JP4972107B2 - Call state determination device, call state determination method, program, recording medium - Google Patents

Description

  The present invention relates to, for example, a call state determination apparatus, a call state determination method, a program, and a recording medium that estimate an emotional state of each speaker from an input voice signal under a situation where two speakers interact.

  In recent years, there has been an active movement to obtain useful information for companies from the raw voices such as requests and dissatisfaction from customers gathering at call centers held by companies. In addition, the call center has started to emphasize the function of the company's face, and companies are also making efforts to improve the service quality of the call center in order to improve the customer's image of the company. Under such circumstances, a technology that automatically finds a call in which a customer is angry (hereinafter referred to as a “claim call”) is desired more than ever. By analyzing complaint calls, you can find out the customer's strong demands and dissatisfactions, product / service defects and problems, operator-related problems that cause complaint calls, and the operator's response status in real time By monitoring, it becomes possible to quickly detect and respond to the occurrence of a complaint.

  As emotion recognition technology for speaker's “angry” voice to find a claim call, conventionally the voice height (pitch frequency) and magnitude (power), or their variation component (Δ component), speech speed, etc. An acoustic feature amount of an utterance speech section focusing on speech features is generally used (see Patent Document 1, hereinafter referred to as “Prior Art 1”).

  In addition, a method has been proposed that pays attention to the operator's interaction in a call (see Patent Document 2, hereinafter referred to as “Prior Art 2”). In this method, a pre-registered companion word is recognized by voice, and it is determined whether or not it is in a complaint state based on the start time when the companion was first heard from the call start time.

JP 2005-345496 A JP 2007-286097 A

  The problem of Prior Art 1 will be described. The customer's voice recorded at the call center is telephone voice, and since the recorded noise and band limiting filter are applied, it is difficult to extract the pitch frequency (pitch pattern) and it is easy to detect it erroneously. Further, even if the speaker speaks with the same voice depending on the volume setting of the telephone or the distance between the speaker's mouth and the receiver, the power value calculated by the receiving recorder is different. The voice of the customer to be analyzed (or the voice of the operator) varies depending on the way of speaking and the telephone environment. Accordingly, it is difficult to accurately calculate acoustic feature quantities such as voice pitch (pitch frequency) and loudness (power), and it is difficult to make a claim determination robustly with high accuracy. Also, the manner of anger varies depending on the speaker, and there is a case where a person gets angry with a angry voice or a person who gets angry with a calm voice. Also, even a speaker who gets angry and angry does not always cry out from the beginning to the end of the conversation. Therefore, it is difficult to determine whether a customer is angry based only on acoustic features such as voice pitch and loudness.

  The problem of Prior Art 2 will be described. With this technology, a complaint cannot be determined unless the call is from the start of the call. Therefore, for example, when the operator's response is poor and the customer gets angry from the middle of the call, it is difficult to recognize the complaint call.

  In order to solve the above-described problem, the call state determination device of the present application is a call state determination device that determines whether the first speaker talking with the second speaker is in the first state or the second state. A detection unit, an extraction unit, a calculation unit, a vectorization unit, a score calculation unit, a determination unit, and an output unit. A detection part detects the speech area of a 1st speaker and a 2nd speaker. The extraction unit extracts a predetermined number of utterance pairs as segments. The calculation unit calculates an interactive feature amount related to the utterance situation for each utterance pair. A vectorization part calculates | requires a feature vector by totaling an interactive feature-value for every segment. The score calculation unit obtains the first state score by substituting each element of the feature vector into a predetermined discriminant. The determination unit determines the segment as the first state segment if the first state score is equal to or greater than a predetermined first threshold. The output unit outputs information indicating that the call state is the first state if the first state segment is equal to or greater than a predetermined second threshold value.

  According to the call determination device of the present invention, whether or not the customer is in a complaint state using an interactive feature quantity that is an interactive feature, instead of an acoustic feature quantity such as voice pitch and loudness as in Prior Art 1. Determine whether or not. Since the interactive feature amount is not changed by the customer's telephone environment, the customer's voice or how to speak, the problem of the prior art 1 can be solved, and the claims can be determined robustly and accurately.

