JP6676009B2 - Speaker determination device, speaker determination information generation method, and program - Google Patents

Description

  The present invention relates to a speaker determination device that determines a speaker from an audio signal, a speaker determination information generation method, and a program.

  2. Description of the Related Art In recent years, customer satisfaction of companies has been viewed as demand. A company's customer service that can directly hear requests is an important point of contact for customer satisfaction. In recent years, analysis of voices according to customer attributes (gender, age, etc.) in customer requests to collect marketing information from customers who have come to the counter and in the training of counter staff to improve customer satisfaction There are also requests to analyze the response quality of the contact person.

  In order to fulfill these requests, place a microphone at the counter to record voice, determine the section where the counter / customer is talking, perform voice recognition of the determined counter / customer, etc. There is a method of analyzing customer requests based on current requests and keywords, and collecting marketing information. For the attribute estimation of the customer, a conventional attribute estimation technology (technology for estimating gender, age, etc.) may be applied to the section where the customer is talking.

To determine the section where the customer is talking, speaker dialization technology can be applied. The speaker dialization technology is a technology for determining “when and who uttered” from a voice such as a dialogue. If the speaker dialization technology is applied, it is possible to determine a section in which the contact person is talking and a section in which the customer is talking. The outline of the conventional speaker dialization technology is shown below.
(I) Register a plurality of contact persons in advance as speaker contact voices in the speaker dialization system.
(Ii) Unnecessary noise / silence portions are removed from the recorded audio signal by an utterance interval detection technique such as VAD to detect an audio interval.
(Iii) For each detected voice section, the similarity with the registered contact person is obtained, and when the similarity is equal to or more than a predetermined threshold, the voice of the contact person and the similarity are less than the predetermined threshold. In this case, since the speaker is not registered, it is determined to be a customer's voice.
(Iv) For the voice section determined as the contact person in (iii), the contact person having the highest similarity is determined to be the speaker of the voice section.
(V) For the voice section determined to be a customer in (iii), speaker clustering is performed using the entire voice section to determine the voice section of each speaker. The speaker clustering is a technique in which a vector (i-vector or the like) indicating a speaker property is obtained for each utterance and plotted in a space, and the utterance of a vector at a short distance is determined as the same speaker.

Takahiro Oku, "Speaker Dialization for Dialogue Speech Recognition," NHK STRL R & D, No.147, September 2014, p.37-44 M. Fujimoto, K. Ishizuka, and T. Nakatani, “A Voice Activity Detection Based on the Adaptive Integration of Multiple Speech Features and a Signal Decision Scheme,” ICASSP2008, pp. 4441--4444, 2008. N. Dehak, P. Kenny, R. Dehak, P. Dumouchel and P. Ouellet, “Front-end factor analysis for speaker verification,” IEEE Transactions on Audio, Speech, And Language Processing, vol.19, pp.788- -798, 2011. D. Pelleg, A. Moore, “X-means: Extending K-means with Efficient Estimation of the Number of Clusters,” Proc. Of ICML, pp.724-734, 2000.

In the conversation at the counter, there are a plurality of customers and a plurality of tellers. In the conventional speaker dialization technology, the following erroneous determination of the speaker may occur due to biased phonemes or short utterances.
(A) An erroneous determination of the utterance of a contact person as an utterance of another contact person (b) An erroneous determination of a certain customer's utterance as an utterance of a contact person not in charge of the customer (c) Erroneous determination of an utterance as an utterance of another customer Therefore, an object of the present invention is to provide a speaker determination device that improves the determination accuracy of a speaker.

  The speaker determination device of the present invention is a speaker determination device that determines a speaker from a voice signal in which a conversation between a contact person and a customer is recorded. The speaker determination device of the present invention includes a similarity calculation unit, a primary speaker determination unit, a secondary speaker determination unit, and a speaker clustering unit.

  The similarity calculating unit calculates a similarity between a speaker feature amount of each divided voice section obtained by dividing a voice section of the voice signal into a predetermined time length and a speaker feature amount generated in advance for each contact person. I do.

  The speaker primary determination unit generates primary determination information indicating the speaker ID of each divided voice section from the similarity.

  The speaker secondary determination unit is configured to determine a speaker feature of a neighboring speaker who is the best-fit speaker in a predetermined number of divided voice sections before or after an arbitrary divided voice section, and a speaker characteristic of an arbitrary divided voice section. When the degree of similarity with the amount satisfies a predetermined condition, the secondary determination information is generated by using the speaker ID of the nearby speaker as the secondary determination information of an arbitrary divided voice section.

