition rate because the band is narrow and the influence of noise is large.

Second Embodiment
Subsequently, a second embodiment of the present invention will be described. In addition, description is abbreviate | omitted about the part which overlaps with 1st Embodiment described above, and it demonstrates centering around difference with 1st Embodiment.

  FIG. 4 is a schematic configuration diagram of a fraud determination apparatus 200 according to the second embodiment. Compared with the fraud determination apparatus 100 in the first embodiment, the fraud determination apparatus 200 corresponds to a general language model holding unit 210 (corresponding to “language model holding unit” in the claims, and hereinafter referred to as “general LM holding unit 210”). ), Instead of the voice recognition unit 160, a first voice recognition unit 220 (corresponding to “voice recognition means” in the claims) and a second voice recognition unit 230 (“voice recognition means” in the claims) And the threshold value holding unit 170 is not provided.

  The general LM holding unit 210 holds a general LM (corresponding to the “general topic language model” in the claims), which is a language model not specialized in the topic of fraud. The general LM may be created in the fraud determination apparatus 200, or may be inputted externally. Since the technique itself for creating the language model is a known technique as disclosed in, for example, the above-mentioned Reference Document 1, description thereof is omitted here. Since general LM is not specialized in fraud topics, the probability of occurrence of words related to fraud and the probability of connection between words are particularly higher than the probability of appearance of words not related to fraud and the probability of connection between words. is not.

  The first voice recognition unit 220 uses the fraud LM held in the fraud LM holding unit 140 and the acoustic model held in the acoustic model holding unit 150 to perform voice recognition on the input voice input from the voice input unit 110 (first 1 speech recognition). After performing the first speech recognition, the speech recognition unit 160 further calculates a score (first score S1) for the result, and outputs the calculated first score S1 to the determination unit 180.

  The second voice recognition unit 230 uses the general LM held in the general LM holding unit 210 and the acoustic model held in the acoustic model holding unit 150 to perform voice recognition on the input voice input from the voice input unit 110 (first 2 speech recognition). After performing the second speech recognition, the speech recognition unit 160 further calculates a score (second score S2) for the result, and outputs the calculated second score S2 to the determination unit 180.

  The determination unit 180 inputs the first score S1 and the second score S2 calculated by the first speech recognition unit 220 and the second speech recognition unit 230, and is input when the first score S1 is equal to or greater than the second score S2. It is determined that the audio content corresponds to the topic of fraud.

(Operation of fraud determination device 200)
Next, an operation performed by the fraud determination apparatus 200 (“topic determination method” in the scope of claims) will be described with reference to FIG. FIG. 5 is a flowchart showing operations performed by the fraud determination apparatus 200. In the following description, fraud data collection processing and fraud LM creation processing by the fraud data collection unit 120 and fraud LM creation unit 130 (“data collection step” and “topic specialization language model creation step” in the claims) The fraud LM created is already held in the fraud LM holding unit 140, and a general LM not specialized in the topic of fraud is already held in the general LM holding unit 210. (Corresponding to “language model holding step” in the claims). Also, the acoustic model is already held in the acoustic model holding unit 150 (corresponding to “acoustic model holding step” in the claims).

  First, the voice input unit 110 receives an input voice and outputs it to the first voice recognition unit 220 and the second voice recognition unit 230. (Step S21).

  Next, the first voice recognition unit 220 uses the fraud LM held in the fraud LM holding unit 140 and the acoustic model held in the acoustic model holding unit 150 to perform the first processing on the input voice input from the voice input unit 110. 1 voice recognition is performed (step S22, corresponding to “voice recognition step” in the claims).

  Next, the first speech recognition unit 220 calculates a first score S1 for the speech recognition result in step S22, and outputs the calculated first score S1 to the determination unit 180 (step S23, claims) Equivalent to “voice recognition step”).

  Next, the second voice recognition unit 230 uses the general LM held in the general LM holding unit 210 and the acoustic model held in the acoustic model holding unit 150 to perform the second processing on the input voice input from the voice input unit 110. 2. Perform voice recognition (step S24, corresponding to “voice recognition step” in the claims).

  Next, the second speech recognition unit 230 calculates a second score S2 for the result of speech recognition in step S24, and outputs the calculated second score S2 to the determination unit 180 (step S25, claims) Equivalent to “voice recognition step”).

  Next, the determination unit 180 inputs the first score S1 and the second score S2 calculated by the first speech recognition unit 220 and the second speech recognition unit 230, and the first score S1 is greater than or equal to the second score S2. On the other hand (step S26: YES), it is determined that the content of the input voice corresponds to the topic of fraud (step S27, corresponding to the “determination step” in the claims). On the other hand, when the first score S1 is not equal to or higher than the second score S2 (step S26: NO), the process is ended as it is. In other words, it is not judged as a topic of fraud.