  Further, as in the prior art 2, it is determined whether or not the complaint state is based on the start time when the first contact is made from the start of the call, but in the present invention, for each utterance pair in the call, Since it is determined whether or not it is a complaint state using the interactive feature amount, even if the customer gets angry from the middle of the call, the complaint can be determined robustly and accurately, and the problem of the prior art 2 The point can be solved.

It is the figure which showed the function structural example of the telephone call state determination apparatus of a present Example. It is the figure which showed the processing flow of the call state determination apparatus of a present Example. A to D are diagrams each showing an utterance pair. It is the figure shown about the segment. A to C are diagrams showing interactive feature amounts.

  In the following description, in the call center, the first speaker is the customer, the second speaker is the operator, the first state is the state where the customer is angry (hereinafter referred to as “claim state”), and the second state is It is assumed that the customer is not angry (that is, a normal state, hereinafter referred to as “non-claimed state”). In addition, what each customer and operator are talking is called utterance, and a set of utterances is called a call.

FIG. 1 shows a functional configuration example of the call state determination device 100 and the like, and FIG. 2 shows a processing flow. When the call voice of the customer and the operator is input to the input terminal 2, the detection unit 4 detects the utterance section between the customer (first speaker) and the operator (second speaker) (step S2). Specifically, the voices of the customer and the operator are separated using existing hardware such as a separation adapter or sound source separation technology, and the start time and end time are obtained for each utterance. As a technique for voice activity detection, and the like method of the third threshold value L ₃ or more a voice power certain time or subsequently interval the speech period predetermined that defines optionally.

  3A to 3D schematically show an operator's utterance and a customer's utterance. The unhatched rectangle indicates the operator's utterance section, the hatched rectangle indicates the customer's utterance section, and the horizontal axis indicates the time axis. As shown in FIG. 3, when the operator and the customer are recorded separately in stereo, it is easier to detect the utterance period than monaural recording. In the case of monaural recording, it is necessary to use in combination with means for identifying the voice between the customer and the operator. For example, the voice and the operator's voice may be separated by using a voice spectrum as a feature quantity and using a GMM (Gaussian Mikuture Model).

  Here, as shown in FIG. 3A, when the continuous utterance (of the operator in the example of FIG. 3A) ends and the other party (customer in the example of FIG. 3A) starts speaking, it is considered that the utterance right is changed. . Then, an utterance section is defined from the utterance start time to the utterance end time. As for the utterance of the operator shown in FIG. 3B, the utterance is stopped halfway, but the operator starts speaking again without changing the utterance right. In this case as well, it is set as one utterance section of the operator. In addition, as shown in FIG. 3C, the customer is in a state of mutual agreement with the operator's utterance, but in this case as well, the utterance right is not changed, and this case is also regarded as one utterance of the operator. A pair of one utterance of one customer and one utterance of an operator is called an utterance pair. That is, in the example of FIGS. 3A to 3C, a pair of an operator's utterance and a customer's utterance is an utterance pair. 3A to 3C show the utterance pairs in the order of the operator → customer, but as shown in FIG. 3D, the utterance pairs in the order of the customer → operator are also called utterance pairs.

  Next, the extraction unit 6 extracts a predetermined number Q of utterance pairs as segments (step S4). FIG. 4 shows a schematic diagram of segment extraction. In the example of FIG. 4, Q = 3, that is, a case where three utterance pairs are set as one segment is shown. First, when three utterance pairs are extracted, a predetermined number Q (three in this example) of utterance pairs are extracted again by sliding at regular intervals or for each utterance section. This process is repeated to extract segments. Further, the predetermined number Q may be set to 1, and in this case, one utterance pair is one segment.

  Next, the calculation unit 8 calculates an interactive feature amount R related to the utterance situation for each utterance pair of the first speaker (customer) and the second speaker (operator) (step S6). The interactive feature R is not an acoustic feature such as the loudness or the height of the voice in the utterance unit as shown in the prior art 1, but the interactive feature amount R is between the first speaker and the second speaker. This is a feature value for a dialogue consisting of a set of utterances.