  The speaker clustering unit clusters the speaker feature amount of the divided voice section corresponding to the secondary determination information indicating that the customer is a customer, that is, a set of customer speaker feature amounts, generates a speaker ID of the customer, Generate judgment information.

  ADVANTAGE OF THE INVENTION According to the speaker determination apparatus of this invention, the determination accuracy of a speaker improves.

FIG. 2 is a block diagram illustrating a configuration of a speaker determination device according to the first embodiment. 5 is a flowchart illustrating the operation of the speaker determination device according to the first embodiment. The figure which shows the example which divides the audio | voice area detected from the audio | voice signal into predetermined time length. FIG. 3 is a block diagram illustrating a configuration of a speaker secondary determination unit according to the first embodiment. 5 is a flowchart illustrating the operation of a speaker secondary determination unit according to the first embodiment. FIG. 3 is a block diagram illustrating a configuration of a speaker clustering unit according to the first embodiment. 5 is a flowchart illustrating the operation of the speaker clustering unit according to the first embodiment. FIG. 9 is a block diagram illustrating a configuration of a speaker determination device according to a second embodiment. 9 is a flowchart illustrating a part of the operation of the speaker determination device according to the second embodiment. FIG. 9 is a block diagram illustrating a configuration of a speaker clustering unit according to the second embodiment. 9 is a flowchart illustrating the operation of the speaker clustering unit according to the second embodiment. FIG. 9 is a block diagram illustrating a configuration of a speaker determination device according to a third embodiment. 9 is a flowchart illustrating a part of the operation of the speaker determination device according to the third embodiment. 13 is a flowchart illustrating the operation of the speaker clustering unit according to the third embodiment.

  Hereinafter, embodiments of the present invention will be described in detail. Note that components having the same functions are given the same reference numerals, and redundant description is omitted.

  Hereinafter, the configuration of the speaker determination device of the first embodiment will be described with reference to FIG. The speaker determination device 1 of the present embodiment includes a voice division unit 11, a speaker characteristic amount extraction unit 12, a contact person speaker characteristic amount storage unit 13A, a similarity calculation unit 13, and a speaker primary determination unit. 14, a speaker secondary determination unit 15, a speaker clustering unit 16, and a speaker utterance section determination unit 17.

Note that the audio signal s (t) described as an input in FIG. 1 is an audio signal in which a conversation between a counterperson and a customer is recorded, and a sampling time t when the sampling frequency is f _s [Hz]. (T = 0, 1,..., S−1). The operation of each component will be described below with reference to FIG.

<Audio division unit 11>
Speech division unit 11 obtains the audio signal s (t), and outputs the time information T _k (S11). The operation of the audio division unit 11 will be further described with reference to FIG. As shown in the figure, the voice division unit 11 first detects a voice section from the voice signal s (t). Subsequently, the voice dividing unit 11 divides the voice section by a predetermined time length (σ seconds, for example, σ = 2). The voice section can be detected by a conventional method as in Non-Patent Document 2, for example.

Each of the voice sections divided into the predetermined time length σ is called a divided voice section. The time information T _k means the start time of the k-th (k = 1,..., K, K is an arbitrary natural number representing the total number of divided voice sections) divided voice sections.

<Speaker feature amount extraction unit 12>
The speaker characteristic amount extraction unit 12 obtains a vector indicating a speaker characteristic for each divided voice section (S12). Hereinafter, a vector indicating a speaker characteristic is referred to as a speaker characteristic amount. The speaker feature amount is extracted by a conventional technique using i-vector (Non-Patent Document 3) or the like. The speaker feature extracted from the k-th divided voice section starting from T _k is denoted as I _k . In order to make the description easy to understand, the description will hereinafter be focused on the k-th divided voice section as appropriate.

<Contact person speaker feature amount storage unit 13A>
The responsible person speaker features storage unit 13A, the speaker characteristic quantity J _m of advance for each responsible person is generated and is registered. m is the speaker ID of the person in charge of the window, and m = 1,..., M. M is the total number of registered contact persons, and is an arbitrary natural number.

