JP2020190669A - Speaker identification device, speaker identification method, and speaker identification program - Google Patents

Abstract

To identify a speaker more easily without need for prior work.SOLUTION: A speaker identification device 50 includes a speakers identification device 50 which identifies a speaker by using a plurality of voice data acquired by a plurality of microphones provided corresponding to a plurality of target persons. The speakers identification device 50 includes: a representative value acquisition unit 52 which acquires a representative value indicating a characteristic of loudness for each voice data for each predetermined first period; a comparison unit 54 which compares two representative values acquired from two different voice data for each first period and obtains a comparison result for each first period; and a specification unit 55 which identifies the speakers in a second period, using a plurality of comparison results acquired in the second period longer than the first period.SELECTED DRAWING: Figure 4

Description

The present invention relates to a speaker identification device, a speaker identification method, and a speaker identification program.

In recent years, with the development of voice recognition technology, a system has been developed that automatically converts the voice of a speaker in an interview or a conference into character data and automatically creates minutes. When performing such character recognition, it is necessary to identify who spoke the voice data acquired by the microphone.

As a method for identifying a speaker, for example, a method disclosed in Patent Document 1 is known. In Patent Document 1, the feature amount of the voice emitted by the speaker is stored in advance, and the feature amount of the voice data acquired by the microphone is compared with the feature amount of the speaker stored in advance. , A method of identifying the speaker is disclosed.

Japanese Unexamined Patent Publication No. 2004-145161

However, in the invention disclosed in Patent Document 1, it is necessary to store the feature amount of the speaker in advance, which is not convenient. In addition, since it is not possible to identify a speaker whose feature amount is not stored, it is not versatile.
In addition, when the voices of the speakers are similar, it is difficult to distinguish them by comparing the features, and the accuracy is low.

The present invention has been made in view of such circumstances, and is a speaker identification device, a speaker identification method, and a speaker identification that do not require prior work and can more easily identify a speaker. The purpose is to provide a program.

A first aspect of the present invention is a speaker identification device that identifies a speaker by using a plurality of voice data acquired by a plurality of microphones provided corresponding to a plurality of subjects, and the voice is described above. For each data, a representative value acquisition unit that acquires a representative value indicating a characteristic regarding loudness for each predetermined first period, and two representative values acquired from two different voice data are said to be the first. A speaker in the second period using a comparison unit that compares each period and obtains a comparison result for each first period and a plurality of the comparison results obtained in a second period longer than the first period. It is a speaker identification device including a specific unit for specifying the above.

According to the speaker identification device, the representative value acquisition unit acquires the representative value of each voice data in each first period, and the comparison unit acquires two different voice data in each first period. The representative values are compared and the comparison result is obtained. Then, the specific unit identifies the speaker in the second period by using a plurality of comparison results obtained in the second period, which is longer than the first period. In this way, instead of using all the sampled voice data, it is decided to acquire the representative value in each first period, so that the noise contained in the voice data can be reduced and the amount of data to be processed. Can be reduced.
Further, according to this aspect, the representative values acquired from different voice data are compared for each first period, and the speaker is specified by using a plurality of comparison results in the second period longer than the first period. .. Since the voice data contains noise, for example, if the speaker is specified for each first period, the speaker may be frequently switched due to the influence of the noise. On the other hand, according to this aspect, since the speaker in the second period is identified by using a plurality of comparison results obtained in the second period longer than the first period, the speakers are frequently switched. Can be suppressed, and the accuracy of identifying the speaker can be improved.

A second aspect of the present invention is a speaker identification method for identifying a speaker by using a plurality of voice data acquired by a plurality of microphones provided corresponding to a plurality of subjects, and the voice is described above. For each data, a representative value acquisition step of acquiring a representative value indicating a characteristic regarding loudness for each predetermined first period, and the first of the two representative values acquired from two different voice data. A speaker in the second period using a comparison step of comparing each period and obtaining a comparison result for each first period and a plurality of the comparison results obtained in a second period longer than the first period. It is a speaker identification method in which a computer executes a specific process for specifying.

A third aspect of the present invention is a speaker identification program for identifying a speaker by using a plurality of voice data acquired by a plurality of microphones provided corresponding to a plurality of subjects. For each of the voice data, a representative value acquisition process of acquiring a representative value indicating a characteristic regarding the loudness of the sound for each predetermined first period, and two representative values acquired from the two different voice data are said to be the same. In the second period, the comparison process of comparing each first period and obtaining the comparison result for each first period and the plurality of the comparison results obtained in the second period longer than the first period are used. It is a speaker identification program for causing a computer to execute a specific process for identifying a speaker.