  In this embodiment, as the interactive feature quantity R, (1) the utterance time A of the customer (first speaker) for each utterance pair (2) the utterance time B (3) of the operator (second speaker) for each utterance pair ) Discreteness of operator utterance time and customer utterance time for each utterance pair C (4) Customer interaction count D for each utterance pair D (5) Operator interaction count E for each utterance pair (6) Customer for each utterance pair Silence time F for the utterances and utterances of the operator F (7) The customer utterance for each utterance pair, the overlap time G for the utterances of the operator, and the like are used. Each will be described in detail. Moreover, the schematic diagram of the interactive feature-value of (1)-(7) is shown to FIG. 5A-FIG. 5C.

(1) Customer utterance time A
FIG. 5A shows the customer utterance time A. FIG. The reason for using the customer's utterance time as an interactive feature is
The customer's utterance time A is long ... the customer is in a complaint state.
Customer's utterance time A is short ... The customer is in a non-claimed state.
This is because this phenomenon is empirically understood and this phenomenon is used. This is because when the customer is in a complaint state, the customer often talks unilaterally to the operator, and the customer's utterance time A tends to be longer.

  In this case, the calculation unit 8 includes first calculation means 81 (see FIG. 1). The first calculation means 81 calculates customer utterance time. As an example of the calculation method of the utterance time, the time from the start time of the customer's utterance to the end time of the utterance is measured.

(2) Operator utterance time B
FIG. 5A shows the utterance time B of the operator. The reason for using the operator's utterance time B as the interactive feature is as follows:
The operator's utterance time B is long ... the customer is in a non-claimed state.
The utterance time B of the operator is short .... The customer is in a complaint state.
This is because this phenomenon is empirically understood and this phenomenon is used. This is because when the customer is in a complaint state, the customer often speaks unilaterally to the operator, the operator does not speak much, and the operator's utterance time B tends to be short.

  In this case, the calculation unit 8 includes second calculation means 82. The second calculation means 82 is for calculating the utterance time of the operator. As an example of the calculation method of the utterance time, the time from the utterance start time of the operator to the utterance end time is measured.

(3) Discreteness C between customer utterance time and operator utterance time
The degree of discreteness C is shown in FIG. 5B. Here, the degree of discreteness C indicates the degree to which the customer's utterance time A is more discrete than A and B compared to the operator's utterance time B, for example, a difference (AB) or a ratio ( A / B). The reason for using the discrete degree C of the customer utterance time and the operator utterance time as the interactive feature amount is as follows:
Discreteness C is large (AB and A / B are large, that is, the customer's utterance time A is large compared to the operator's utterance time B)...

Discreteness C is small (values of AB and A / B are small, that is, the customer's speech time A is small compared to the operator's speech time B)... The customer is in a non-claimed state.
This is because this phenomenon is empirically understood and this phenomenon is used. This is because when the customer is in a complaint state, the customer speaks unilaterally, so the customer's utterance time becomes longer and the operator's utterance time becomes shorter, so that the customer's utterance time A and the operator's utterance time B This is because the degree of discreteness C increases.

  Further, in a call, when the customer's utterance time is long (for example, when the customer's question is long), the utterance time of the operator who answers the long utterance may be long. In this case, the customer is not actually in the complaint state, but if the customer's utterance time A is used as the dialogue feature amount R, the customer's utterance time is long, so that the customer is in the complaint state ( The determination method will be described later. However, by using the degree of discreteness C, the customer's utterance time and the operator's utterance time are normalized, and even in such a case, there is no misjudgment of being in a complaint state.

  On the other hand, when the customer's utterance time is short, the utterance time of the operator who responds to the short utterance may be short. In this case, the customer is not actually in the complaint state. However, if the operator's utterance time B is used as the dialogue feature amount R, an erroneous determination that the customer is in the complaint state is made because the operator's utterance time is short. May do. However, by using the degree of discreteness C, the customer's utterance time and the operator's utterance time are normalized, and even in such a case, there is no misjudgment of being in a complaint state.

  In this case, the calculation unit 8 includes third calculation means 83. The third calculation means 83 calculates the customer utterance time A and the operator utterance time B, and obtains the discreteness (A−B, A / B) of A and B.

(4) Number of customer consultations D
FIG. 5A shows the number of customer consultations D. The reason for using the number of customer interactions D as an interactive feature is as follows:
The number of customer consultations D is large ... The customer is in a non-claimed state.
The number of customer consultations D is small ... the customer is in a complaint state.
This is because this phenomenon is empirically understood and this phenomenon is used. This is because when the customer is in a complaint state, the customer does not listen to the operator and speaks unilaterally, so that the number of customer interactions decreases.