<Similarity calculator 13>
The similarity calculation unit 13 calculates the similarity S _k = (S _k1 , S _k2 , S _k3 ,..., S _kM ) using the speaker feature amount I _k and the speaker feature amount J _m (S13). ). Note that S _km is a value of the speaker feature amount I _k of the k-th divided voice section and the speaker feature amount J _m of the m-th contact person registered in advance in the contact person speaker feature amount storage unit 13A. It is a similarity and uses a Euclidean distance or a cosine distance.

<Speaker primary determination unit 14>
Speaker primary determination unit 14 generates a primary determination information SP _k from the similarity S _k (S14). The primary determination information SP _k represents the speaker ID of the k-th divided voice section, SP _k ∈ {1,2,3,..., M,..., M} + {0}, and SP _k = 0 In this case, it indicates that the customer is not the contact person. SP _k is _obtained from the following equation using the threshold δ.

When the maximum value of the similarities S _km (m = 1,..., M) is equal to or greater than a predetermined threshold δ, the above equation is used to primarily determine the speaker ID corresponding to the maximum value. This represents information SP _k . On the other hand, when the maximum value of the similarities S _km (m = 1,..., M) is less than the predetermined threshold value δ, no contact person is applicable, indicating that the customer is a customer. the speaker ID = 0 indicating that the primary determination information SP _k. Note that when the cosine distance is used in the similarity calculation unit 13, it is preferable that the threshold value δ be a value near 0.5.

<Speaker secondary determination unit 15>
The determination by the speaker primary determination unit 14 may be an erroneous determination due to the stagnation of the voice or the bias of the phoneme. Therefore, the speaker secondary determination unit 15 corrects the primary determination information SP _k, to produce a secondary determination information SP _'k (S15).

  The configuration and operation of the secondary speaker determination unit 15 will be described with reference to FIGS. As illustrated in FIG. 4, the secondary speaker determination unit 15 includes a nearby speaker detection unit 151 and a secondary determination information generation unit 152.

  This applies most to a predetermined number (L) of divided voice sections before and after an arbitrary divided voice section of interest (here, the k-th divided voice section is continued) (the determination result was the largest). The speaker is called a neighbor speaker, and the speaker ID of the neighbor speaker is

And The nearby speaker detection unit 151 detects the speaker ID of the nearby speaker as follows (S151).

  That is, the above equation means that the speaker ID that is the most applicable (τ (0) = 1) in the kL-th to k + L-th divided voice sections is obtained as the speaker ID of the nearby speaker. . It should be noted that neighbor speakers may be obtained only for a predetermined number (L) of divided voice sections before the k-th divided voice section, or a predetermined number (L) of separated voice sections after the k-th divided voice section. May be obtained for only the divided voice section of. In this case, the above equation is modified as follows.

Next, the secondary determination information generation unit 152 generates the secondary determination information SP ′ _k using the similarity S _{km ^} between the speaker feature of the nearby speaker and the speaker feature of the k-th divided voice section. (S152). Specifically secondary decision information generating unit 152 generates a secondary determination information SP _'k as follows.

When the similarity S _{km ^} is equal to or more than a predetermined threshold δ ′, the above equation sets the speaker ID corresponding to the nearby speaker as the secondary determination information SP ′ _k , and If it becomes represents that to the speaker ID = 0 corresponding to the customer and the secondary determination information SP _'k. Note that the threshold δ ′ may be set to a value equal to or larger than δ used in step S14, and the determination may be strict. Also, near the speaker may be determined using a similarity S _k a constant window width of the front and rear. Specifically, neighboring speakers may be calculated as in the following equation.

<Speaker clustering unit 16>
In the following description, the speaker feature amount of the divided voice section corresponding to the secondary determination information (SP ′ _k = 0) indicating the customer is referred to as a customer speaker feature amount.

The speaker clustering unit 16 clusters a set of the customer speaker feature amounts, collectively divides the divided voice sections considered to be speaking by the same customer, and generates and assigns a speaker ID of the customer. The tertiary determination information SP ″ _k is generated (S16).

  The configuration and operation of the speaker clustering unit 16 will be described with reference to FIGS. As shown in FIG. 6, the speaker clustering unit 16 includes a customer speaker feature amount set generation unit 161, a feature cluster generation unit 162, and a tertiary determination information generation unit 163.

The customer speaker feature set generating unit 161 generates a set of customer speaker feature V = {I _k | SP ′ _k = 0} to be clustered (S161).