In the above flow, it is assumed that, for example, the following input voice is received in step S21.
Example of input speech: “We spent company money. If we don't pay, we wo n’t be in time for the audit.
It is assumed that the speech recognition process in step S22 and the speech recognition process in step S24 are performed on such input speech, and the following results are obtained.
Example of S22 result: “I used my card money. If I don't pay, I can't make it in time for the test.
Example of S24 result: “You have used your company's money. If you do not pay, you will not be in time for the inspection.
If the first score S1 is 0.8, for example, and the second score S2 is 0.4, for example, the score calculation processing of step S23 and step S25 is performed on the result of such voice recognition processing, and the above-mentioned in step S27. The input voice is determined to be a topic of fraud. In other words, according to the present embodiment, keywords such as “used”, “audit”, and “embedding” are not properly recognized in the case of telephone voice with a particularly low recognition rate due to the narrow band and the influence of noise. Nevertheless, it is appropriately determined whether or not the input voice is a topic of fraud.

  According to the second embodiment of the present invention described above, the first speech recognition for the input speech is performed using the fraud LM created from the data specialized for the specific topic of fraud, and the specific topic of fraud The second speech recognition is performed on the input speech using a general LM that is not specialized for the topic, and topic determination is performed by comparing respective scores for the results. This makes it possible to perform topic determination with higher accuracy than the topic determination method based on the number of appearances of a keyword. For example, even when a specific keyword is not recognized, appropriate topic determination is performed. Further, there is an advantage that it is not necessary to separately set a threshold value.

<Third Embodiment>
Subsequently, a third embodiment of the present invention will be described. In addition, description is abbreviate | omitted about the part which overlaps 1st Embodiment described above, or 2nd Embodiment, and it demonstrates centering around difference with 1st Embodiment or 2nd Embodiment.

  FIG. 6 is a schematic configuration diagram of a fraud determination apparatus 300 according to the third embodiment. Compared to the fraud determination apparatus 200 in the second embodiment, the fraud determination apparatus 300 further includes a keyword holding unit 310 (corresponding to “keyword holding means” in the claims). The keyword holding unit 310 holds a keyword specialized in a specific topic such as fraud, for example.

  The determination unit 180 of the fraud determination apparatus 300 according to the third embodiment inputs the first score S1 and the second score S2 calculated by the first speech recognition unit 220 and the second speech recognition unit 230, and the first score S1 is In addition to being above the second score S2, if the keyword is detected more than a certain number of times in the first speech recognition result or the second speech recognition result, it is determined that the content of the input speech corresponds to the topic of fraud .

(Operation of fraud determination device 300)
Next, an operation performed by the fraud determination apparatus 300 (“topic determination method” in the scope of claims) will be described with reference to FIG. FIG. 7 is a flowchart showing operations performed by the fraud determination apparatus 300. In the following description, fraud data collection processing and fraud LM creation processing by the fraud data collection unit 120 and fraud LM creation unit 130 (“data collection step” and “topic specialization language model creation step” in the claims) The fraud LM created is already held in the fraud LM holding unit 140, and a general LM not specialized in the topic of fraud is already held in the general LM holding unit 210. (Corresponding to “language model holding step” in the claims). Also, the acoustic model is already held in the acoustic model holding unit 150 (corresponding to “acoustic model holding step” in the claims).

  First, the processes of steps S21 to S25 described in the second embodiment are performed (corresponding to the “voice recognition step” in the claims).

  Next, the determination unit 180 inputs the first score S1 and the second score S2 calculated by the first speech recognition unit 220 and the second speech recognition unit 230, and the first score S1 is greater than or equal to the second score S2. (Step S26: YES), it is tentatively determined that the content of the input voice corresponds to the topic of fraud. This is called the first stage of fraud topic determination.

  Next, when the determination unit 180 detects the keyword more than a certain number of times in the first speech recognition result or the second speech recognition result (step S31: YES), the content of the input speech corresponds to the topic of fraud. Then, finally judge. This is called the second stage of fraud topic determination. In the present embodiment, the predetermined number of times is set to one time, for example.

  On the other hand, when the first score S1 is not equal to or higher than the second score S2 (step S26: NO), or when the first score S1 is equal to or higher than the second score S2 but no keyword is found (step S31: NO), processing is performed. Ends as is. In other words, it is not judged as a topic of fraud.