  In this case, the calculation unit 8 includes fourth calculation means 84. The fourth calculation means 84 obtains the customer's consultation count D. Here, an example of a method for obtaining the number of conflicts will be described. For example, a word (for example, “Yes”, “Ah”, “Yes”, etc.) that the customer will utter when competing is determined, the customer's voice is recognized by voice recognition means (not shown), and the companion word A method of measuring the number of the above may be used.

(5) Number of operator consultations E
FIG. 5A shows the number E of the operator's discussions. The reason for using the operator's number of times of interaction E as the interactive feature is as follows:
Operator's number of times of interaction E is small .... The customer is in a non-claimed state.
Operator's number of times of interaction E is large ... The customer is in a complaint state.
This is because this phenomenon is empirically understood and this phenomenon is used. This is because when the customer is in a complaint state, the customer does not listen to the operator but speaks unilaterally, so that the number of times the operator is in conflict increases.

  In this case, the calculation unit 8 includes fifth calculation means 85. The fifth calculation means 85 obtains the number of times E of the operator's discussion. Here, an example of a method for obtaining the number of conflicts will be described. For example, a word (for example, “Yes”, “Yes”, “Sorry”, etc.) that the operator will utter when competing is determined, and voice recognition means (not shown) (“(4) Number of customer's conflicts” In D), a method of recognizing the operator's voice and measuring the number of conflicting words may be used.

(6) Silent time F
FIG. 5A shows the silent period F. Here, the silent time refers to the time during which the customer and the operator are not speaking for the utterance pair. The reason for using the silent time F for the customer utterance and the operator utterance for the utterance pair is as follows:
Silent time F is long ... The customer is in a complaint state.
Silent time F is short ... the customer is in a non-claimed state.
This is because this phenomenon is empirically understood and this phenomenon is used. This is because when the customer is angry, the operator often keeps silent (silence time F becomes longer), and when the customer is not angry, no silence is generated between the operator's utterance and the customer's utterance. This is because there are many cases where the silent time F is shortened.

In this case, the calculation unit 8 includes sixth calculation means 86. The sixth calculation means 86 obtains the silent time F. Here, a method example for obtaining the silent time F will be described. For example, the sixth calculating means 86, a speaker that speech (in the example of FIG. 5A, the operator) and the speech end time T ₁ of the utterance of the speaker is finished, the other speaker (Fig. 5 In this example, the customer's utterance start time T ₂ is measured. In FIG. 5A, T ₂ > T ₁ , and in this case, the silent time F is generated, and the sixth calculating unit 86 may calculate T ₂ −T ₁ . T ₂ −T ₁ is a positive value, and is a value of the silent time F.

(7) Overlap time G
The overlap time G is shown in FIG. 5C. Here, the overlap time G refers to the time during which both the customer and the operator speak in an overlapping manner with respect to the utterance pair. The reason for using overlap time G is
Overlap time G is long ... The customer is in a complaint state.
Overlapping time G is short ... the customer is in a non-claimed state.
This is because this phenomenon is empirically understood and this phenomenon is used. This is because when the customer is angry, the customer often interrupts the operator's speech, and when the customer is not angry, there is no overlap between the operator's speech and the customer's speech ( This is because there are many cases where the overlap time G is shortened).

In this case, the calculation unit 8 includes seventh calculation means 87. The seventh calculating means 87 calculates the overlap time G. Here, a method example for obtaining the overlap time G will be described. For example, the seventh calculating means 87 is the utterance end time T ₁ of the speaker who is speaking (the operator in the example of FIG. 5), and the speaker's utterance is terminated, and the other speaker (of FIG. 5). In the example, the customer's utterance start time T ₂ is measured. In FIG. 5C, T ₂ <T ₁ , and in this case, the overlap time G occurs, and T ₂ −T ₁ becomes a negative value and the overlap time F.