According to the present invention, there is an effect that the speaker can be identified more easily without the need for prior work.

It is a figure which showed the structure which assumed the case of two subjects in the speaker identification system which concerns on one Embodiment of this invention. It is a figure for demonstrating the correspondence relationship between a microphone and a subject person in the case of assuming the case of two subjects in the speaker identification system which concerns on one Embodiment of this invention. It is a schematic block diagram which showed an example of the hardware configuration of the speaker identification apparatus which concerns on one Embodiment of this invention. It is a functional block diagram which shows typically the function which the speaker identification apparatus which concerns on one Embodiment of this invention has. It is a figure which showed typically an example of the waveform of the voice data acquired by the microphone which concerns on one Embodiment of this invention. It is a figure for demonstrating the process realized by the representative value acquisition part which concerns on one Embodiment of this invention. It is a figure for demonstrating the process realized by the specific part which concerns on one Embodiment of this invention. It is a flowchart which showed an example of the procedure of the speaker identification method which concerns on one Embodiment of this invention. It is a figure which showed the structure which assumed the case of three subjects in the speaker identification system which concerns on one Embodiment of this invention. It is a figure for demonstrating the correspondence relationship between a microphone and a subject person in the case of assuming the case of three subjects in the speaker identification system which concerns on one Embodiment of this invention. It is a figure for demonstrating the process realized by the specific part which concerns on one Embodiment of this invention.

Hereinafter, an embodiment of a speaker identification device, a speaker identification method, and a speaker identification program according to the present invention will be described with reference to the drawings. Hereinafter, for convenience of explanation, first, a case where there are two persons will be described as an example, and then, a case where there are three persons will be illustrated and described.

FIG. 1 is a diagram schematically showing a configuration of a speaker identification system 1 according to an embodiment of the present invention. As shown in FIG. 1, the speaker identification system 1 according to the present embodiment is based on a plurality of microphones (hereinafter, simply referred to as “microphones”) 10 and 20 and voice data acquired by the microphones 10 and 20. It is provided with a speaker identification device 50 for identifying a speaker.

The microphones 10 and 20 are provided corresponding to the persons A and B, for example, as shown in FIG. In FIG. 2, a microphone 10 is provided corresponding to the person A, and a microphone 20 is provided corresponding to the person B. The microphone 10 is arranged at a position where the distance La10 from the person A and the distance Lb10 from the person B are different. Similarly, the microphone 20 is arranged at different positions from the distance La20 from the person A and the distance Lb20 from the person B. The microphones 10 and 20 may be attached to the persons A and B.

In FIG. 2, the microphone 10 is arranged at a position closer to the person A than the person B, and the microphone 20 is arranged at a position closer to the person B than the person A. That is, the relationships shown in the following equations (1) and (2) are established for the distances from the persons A and B to the microphones 10 and 20.

La10 <Lb10 (1)
Lb20 <La20 (2)

FIG. 3 is a schematic configuration diagram showing an example of the hardware configuration of the speaker identification device 50. As shown in FIG. 3, the speaker identification device 50 is, for example, a CPU 11, an auxiliary storage device 12 for storing a program executed by the CPU 11, a data referenced by the program, and the like, and as a work area at the time of executing each program. It includes a functioning main storage device 13, an external device (for example, microphones 10, 20, etc.), at least one communication interface 14 for connecting to a network, and the like. Further, the speaker identification device 50 may include, for example, an input unit 15 that functions as a user interface for a keyboard, a mouse, a pointing device, and the like, a display unit 16 for a liquid crystal display, and the like.
Each of these parts is connected via, for example, a bus 18. Examples of the auxiliary storage device 12 include magnetic disks such as HDDs (Hard Disk Drives), magneto-optical disks, and semiconductor memories such as SSDs (Solid State Drives).

As an example, a series of processes for realizing various functions described later are stored in the auxiliary storage device 12 in the form of a program (for example, a speaker identification program), and the CPU 11 reads this program into the main storage device 13. Therefore, various functions are realized by executing information processing and arithmetic processing. The program is pre-installed in the auxiliary storage device 12, provided in a state of being stored in another computer-readable storage medium, or distributed via a wired or wireless communication means. Etc. may be applied. Computer-readable storage media include magnetic disks, magneto-optical disks, CD-ROMs, DVD-ROMs, semiconductor memories, and the like.