The feature cluster generation unit 162 clusters V using a general clustering technique such as k-means, and generates a set of feature clusters V ₁ , V ₂ ,..., V _H (H is the total number of feature clusters) ( S162).

The tertiary determination information generator 163 generates tertiary determination information SP ″ _k (S163). The tertiary determination information generation unit 163 generates SP ″ _k , for example, as follows.

_{I k ∈V h (h = 1,2} , ..., H) is a function representing the set of features clusters belonging to become I _k. The speaker ID of the customer was set to a negative value to distinguish it from the speaker ID of the contact person.

  In the k-means method of clustering in step S162, the number of customers needs to be given in advance as the number of feature clusters. If the number of customers is unknown, it is possible to use the x-means method using BIC as in Non-Patent Document 4.

<Speaker utterance section determination unit 17>
The speaker utterance section determination unit 17 integrates the speaker and the divided voice section as a final processing result. The speaker utterance section determination unit 17 integrates the speaker and the divided voice section corresponding to the speaker using, for example, the following formula, and generates and outputs speaker determination information _Dk (S17).
D _k = {SP '' _k , T _k , T _{k + 1} -1}
According to the speaker determination device 1 of the present embodiment, when determining the speaker of a certain divided voice section, not only the divided voice section but also a predetermined number of divided voice sections before or after the divided voice section are determined. Since the similarity is used, the accuracy of speaker determination is improved.

  Hereinafter, the configuration of the speaker determination device of the second embodiment will be described with reference to FIG. The speaker determination device 2 of the present embodiment includes a speaker clustering unit 26 different from that of the first embodiment. The components other than the speaker clustering unit 26 are the same as those in the first embodiment. Hereinafter, only the speaker clustering unit 26 will be described.

<Speaker clustering unit 26>
The speaker clustering unit 26 of this embodiment is characterized in that a customer-coupled speaker characteristic amount obtained by vector-connecting a customer speaker characteristic amount and corresponding time information is used.

As shown in FIG. 9, the speaker clustering unit 26 clusters a set of customer-coupled speaker features to group divided voice sections that are considered to be speaking by the same customer, and An ID is generated and assigned, and tertiary determination information SP ″ _k is generated (S26).

  The configuration and operation of the speaker clustering unit 26 will be described with reference to FIGS. As illustrated in FIG. 10, the speaker clustering unit 26 includes a combined speaker feature amount generation unit 261, a customer combined speaker feature amount set generation unit 262, a feature cluster generation unit 263, and a tertiary determination information generation unit 264. .

Binding speaker feature amount generating unit 261 combines the speaker characteristic quantity I _k and time information T _k corresponding thereto, binding speaker feature amount _{I 'k = (I k,} T k) to generate (S261 ). As binding methods, usually but a single vector (I _k, T _k), using the weight coefficient alpha, may be coupled in terms of was (I _k, αT _k). When clustering was performed by combining (I _k , αT _k ), it was found that the accuracy was improved as compared with the case where clustering was performed by combining (I _k , T _k ).

Hereinafter, the combined speaker feature amount of the divided voice section corresponding to the secondary determination information (SP ′ _k = 0) indicating the customer is referred to as a customer combined speaker feature amount.

The customer-joined speaker feature set generation unit 262 generates a set of customer-joined speaker feature V ′ = {I ′ _k | SP ′ _k = 0 (k = 1, 2, 3,..., K)}. It is generated (S262).

The feature cluster generation unit 263 clusters V ′ by a general clustering method such as k-means or the like, and generates a set of feature clusters V ′ ₁ , V ′ ₂ ,..., V ′ _H (S263).

The tertiary determination information generation unit 264 generates tertiary determination information SP ″ _k (S264). For example, SP ″ _k is generated as follows.

It is a function indicating a set of feature clusters to which I ′ _k belongs such that I ′ _k ∈V ′ _h (h = 1, 2,..., H).

  According to the speaker determination device 2 of the present embodiment, the time information is included in the customer-joined speaker feature quantity to be clustered, so that it is easy to determine a speaker who is speaking at a close time as the same speaker. And the accuracy of speaker determination is improved.

  Hereinafter, the configuration of the speaker determination device of the third embodiment will be described with reference to FIG. The speaker determination device 3 of the present embodiment includes a speaker clustering unit 36 different from the first and second embodiments. The components other than the speaker clustering unit 36 are the same as those in the first and second embodiments. Hereinafter, only the speaker clustering unit 36 will be described.