  Therefore, as described above, the interactive feature amount R includes the customer utterance time A, the operator utterance time B, the discreteness C, the customer interaction count D, the operator interaction count E, the silence time F, and the overlap time G. The calculation unit 8 may be provided with at least one of the first calculation unit 81 to the seventh calculation unit 87. Which interactive feature is to be used among these A to G may be appropriately determined. Further, the interactive feature amount R is not limited to these A to G.

  Moreover, the process of the calculation part 8 may be performed about the utterance pair after extracting as a segment, and may be performed about the utterance pair before extracting as a segment. In addition to the interactive feature amount, a conventional acoustic feature amount (for example, voice pitch (pitch frequency), loudness (power), speech speed, etc.) may be calculated.

The vectorization unit 10 obtains a feature vector by summing up interactive feature values for each segment (step S8). Specifically, vector elements of feature vectors are obtained from interactive feature values R (at least one of A to G) for Q (number of utterance pairs constituting a segment) utterance pairs. As a method for obtaining a vector element, for example, an average value, variance, maximum value, and minimum value may be obtained for Q utterance pairs. Among the Q utterance pairs, the values of the interactive feature quantities A to G of the q-th (1,..., Q,..., Q) utterance pair are respectively a _q , b _q , c _q , d. Assuming _{q 1} , e _q , f _q , and g _q , when all the interactive feature amounts A to G are used, the feature vector for each segment obtained by the vectorization unit 10 is
(Average of all _{a q,} the dispersion of all _{a q,} the maximum value of all the _{a q,} the minimum value of all _{a q,} ..., the average value of all _{g q,} all dispersion of g _q, maximum value of all g _q, the minimum value) of all g _q. In this case, the feature vector has 28 elements. As described above, since the interactive feature value to be used is at least one of A to G, a feature vector corresponding to the used interactive feature value is generated.

  The score calculation unit 12 obtains a first state score (claim score) by substituting each element of the feature vector into a predetermined discriminant F (X) (step S10). The discriminant H is obtained in advance by the learning device 200. The learning device 200 includes the same detection unit 4, extraction unit 6, calculation unit 8, vectorization unit 10, and learning unit 18 of the call state determination device 100. Hereinafter, how to obtain the discriminant F (X) will be described.

  The call database storage unit 20 stores a plurality of complaint calls and non-claimed calls. And about the some claim call and non-claim call, the process of the detection part 4-the vectorization part 10 is performed similarly to the call state determination apparatus 100. FIG. Further, the feature quantity vector R (A to G described above) used in the learning device 200 and the vector elements of the feature vector R (the above-described average value, variance, maximum value, and minimum value) are used in the call state determination device 100. Must be identical to them. The learning unit 18 performs machine learning of the feature vector from the vectorization unit 10. There are various learning methods. For example, a linear discrimination method, a support vector machine, a neural network, or the like may be used.

And the discriminant F (X) calculated | required from the learning apparatus 200 becomes the following formulas, when a linear discriminant method is used, for example.

Here, the value of M becomes the number of elements of the feature vector, each element of the feature vector is assigned to X _m, respectively. Further, when each element of the feature vector of the segment in which the customer is in the claim state is substituted as the discriminant F (X), the calculated claim score is large, and each of the feature vector of the segment in the non-claim state When an element is substituted, an expression that reduces the calculated claim score is obtained by learning. The learning apparatus 200 obtains the weighting coefficient α _m (m = 1,..., M) of the above equation (1) by learning.

And the determination part 14 determines whether a claim score is more than a predetermined 1st threshold value for every segment (step S12). Greater than the first threshold value, and the segment as claimed segments (intervals that were claim state), smaller than the first threshold value L _1, to the segment as a non-complaint state segment (segment is non claim state).

Output unit 16, if the second threshold value L ₁ or the state of Claims segment (first state segment) is predetermined, communication state outputs information indicating that the first state (step S14). The state of the claim segment indicates the number of claim segments and the ratio of the claim segments to the total segments. For example, if the second threshold L ₁ is 1, and there is at least one claim segment, the call may be regarded as a claim call (that is, the customer is angry). Further, and when the number of claims segment is the second threshold value L ₁ or more, the proportion of claims segment in the total segment may be regarded also claims call the call if the second threshold value L ₁ or more.