FIG. 4 is a functional block diagram schematically showing the functions of the speaker identification device 50 according to the present embodiment. As shown in FIG. 4, the speaker identification device 50 includes, for example, a voice data storage unit 51, a representative value acquisition unit 52, a standardization processing unit 53, a comparison unit 54, and a identification unit 55.

The voice data storage unit 51 stores the voice data D1 and D2 acquired by the microphones 10 and 20.
For example, when the microphones 10 and 20 and the speaker identification device 50 are connected via a wired line or a wireless line, the speaker identification device 50 transmits the voice data output from the microphones 10 and 20 to the wired line or wireless. It is acquired via a line and stored in the voice data storage unit 51. Further, the speaker identification device 50 and the microphones 10 and 20 may be indirectly connected via another terminal or the like.

For example, the voice data acquired by the microphones 10 and 20 is temporarily stored in the storage medium connected to the microphones 10 and 20, and then the voice data stored in the storage medium is stored in the communication medium at a predetermined timing. It may be transferred to the speaker identification device 50 via the speaker identification device 50. In this case, the microphones 10 and 20 and the storage medium may be provided separately, or may be a device in which the voice acquisition function and the storage medium are integrated, such as a voice recorder. Further, the speaker identification device 50 may exist on the cloud.

Further, the above example is an example, and a wide variety of known methods can be adopted as a method for the speaker identification device 50 to acquire the voice data acquired by the microphones 10 and 20.
As described above, the voice data storage unit 51 stores the voice data D1 acquired by the microphone 10 and the voice data D2 acquired by the microphone 20.
FIG. 5 is a diagram schematically showing an example of waveforms of the voice data D1 acquired by the microphone 10 and the voice data D2 acquired by the microphone 20.

The representative value acquisition unit 52 acquires a representative value indicating a feature regarding the loudness of each of the voice data D1 and D2 stored in the voice data storage unit 51 for each predetermined first period. FIG. 6 is an enlarged view of the time axis of the voice data D1 shown in FIG. As shown in FIG. 6, the voice data D1 is data including at least the voice intensity associated with the time when the voice was acquired.

The representative value acquisition unit 52 acquires the representative value Ui from the voice data D1 for each predetermined first period TWi (i = 1 to n), for example. The first period TWi is set to, for example, a period longer than one cycle of the sound wave waveform based on the voice data. As an example, when the sampling frequency of voice data is set to 44.1 kHz, the first period TWi is set within the range of 1 ms or more and 30 ms or less, preferably 5 ms or more and 20 ms or less.

By setting the first period TWi in such a range, it is possible to acquire an appropriate representative value. The number of sampling data included in the first period TWi changes according to the pitch of the speaker's voice. If the voice is high, the number of sampling data included in the first period TWi increases. Therefore, for example, when the speaker is identified after the fact, the first is such that the number of data included in the first period TWi is substantially equal according to the frequencies of the acquired voice data D1 and D2. The period TWi may be changed dynamically.

The representative value is, for example, the maximum value in each first period TWi. Further, the representative value may be the maximum amplitude value or the difference between the minimum intensity value and the maximum intensity value. Further, for example, among the absolute values of each voice intensity, a value equal to or higher than a predetermined threshold value may be extracted, and the average value of the extracted values may be used as a representative value. As described above, the representative value may be a feature amount indicating the characteristic of loudness in each first period TWi, and a specific determination method can be appropriately set. In the following, for convenience of explanation, a case where the maximum value is adopted as a representative value will be described as an example.

As described above, the representative value acquisition unit 52 acquires the maximum value Ui in each first period TWi as the representative value in the voice data D1. Further, the representative value acquisition unit 52 similarly acquires the maximum value Vi in each first period TWi (i = 1 to n) as the representative value in the voice data D2.

The standardization processing unit 53 standardizes the representative values Ui and Vi acquired by the representative value acquisition unit 52. For example, if the loudness of the speaker's voice is different, the intensity of the sound waves acquired by the microphones 10 and 20 will be imbalanced. The standardization processing unit 53 performs standardization processing in order to eliminate the imbalance of voice intensity due to the loudness of the speaker's voice.

For example, the standardization process is performed as follows.
First, the standardization processing unit 53 calculates the average value X of the representative values Ui (i = 1 to n) of the audio data D1 and the average value Y of the representative values Vi (i = 1 to n) of the audio data D2.