<Speaker clustering unit 36>
The configuration and operation of the speaker clustering unit 36 will be described with reference to FIGS. As shown in FIG. 12, the speaker clustering unit 36 includes a feature cluster updating unit 361 and a time cluster updating unit 362.

  As shown in FIG. 13, the speaker clustering unit 36 obtains each feature cluster by clustering a set of customer speaker features, obtains each time cluster based on the centroid of each feature cluster in the time direction, and Update each feature cluster based on the centroid in the feature direction of the time cluster, update each time cluster based on the centroid in the time direction of each updated feature cluster, and converge the update in the time direction with the centroid (S36). Hereinafter, step S36 will be described more specifically with reference to FIG.

First, the feature cluster updating unit 361 generates a set V = {I _k | SP ′ _k = 0 (k = 1, 2, 3,..., K)} of the customer speaker feature amounts to be clustered (K = 1, 2, 3,..., K)}. S361-1).

Next, the time cluster updating unit 362 generates a set of customer time information U = {T _k | SP ′ _k = 0 (k = 1, 2, 3,..., K)} (S362-1).

Next feature cluster updating unit 361, in the same manner as in Example 1, clustering V to the feature direction, set V _1, V ₂ feature cluster, ..., and generates a V _H, the characteristic directions of the characteristic cluster St. Lloyd c _1, c _2, ..., determine the c _H (S361-2).

Next time cluster update unit 362, the set V _1, V ₂ feature cluster, ..., using a V _H, St. time direction Lloyd t _1, t _2, ..., obtaining the t _H. Specifically, it is obtained as follows.

In addition, the time cluster update section 362, centroid t _1, t ₂ in the time direction of each feature cluster, ..., from t _H, boundary r _h in the time direction of the _speaker, asked the _{h + 1,} a set of time cluster U ₁ , U ₂ ,..., U _H are generated (S362-2). For example, the boundary of the speaker in the time direction can be obtained by giving the midpoint of the centroid in the time direction as described below, or by maximizing the number of matches between the belonging cluster in the feature cluster and the belonging cluster in the time cluster. .

(1) Method based on midpoint in time direction
(1-1) The boundary in the time direction is obtained as follows.
r _{h, h + 1} = (t _h + t _{h + 1} ) / 2
(1-2) The time cluster is obtained as follows.
U _h = {T _k | T _k ≧ r _{h, h + 1} ∧T _k <r _{h + 1, h + 2} }

(2) Method based on maximizing the number of matches (2-1) A boundary in the time direction is obtained as follows.
r _{h, h + 1} = arg _t max {| g _h (t) | + | g ' _h (t) |}
Where g _h (t) = {k | I _k ∈V _h ∧T _k ≤t (k = 1,2,…, K)}, g ' _h (t) = {k | I _k ∈V _{h + 1} {T _k > t (k = 1, 2,..., K)}, and g and g ′ are equal to or smaller than the time boundary t and larger than the number of voices belonging to the same cluster.
(2-2) A time cluster is obtained in the same manner as in (1-2).
U _h = {T _k | T _k ≧ r _{h, h + 1} ∧T _k <r _{h + 1, h + 2} }

Next, the feature cluster updating unit 361 obtains initial centroids c ₁ , c ₂ ,..., C _H in the feature direction from each time cluster, performs clustering in the feature direction, and performs centroid c ₁ , c ₂ , , C _H and the set V ₁ , V ₂ ,..., V _H of the feature clusters are updated (S361-3). The feature cluster updating unit 361 obtains an initial centroid in the feature direction as follows.

Next, the time cluster updating unit 362 obtains the speaker boundaries rh _{, h + 1} in the time direction from the centroids t ₁ , t ₂ ,..., T _{H in} the time direction of each feature cluster, and sets a set of time clusters. U _1, U _2, ..., updates the U _H (S362-3).

Step S361-3, S362-3 is converged centroid t _h time direction, it is repeatedly executed alternately until no change.

After that, the time cluster updating unit 362 reflects the set U ₁ , U ₂ ,..., U _H of the time clusters on the set V ₁ , V ₂ ,..., V _H of the feature clusters (V _h = {I _k | T _k ∈U _h (k = 1, 2,..., K)}), and generates the tertiary determination information SP ″ _k (S362-4). The time cluster updating unit 362 generates SP ″ _k , for example, as follows.