  As described above, according to the call determination device of the present invention, not using the acoustic feature quantity such as the voice pitch and loudness as in the prior art 1, but using the interactive feature quantity that is an interactive feature, Determines whether it is in a claim state. The interactive feature amount is not changed by the customer's telephone environment, the customer's voice or the way of speaking, and the variation is small, so that the claims can be determined robustly and accurately. Further, according to the present invention, it is possible to determine the customer's complaint state who is angry.

  Further, in the present invention, since it is determined whether or not it is a complaint state using the interactive feature amount for each utterance pair during a call, even if the customer gets angry during the call, Claims can be determined with high accuracy.

  In addition, since it captures the phenomenon that appears when the customer is angry, it is suitable for determining whether or not it is a set of angry utterances (= call).

<Hardware configuration>
The present invention is not limited to the above-described embodiment. In addition, the various processes described above are not only executed in time series according to the description, but may be executed in parallel or individually according to the processing capability of the apparatus that executes the processes or as necessary. Needless to say, other modifications are possible without departing from the spirit of the present invention.

  Further, when the above configuration is realized by a computer, the processing contents of the functions that the call state determination apparatus 100 should have are described by a program. The processing function is realized on the computer by executing the program on the computer.

  The program describing the processing contents can be stored in a computer-readable recording medium. The computer-readable recording medium may be any medium such as a magnetic storage device, an optical disk, a magneto-optical recording medium, or a semiconductor memory. Specifically, for example, as the magnetic storage device, a hard disk device, a flexible Discs, magnetic tapes, etc. as optical disks, DVD (Digital Versatile Disc), DVD-RAM (Random Access Memory), CD-ROM (Compact Disc Read Only Memory), CD-R (Recordable) / RW (ReWritable), etc. As the magneto-optical recording medium, MO (Magneto-Optical disc) or the like can be used, and as the semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory) or the like can be used.

  The program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM storing the program, for example. Furthermore, the program may be distributed by storing the program in a storage device of the server computer and transferring the program from the server computer to another computer via a network.

  A computer that executes such a program first stores, for example, a program stored in a portable recording medium or a program transferred from a server computer in its own storage device. When executing the process, the computer reads a program stored in its own recording medium and executes a process according to the read program. As another execution form of the program, the computer may directly read the program from a portable recording medium and execute processing according to the program, and the program is transferred from the server computer to the computer. Each time, the processing according to the received program may be executed sequentially. Also, the program is not transferred from the server computer to the computer, and the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes the processing function only by the execution instruction and result acquisition. It is good. Note that the program in this embodiment includes information that is used for processing by an electronic computer and that conforms to the program (data that is not a direct command to the computer but has a property that defines the processing of the computer).

  In this embodiment, the present apparatus is configured by executing a predetermined program on a computer. However, at least a part of these processing contents may be realized by hardware.

  The call state determination device 100 described in this embodiment includes a CPU (Central Processing Unit), an input unit, an output unit, an auxiliary storage device, a RAM (Random Access Memory), a ROM (Read Only Memory), and a bus. (Both not shown).

  The CPU executes various arithmetic processes according to the read various programs. The auxiliary storage device is, for example, a hard disk, an MO (Magneto-Optical disc), a semiconductor memory, or the like, and the RAM is an SRAM (Static Random Access Memory), a DRAM (Dynamic Random Access Memory), or the like. The bus connects the CPU, the input unit, the output unit, the auxiliary storage device, the RAM, and the ROM so that they can communicate with each other.

<Cooperation between hardware and software>
The word adding device of this embodiment is constructed by reading a predetermined program into the hardware as described above and executing it by the CPU. The functional configuration of each device constructed in this way will be described below.

  The input unit and output unit of the call state determination device 100 are communication devices such as a LAN card and a modem that are driven under the control of a CPU loaded with a predetermined program. The other calculation units 8 and the like are calculation units constructed by reading a predetermined program into the CPU and executing it. The storage unit functions as the auxiliary storage device.