In the equations (3) and (4), n is the number of representative values in the voice data D1 and D2, in other words, the number of TWi in the first period.

Subsequently, the standardization processing unit 53 uses the average values X and Y of the representative values to standardize the representative values Ui (i = 1 to n) and Vi (i = 1 to n) as standardized representative values Ui_st ( i = 1 to n) and Vi_st (i = 1 to n) are calculated. The calculation formula for standardization processing is as follows.

The comparison unit 54 compares the two standardized representative values acquired from the different voice data D1 and D2 for each first period TWi, and obtains a comparison result for each first period TWi.
Specifically, the comparison unit 54 calculates the difference di (i = 1 to n) between the standardized representative values Ui_st and Vi_st obtained in the same first period TWi (i = 1 to n). The calculation formula is represented by the following formula (7).

di = Ui_st-Vi_st (7)

The identification unit 55 identifies the speaker by using a plurality of comparison result dis obtained in the second period TYj for each second period TYj (j = 1 to m) longer than the first period TWi.
For example, the specific unit 55 calculates the difference average value dj_ave, which is the average value of the difference di, for each second period TYj.

For example, the second period TYj is set to a period five times or more that of the first period TWi. In the present embodiment, as an example, the second period TYj is set to 250 ms. For example, when the first period TWi is set to 10 ms, there are 25 difference dis in each second period TYj, so the average value dj_ave of the 25 difference dis is calculated.
Here, the above 250 ms is an example of the second period TYj, and may be set in a range in which the speaker can be specified with an accuracy that satisfies the required accuracy, for example. Specifically, the second period TYj is set to a period five times or more the first period.

Subsequently, the specific unit 55 determines whether the difference average value dj_ave is a positive value (that is, dj_ave> 0) or a negative value (dj_ave <0). If the absolute value of the difference average value is equal to or less than a predetermined threshold value, it is determined to be zero. That is, in this case, it is determined that no one is speaking or that it is noise. Then, as shown in FIG. 7, when the same positive / negative sign or zero continues for a predetermined number of times (6 times in FIG. 7) or for a predetermined period or more, the person corresponding to the sign is identified as the speaker.

For example, as shown in FIG. 7, the positive and negative signs in the second period TYj are positive 6 times, negative 2 times, 0 (zero) 2 times, negative 1 time, positive 7 times, negative 1 It is assumed that the number of times, the positive number is 2 times, the negative number is 8 times, the positive number is 1 time, and 0 (zero) is 6 times. In such a case, the speaker is identified as person A until a positive appears six times and then a negative or zero appears six times in a row. Then, for example, when the negative sign continues 6 times, it is determined that the speaker has switched from the person A to the person B, and the speaker is identified as the person B from the time when the negative sign first appears. Will be done. Then, the speaker is identified as person B unless a positive sign or zero continues six or more times. Further, for example, when zero continues 6 times, it is determined that no speaker has spoken.
As a result, for example, speaker A, speaker B, and no utterance are specified as illustrated in FIG. 7.
In this way, when the positive, negative, or zero signs are continuous for a predetermined number of times or for a predetermined period or more, it is determined that the speaker has been switched or none of the speakers is speaking, and thus the influence of noise It is possible to suppress the erroneous discrimination of the speaker due to.

Next, the speaker identification method according to the present embodiment described above will be briefly described with reference to FIG.
First, the voice data D1 and D2 spoken by the persons A and B are acquired by the microphones 10 and 20, and stored in the voice data storage unit 51 of the speaker identification device 50 (SA1).
Subsequently, in each of the audio data D1 and D2, the representative values Ui and Vi for each TWi in the first period are acquired (SA2), and further, the standardized representative values Ui_st and Vi_st are calculated by standardizing the representative values Ui and Vi. (SA3).
Subsequently, the difference di of the standardized representative value for each first period TWi is calculated (SA4), and the difference average value dj_ave is calculated for each second period TYj longer than the first period TWi (SA5).

Subsequently, the positive / negative of the difference mean value dj_ave is determined for each second period TYj (SA6), and depending on the continuous situation of positive / negative, for example, when the same code is consecutive a predetermined number of times (for example, 6), the positive / negative It is specified that the speaker is speaking according to the above (SA7). For example, in the present embodiment, as shown in the above equation (7), when the difference di is calculated, the standardized representative value Vi_st of the voice data D2 is subtracted from the standardized representative value Ui_st of the voice data D1. When the code is positive, it is determined that the person A corresponding to the voice data D1 is speaking, and when the code is negative, it is determined that the person B corresponding to the voice data D2 is speaking.