_{I k ∈V h (h = 1,2} , ..., H) is a function representing the set of features clusters belonging to become I _k.

  Similarly, in the alternate speaker clustering in the present embodiment, the number of speakers may be automatically determined by applying the x-means method. Further, by combining the second embodiment and the third embodiment, the clustering in the feature direction and the clustering in the time direction may be alternately performed on the set of the customer-coupled speaker feature amounts.

  According to the speaker determination device 3 of the present embodiment, erroneous determination can be reduced by alternately performing clustering in the feature direction and clustering in the time direction. For temporal clustering, we make the assumption that customers will not change frequently at the window, and since this assumption is often correct, by performing one-dimensional clustering in the temporal direction, The determination accuracy of the speaker is improved.

<Supplementary note>
The device of the present invention includes, for example, an input unit to which a keyboard or the like can be connected, an output unit to which a liquid crystal display or the like can be connected, and a communication device (for example, a communication cable) that can communicate outside the hardware entity as a single hardware entity , A communication unit, a CPU (which may include a Central Processing Unit, a cache memory and a register), a RAM or ROM as a memory, an external storage device as a hard disk, and an input unit, an output unit, and a communication unit thereof. , A CPU, a RAM, a ROM, and a bus connected so that data can be exchanged between the external storage devices. If necessary, the hardware entity may be provided with a device (drive) that can read and write a recording medium such as a CD-ROM. A physical entity provided with such hardware resources includes a general-purpose computer.

  The external storage device of the hardware entity stores a program necessary for realizing the above-described functions, data necessary for processing the program, and the like. It may be stored in a ROM that is a dedicated storage device). Data obtained by the processing of these programs is appropriately stored in a RAM, an external storage device, or the like.

  In the hardware entity, each program stored in the external storage device (or ROM or the like) and data necessary for processing of each program are read into the memory as needed, and interpreted and executed / processed by the CPU as appropriate. . As a result, the CPU realizes predetermined functions (the above-described components, such as components, means, etc.).

  The present invention is not limited to the embodiments described above, and can be appropriately modified without departing from the spirit of the present invention. In addition, the processes described in the above embodiments may be performed not only in chronological order according to the order described, but also in parallel or individually according to the processing capability of the device that executes the processes or as necessary. .

  As described above, when the processing function of the hardware entity (the device of the present invention) described in the above embodiment is implemented by a computer, the processing content of the function that the hardware entity should have is described by a program. By executing this program on a computer, the processing functions of the hardware entity are realized on the computer.

  A program describing this processing content can be recorded on a computer-readable recording medium. As a computer-readable recording medium, for example, any recording medium such as a magnetic recording device, an optical disk, a magneto-optical recording medium, and a semiconductor memory may be used. Specifically, for example, a hard disk device, a flexible disk, a magnetic tape, or the like is used as a magnetic recording device, and a DVD (Digital Versatile Disc), a DVD-RAM (Random Access Memory), or a CD-ROM (Compact Disc Read Only) is used as an optical disk. Memory), CD-R (Recordable) / RW (ReWritable), etc., magneto-optical recording media, MO (Magneto-Optical disc), semiconductor memory, EEP-ROM (Electronically Erasable and Programmable-Read Only Memory), etc. Can be used.

  The distribution of the program is performed by, for example, selling, transferring, lending, or the like, a portable recording medium such as a DVD or a CD-ROM on which the program is recorded. Further, the program may be stored in a storage device of a server computer, and the program may be distributed by 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 recorded on a portable recording medium or a program transferred from a server computer in its own storage device. Then, when executing the processing, the computer reads the program stored in its own recording medium and executes the processing according to the read program. Further, as another execution form of the program, the computer may directly read the program from the portable recording medium and execute processing according to the program, and further, the program may be transferred from the server computer to the computer. Each time, the processing according to the received program may be sequentially executed. A configuration in which the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes a processing function only by executing an instruction and acquiring a result without transferring a program from the server computer to the computer. It may be. It should be noted that the program in the present embodiment includes information used for processing by the computer and which is similar to the program (data that is not a direct command to the computer but has characteristics that define the processing of the computer).

  Further, in this embodiment, a hardware entity is configured by executing a predetermined program on a computer. However, at least a part of the processing contents may be realized by hardware.