  In the above flow, the reason for performing the topic determination in two stages of step S26 and step S31 is as follows. That is, for example, in a fraud case such as “Ole scam”, a person who commits a fraud first explains the background. For example, deceive the police and explain the occurrence of criminal cases in the victim's family and their progress and results. If this stage is the first stage of fraud, through this first stage, the person who commits the fraud gains the trust of the victim and finally enters the main topic, that is, the actual transfer. If the stage where a person who commits a fraud talks about the subject matter is the second stage of fraud, the content of the utterance in this second stage tends to be a fixed phrase. For example, “Please transfer **** yen money to your account ***!”. The words used in this second stage are often very important words in fraud, and the types and numbers thereof are relatively small. That is, regardless of the content of the background explanation in the first stage, the utterance in the second stage tends to be a fixed phrase that can be predicted in advance. Therefore, unlike the keyword in the first stage, the keyword in the second stage is not difficult to set in advance. Moreover, the accuracy of topic determination is further enhanced by the keywords in the second stage. In the prior art, it can be said that the accuracy of the topic is not high because the topic is determined by the keyword without considering the concept of the first stage and the second stage in the present embodiment.

  According to the third embodiment of the present invention described above, when a certain degree of topic determination has progressed by score comparison, further topic determination by keyword is performed at this point. That is, the first topic determination based on the language model and the second topic determination based on the keyword are repeated. Since the second topic determination by keyword is performed when the first topic determination is completed, it is relatively easy to set the keyword for the topic. Moreover, the accuracy of topic determination can be further increased by performing topic determination multiple times.

  As mentioned above, although preferred embodiment of this invention was described, it cannot be overemphasized that this invention is not limited to the said embodiment.

  For example, in the above embodiment, “fraud” is taken as an example of “specific topic”, but the present invention is not limited to this, and may be a topic other than “fraud”.

  In the above embodiment, uni-gram and n-gram have been described as examples of language models. However, the present invention is not limited to this, and other language models other than uni-gram and n-gram are used in the present invention. You may use suitably for this concept.

DESCRIPTION OF SYMBOLS 100,200,300 ... Fraud judging device, 110 ... Voice input part, 120 ... Fraud data collection part, 130 ... Fraud language model creation part (fraud LM creation part), 140 ... Fraud language model holding part (fraud LM holding part) , 150 ... acoustic model holding unit, 160 ... voice recognition unit, 170 ... threshold holding unit, 180 ... determination unit, 210 ... general language model holding unit (general LM holding unit), 220 ... first voice recognition unit, 230 ... first 2 voice recognition unit, 310... Keyword holding unit.

Claims (2)

A data collection means for collecting data specific to a particular topic;
Using the data collected by the data collection means, topic-specific language model creation means for creating a topic-specific language model specialized for the specific topic;
A language model holding means for holding a topic-specific language model created by the topic-specific language model creating means, and a general topic language model not specialized for the specific topic;
Acoustic model holding means for holding an acoustic model for voice recognition;
The first speech recognition is performed on the input speech using the topic-specific language model held in the language model holding unit and the acoustic model held in the acoustic model holding unit, and the general language held in the language model holding unit Second speech recognition is performed on the input speech using a topic language model and an acoustic model held in the acoustic model holding unit, and further, a first score for the result of the first speech recognition and a result of the second speech recognition are Voice recognition means for calculating a second score;
A determination unit that inputs the first score and the second score calculated by the voice recognition unit, and determines that the content of the input voice corresponds to the specific topic when the first score is equal to or higher than the second score. When,
Comprising a keyword holding means for holding a keyword specialized for the specific topic,
When the determination unit detects that the keyword is detected more than a certain number of times in the result of the first speech recognition or the result of the second speech recognition after it is determined that the first score is greater than or equal to the second score. And determining that the content of the input voice corresponds to the specific topic,
The specific topic is a topic of fraud,
The topic determination device according to claim 1, wherein the input voice is utterance using a telephone . A data collection step in which the data collection means collects data specific to a specific topic;
Topic-specific language model creation means uses the data collected by the data collection means to create a topic-specific language model creation step for creating a topic-specific language model specialized for the specific topic;
A language model holding unit that holds a topic-specific language model created by the topic-specific language model creation unit and a general topic language model that is not specialized for the specific topic;
An acoustic model holding unit that holds an acoustic model for performing speech recognition; and
The speech recognition means performs first speech recognition on the input speech using the topic-specific language model held in the language model holding means and the acoustic model held in the acoustic model holding means, and the language model holding means The second speech recognition for the input speech is performed using the general topic language model held in the sound model and the acoustic model held in the acoustic model holding means, and the first score and the second score for the result of the first speech recognition are further performed. A speech recognition step of calculating a second score for the result of speech recognition;
When the determination unit inputs the first score and the second score calculated by the speech recognition unit, and the first score is equal to or higher than the second score, the content of the input speech corresponds to the specific topic. A determination step for determining;
The keyword holding means includes a keyword holding step for holding a keyword specialized for the specific topic,
In the determination step, after the determination means determines that the first score is greater than or equal to the second score, the keyword is added to the result of the first speech recognition or the result of the second speech recognition a predetermined number of times. When detected above, it is determined that the content of the input voice corresponds to the specific topic,
The specific topic is a topic of fraud,
A topic determination method , wherein the input voice is utterance using a telephone .

Images