As described above, according to the utterance identification device, the utterance identification method, and the utterance identification program according to the present embodiment, the representative value acquisition unit 52 sets the representative values Ui and Vi in the voice data D1 and D2 in the first period. It is acquired for each TWi, and the representative values Ui and Vi are standardized by the standardization processing unit 53. Then, the comparison unit 54 compares the two standardized representative values Ui_st and Vi_st acquired from the two different voice data D1 and D2 for each TWi in the first period, and obtains a comparison result. Then, the identification unit 55 identifies the speaker in the second period TYj by using a plurality of comparison results obtained in the second period TYj longer than the first period TWi. In this way, instead of using all the sampled audio data D1 and D2, it was decided to acquire the representative values Ui and Vi in each first period TWi and use them to identify the speaker. Therefore, the audio data D1 and The noise contained in D2 can be reduced, and the amount of data to be processed can be reduced.

Further, since the representative values Ui and Vi are standardized, it is possible to suppress the imbalance of voice intensity between the voice data.
Further, the speaker in the second period TYj using the difference average value dj_ave obtained by averaging the difference di of the two standardized representative values Ui_st and Vi_st obtained for each TWi in the first period by the comparison unit 54 in the second period TYj. Therefore, it is possible to suppress frequent switching of the speaker due to the influence of noise contained in the voice data D1 and D2, and it is possible to improve the identification accuracy of the speaker.

Further, according to the present embodiment, the specific unit 55 determines whether the difference average value dj_ave is positive or negative, and when the same code is consecutive for a predetermined number of times or for a predetermined period, the speaker corresponding to the code is speaking. Since the determination is made, it is possible to further suppress frequent switching of the speaker. That is, in the present embodiment, the second period TYj is set to 250 ms, but it is not very realistic that the speaker can be switched in 250 ms units. According to the present embodiment, when the same positive and negative signs are consecutive for a predetermined number of times or for a predetermined period or more in the difference average value dj_ave obtained for each second period TYj, it is determined that the person corresponding to the sign is speaking. Therefore, the accuracy of identifying the speaker can be further improved. Note that this specific method is an example. For example, although the accuracy of identifying the speaker is reduced, it is possible to specify the speaker for each TYj in the second period.

In the above embodiment, the standardized representative values Ui_st and Vi_st are used to obtain the comparison result, but the standardization is not essential, and the comparison unit 54 uses the non-standardized representative values Ui and Vi to obtain the difference di. May be obtained.

Next, the utterance identification device, the utterance identification method, and the utterance identification program according to the embodiment of the present invention will be described by way of example when there are three persons. Regarding the following explanations, the points common to the above-mentioned case of two speakers will be omitted, and the differences will be mainly explained.

FIG. 9 is a diagram schematically showing the configuration of the speaker identification system 1'when there are three speakers. As shown in FIG. 9, in the speaker identification system 1', in addition to the configuration shown in FIG. 1, a microphone (microphone) 30 corresponding to the person C is added, and the voice data D3 acquired by the microphone 30 speaks. It is designed to be input to the person identification device 50.

FIG. 10 is a diagram showing the arrangement of the persons A to C and the microphones 10, 20, and 30. In FIG. 10, a microphone 30 is provided corresponding to the person C. The microphone 10 is arranged at different positions from the distance LA10 from the person A, the distance Lb10 from the person B, and the distance Lc10 from the person C. Similarly, the microphone 20 is arranged at different positions from the distance La20 from the person A, the distance Lb20 from the person B, and the distance Lc20 from the person C. Further, the microphone 30 is arranged at a position where the distance La30 from the person A, the distance Lb30 from the person B, and the distance Lc30 from the person C are different.
The microphones 10, 20, and 30 may be attached to the persons A to C, respectively.

In FIG. 10, the microphone 10 is arranged at a position closer to the person A than any of the persons B and C, the microphone 20 is arranged at a position closer to the person B than any of the persons A and C, and the microphone 30 is arranged. It is arranged closer to the person C than any of the persons A and B. That is, the relationships shown in the following equations (8) to (13) are established for the distances from the persons A, B, and C to the microphones 10, 20, and 30.