  Each step described in the specification and the claims corresponds to each step of a method for generating various information. The various types of information referred to here correspond to programs and the like (programs and other information used for processing by a computer and conform to the programs) prescribed in Article 2, paragraph 4 of the Patent Act. The information in (1) corresponds to the provisions of Article 2, Paragraph 3, Item 1 of the Patent Act. Therefore, it goes without saying that the method of generating various kinds of information described in the description and the claims corresponds to the method of producing a product specified in Article 2, Paragraph 3, Item 3 of the Patent Act. Absent.

Claims (8)

A speaker determination device that determines a speaker from an audio signal in which a conversation between a contact person and a customer is recorded,
A similarity calculation for calculating a similarity between a speaker feature of each divided voice section obtained by dividing a voice section of the voice signal into a predetermined time length and a speaker feature generated in advance for each contact person. Department and
A speaker primary determination unit that generates primary determination information indicating a speaker ID of each divided voice section from the similarity;
The similarity between the speaker feature of the neighboring speaker who is the most applicable speaker in the predetermined number of divided speech sections before or after the arbitrary divided speech section and the speaker feature of the arbitrary divided speech section is predetermined. When the condition of is satisfied, by making the speaker ID of the nearby speaker the secondary determination information of the arbitrary divided voice section, a speaker secondary determination unit that generates secondary determination information,
Clustering a set of speaker features of the divided voice section corresponding to the secondary determination information indicating the customer, that is, a set of customer speaker features, generates a speaker ID of the customer, and generates tertiary determination information A speaker determination device including a speaker clustering unit for performing the above. The speaker determination device according to claim 1,
The speaker clustering unit includes:
A speaker judging device for clustering a set of customer combined speaker features obtained by combining the customer speaker features and time information corresponding thereto. The speaker determination device according to claim 1,
The speaker clustering unit includes:
Clustering the set of customer speaker feature amounts to obtain each feature cluster, obtaining each time cluster based on the centroid in the time direction of each feature cluster, and based on the centroid in the feature direction of each time cluster A speaker determination device that updates each of the feature clusters, updates each time cluster based on the updated centroid of each feature cluster, and repeatedly executes the update until the centroid in the time direction converges. . The speaker determination device according to claim 2,
The speaker clustering unit includes:
The set of customer-coupled speaker features is clustered to determine each feature cluster, each time cluster is determined based on the centroid in the time direction of each feature cluster, and based on the centroid in the feature direction of each time cluster. Speaker determination that updates each of the feature clusters, updates each time cluster based on the updated centroid of each feature cluster, and repeatedly executes the update until the centroid in the time direction converges. apparatus. A speaker determination information generation method for determining a speaker from a voice signal in which a conversation between a contact person and a customer is recorded,
A similarity calculation for calculating a similarity between a speaker feature of each divided voice section obtained by dividing a voice section of the voice signal into a predetermined time length and a speaker feature generated in advance for each contact person. Steps and
A speaker primary determination step of generating primary determination information representing a speaker ID of each divided voice section from the similarity;
The similarity between the speaker feature of the neighboring speaker who is the most applicable speaker in the predetermined number of divided speech sections before or after the arbitrary divided speech section and the speaker feature of the arbitrary divided speech section is predetermined. When the condition of is satisfied, by making the speaker ID of the nearby speaker the secondary determination information of the arbitrary divided voice section, a speaker secondary determination step of generating secondary determination information;
Clustering a set of speaker features of the divided voice section corresponding to the secondary determination information indicating the customer, that is, a set of customer speaker features, generates a speaker ID of the customer, and generates tertiary determination information Speaker clustering step
A speaker determination information generation method executed by the speaker determination device. A method for generating speaker determination information according to claim 5, wherein
The speaker clustering step includes:
A speaker determination information generating method for clustering a set of customer combined speaker features obtained by combining the customer speaker features and time information corresponding thereto. A method for generating speaker determination information according to claim 5, wherein
The speaker clustering step includes:
Clustering the set of customer speaker feature amounts to obtain each feature cluster, obtaining each time cluster based on the centroid in the time direction of each feature cluster, and based on the centroid in the feature direction of each time cluster Speaker determination information that updates each feature cluster, updates each time cluster based on the updated centroid of each feature cluster, and repeatedly executes the update until the centroid in the time direction converges Generation method.   A program for causing a computer to function as the speaker determination device according to any one of claims 1 to 4.

Images