Claims (8)

As input a voice call between the first speaker and the second speaker, a determining call status determination device whether a claim state is a state where one of the speaker angry,
A detection unit asking you to start time and end time of each utterance of each utterance of the first speaker and the second speaker,
Wherein to one utterance of the first speaker included in the telephone call voice and a pair utterances pair of one utterance of the second speaker, and extracts a set according to predetermined number the number of speech-to as a segment extraction unit When,
For each of the utterance pair, a calculation unit both the first speaker of the utterance and the utterance of the second speaker is calculated as an interactive feature amount time overlap,
Average value of the interactive features of all utterance pairs constituting the segment, variance of the interactive features of all utterance pairs constituting the segment, interactive of all utterance pairs constituting the segment A vectorization unit for obtaining a maximum value of feature values and a minimum value of the interactive feature values of all utterance pairs constituting the segment as elements of a feature vector;
By substituting each element of the feature vector previously stored more claims state a is call voice, and the voice of a prestored plurality of non claims state discriminant predetermined using A score calculation unit for obtaining a claim state score;
If the claim status score is greater than or equal to a predetermined first threshold, a determination unit that determines the segment as a claim status segment ;
Call state determination apparatus having. The call state determination device according to claim 1,
An output unit that outputs information indicating that the call is in a claim state if the number of the claim state segments or the ratio of the claim state segments in all the segments is equal to or greater than a predetermined second threshold;
A call state determination device characterized by further comprising : The call state determination device according to claim 1 or 2,
The calculation unit includes : the first speaker's speech time , the second speaker's speech time , the difference between the first speaker's speech time and the second speaker's speech time, the first speaker's speech time and the second speech Any one or more of the ratio of the speaker's utterance time, the silent time between the utterance of the first speaker and the utterance of the second speaker, the number of times of the first speaker's conversation, or the number of times of the second speaker's conversation Is further calculated as the interactive feature amount,
The vectorization unit comprises an average value for each interactive feature of all utterance pairs constituting the segment, a variance for each interactive feature of all utterance pairs constituting the segment, and the segment. Calling characterized in that a maximum value for each interactive feature value of all utterance pairs and a minimum value for each interactive feature value of all utterance pairs constituting the segment are obtained as elements of the feature vector. State determination device. As input a voice call between the first speaker and the second speaker, a one speaker angry or determining call status determination method is claims state is a state,
And the detection process asking you to start time and end time of each utterance of each utterance of the first speaker and the second speaker,
Extraction process of the one utterance of the first speaker included in the telephone call voice and a pair of one utterance of the second speaker and the utterance pair, extracts a set according to predetermined number the number of speech-to as a segment When,
For each utterance pair, a calculation process for calculating, as an interactive feature, a time during which both of the first speaker's utterance and the second speaker's utterance overlap .
Average value of the interactive features of all utterance pairs constituting the segment, variance of the interactive features of all utterance pairs constituting the segment, interactive of all utterance pairs constituting the segment A vectorization process for obtaining a maximum value of feature values and a minimum value of the interactive feature values of all utterance pairs constituting the segment as elements of a feature vector;
By substituting each element of the feature vector previously stored more claims state a is call voice, and the voice of a prestored plurality of non claims state discriminant predetermined using A score calculation process to obtain a claim status score;
If the claim status score is greater than or equal to a predetermined first threshold, a determination process for determining the segment as a claim status segment ;
Call state determination method having. The call state determination method according to claim 4,
An output process for outputting information indicating that the call is in a claim state if the number of the claim state segments or the ratio of the claim state segments in all the segments is equal to or greater than a predetermined second threshold;
The call state determination method further comprising: The call state determination method according to claim 4 or 5,
The calculation process includes the first speaker's speech time, the second speaker's speech time, the difference between the first speaker's speech time and the second speaker's speech time, the first speaker's speech time and the second speech. Any one or more of the ratio of the speaker's utterance time, the silent time between the utterance of the first speaker and the utterance of the second speaker, the number of times of the first speaker's conversation, or the number of times of the second speaker's conversation Is further calculated as the interactive feature amount,
  The vectorization process includes an average value for each interactive feature of all utterance pairs constituting the segment, a variance for each interactive feature of all utterance pairs constituting the segment, and the segment. The maximum value for each interactive feature value of all utterance pairs and the minimum value for each interactive feature value of all utterance pairs constituting the segment are obtained as elements of the feature vector.
A call state determination method characterized by the above. The program which operates a computer as a telephone call state determination apparatus in any one of Claims 1-3 . A computer-readable recording medium on which the program according to claim 7 is recorded.

Images