La10 <Lb10 (8)
La10 <Lc10 (9)
Lb20 <La20 (10)
Lb20 <Lc20 (11)
Lc30 <La30 (12)
Lc30 <Lb30 (13)

Next, a method of identifying the speaker based on the voice data D1 to D3 acquired by the microphones 10, 20, and 30 will be described. The functional block included in the speaker identification device 50 is the same as the function shown in FIG. 4, but the point that each unit performs the same processing for the voice data D3 and the speaker identification method in the specific unit 55 are slightly different. ..

First, the voice data storage unit 51 stores the voice data D1 acquired by the microphone 10, the voice data D2 acquired by the microphone 20, and the voice data D3 acquired by the microphone 30.

The representative value acquisition unit 52 acquires the representative value for each voice data D1 to D3 for each TWi in the first period by the same method as in the case of the two speakers described above. For example, the representative value acquisition unit 52 acquires the maximum value Ui in each first period TWi (i = 1 to n, the same applies hereinafter) as a representative value for the voice data D1, and in the voice data D2, each first period TWi. The maximum value Wi in each first period TWi is acquired as a representative value in the voice data D3.

The standardization processing unit 53 standardizes the representative values Ui, Vi, and Wi acquired by the representative value acquisition unit 52, and acquires the standardized representative values Ui_st_uv, Ui_st_uu, Vi_st_uv, Vi_st_vw, Wi_st_uu, and Wi_st_vw, respectively. For example, the standardization process is performed as follows.
First, the standardization processing unit 53 sets the average value X of the representative value Ui (i = 1 to n) of the audio data D1, the average value Y of the representative value Vi (i = 1 to n) of the audio data D2, and the audio data D3. The average value Z of the representative value Wi (i = 1 to n) is calculated.

In the equations (14) to (16), n is the number of representative values in the voice data D1, D2, and D3, in other words, the number of the first period TWi.

Subsequently, the standardization processing unit 53 uses the average values X, Y, and Z of the representative values to combine the voice data D1 and the voice data D2, the voice data D1 and the voice data D3, and the voice data D2. A standardized representative value is calculated for each combination of the voice data and the voice data D3.

For example, the data standardization in the combination of the voice data D1 and the voice data D2 is as follows.

For example, the data standardization in the combination of the voice data D1 and the voice data D3 is as follows.

For example, the data standardization in the combination of the voice data D2 and the voice data D3 is as follows.

Next, the comparison unit 54 determines the standardized representative value between the audio data D1 and the audio data D2 among the standardized representative values acquired from the different audio data D1, D2, and D3 for each first period TWi. The difference di_uv, the difference di_uw of the standardized representative value between the voice data D1 and the voice data D3, and the difference di_vw of the standardized representative value between the voice data D2 and the voice data D3 are calculated, respectively. The calculation formula is as follows.

di_uv = Ui_st_uv-Vi_st_uv (23)
di_uuw = Ui_st_uuw-Wi_st_uuw (24)
di_vw = Vi_st_vw-Wi_st_vw (25)

The identification unit 55 identifies the speaker by using a plurality of comparison results obtained within the second period TYj for each second period TYj longer than the first period TWi.
For example, the specific unit 55 calculates the difference average values dj_uv_ave, dj_uu_ave, and dj_vw_ave, which are the average values of the differences di_uv, di_uu, and di_vw for each second period TYj.

Subsequently, the specific unit 55 determines whether the difference average values dj_uv_ave, dj_uw_ave, and dj_vw_ave are positive or negative values. If the absolute value of the difference average value is equal to or less than a predetermined threshold value, it is determined to be zero. That is, in this case, it is determined that a speaker other than the speaker corresponding to the voice data used to calculate the difference is speaking.

Even when there are three persons, as shown in FIG. 11, when the same code occurs continuously for a predetermined number or more (or continuously for a predetermined period or more), the person corresponding to the code is spoken. Identify as a person. Further, in any combination (dj_uv_ave, dj_uw_ave, dj_vw_ave), when zeros occur continuously for a predetermined number or more, it is determined that no person is speaking. Thereby, as shown in FIG. 11, it is specified whether the speaker is person A, person B, person C, or none of the persons speaks (no utterance). Can be done.

As described above, when there are three or more persons, the comparison unit 54 generates a combination of two different voice data, and calculates the difference between the standardized representative values of the voice data in the generated combination. Then, the specific unit 55 calculates the difference average value by averaging the difference of the calculated standardized representative values for each TYj in the second period in each combination of audio data, and positive, negative, or zero of the difference average value. Identify the speaker by comprehensively judging the situation.

As described above, even when there are three or more persons, a plurality of combinations consisting of two different voice data are created, the difference average value is calculated for each combination, and the code of the difference average value is calculated. It is possible to identify the speaker by comprehensively judging the situation of.

Although the present invention has been described above using the embodiments, the technical scope of the present invention is not limited to the scope described in the above embodiments. Various changes or improvements can be made to the above embodiments without departing from the gist of the invention, and the modified or improved forms are also included in the technical scope of the present invention. Moreover, you may combine the said embodiment as appropriate.
Further, the flow of the speaker identification process described in the above embodiment is also an example, and unnecessary steps may be deleted, new steps may be added, or the processing order may be changed within a range not deviating from the gist of the present invention. You may.

1, 1': Speaker identification system 10: Microphone (microphone)
11: CPU
12: Auxiliary storage device 13: Main storage device 14: Communication interface 15: Input unit 16: Display unit 18: Bus 20: Microphone (microphone)
30: Microphone (microphone)
50: Speaker identification device 51: Voice data storage unit 52: Representative value acquisition unit 53: Standardization processing unit 54: Comparison unit 55: Identification unit

Claims (10)

It is a speaker identification device that identifies a speaker by using a plurality of voice data acquired by a plurality of microphones provided corresponding to a plurality of target persons.
For each voice data, a representative value acquisition unit that acquires a representative value indicating a feature regarding loudness for each predetermined first period, and a representative value acquisition unit.
A comparison unit that compares two representative values acquired from two different voice data for each first period and obtains a comparison result for each first period.
A speaker identification device including a specific unit for identifying a speaker in the second period by using a plurality of the comparison results obtained in a second period longer than the first period. A standardization processing unit for calculating a standardized representative value obtained by standardizing the representative value between the voice data is provided.
The speaker identification device according to claim 1, wherein the comparison unit uses the standardized representative value as the representative value to obtain the comparison result. The comparison unit calculates the difference between the two representative values acquired from the two different voice data for each first period.
The speaker identification device according to claim 1 or 2, wherein the identification unit calculates a difference average value obtained by averaging the difference for each second period, and identifies a speaker based on the positive or negative of the difference average value. .. The story according to claim 3, wherein the specific unit determines whether the difference average value is positive or negative, and when the same code is consecutive for a predetermined number of times or for a predetermined period of time, it is determined that the speaker corresponding to the code is speaking. Person identification device. When three or more of the microphones are installed, the comparison unit generates a combination of two different microphones, and obtains the voice data obtained from the two voice data corresponding to each of the generated combinations. The representative values are compared for each of the first periods, and the comparison result is obtained.
The story according to any one of claims 1 to 4, wherein the specific unit identifies a speaker in the second period by using a plurality of the comparison results obtained in the second period in each of the combinations. Person identification device. The speaker identification device according to any one of claims 1 to 5, wherein the first period is set to a period longer than one cycle of the sound wave waveform based on the voice data. The sampling frequency of the audio data is 44.1 kHz.
The speaker identifying device according to any one of claims 1 to 6, wherein the first period is set within a range of 1 ms or more and 30 ms or less. The speaker identifying device according to any one of claims 1 to 7, wherein the second period is set to a period five times or more as long as the first period. It is a speaker identification method that identifies a speaker by using a plurality of voice data acquired by a plurality of microphones provided corresponding to a plurality of target persons.
A representative value acquisition step of acquiring a representative value indicating a characteristic of loudness for each of the voice data for each predetermined first period,
A comparison step of comparing two representative values acquired from two different voice data for each first period and obtaining a comparison result for each first period.
A speaker identification method in which a computer executes a specific step of identifying a speaker in the second period by using a plurality of the comparison results obtained in a second period longer than the first period. It is a speaker identification program for identifying a speaker by using a plurality of voice data acquired by a plurality of microphones provided corresponding to a plurality of target persons.
For each voice data, a representative value acquisition process for acquiring a representative value indicating a characteristic regarding loudness for each predetermined first period, and a representative value acquisition process.
A comparison process in which two representative values acquired from two different voice data are compared for each first period and a comparison result is obtained for each first period.
A speaker identification program for causing a computer to perform a specific process for identifying a speaker in the second period by using a plurality of the comparison results obtained in a second period longer than the first period.

Images