Identification device, identification method, and recording medium

Abstract

An identification device (10) includes an acquisition unit (11) that acquires voice data, an identification unit (12) that obtains a score indicating a similarity between the voice data acquired by the acquisition unit (11) and voice data involved in speech of a predetermined speaker, by speaker identification processing, and a correction unit (13) that, in a case where it is determined that the voice data acquired by the acquisition unit (11) has a feature that degrades the identification performance of the speaker identification processing by the identification unit (12), performs a correction process that reduces the effect of degradation of the identification performance on the score on the score and outputs the score.

Description

Technical Field

[0001] This invention relates to an identification device, an identification method, and a recording medium. Background Technology

[0002] There is a technique for identifying the speaker of an input voice data by using the similarity between registered voice data and input voice data (see Patent Document 1).

[0003] (Existing technical literature)

[0004] (Patent Documents)

[0005] Patent Document 1: Japanese Patent Application Publication No. 2015-55835 Summary of the Invention

[0006] The problem that the invention aims to solve

[0007] However, depending on the content of the audio data, the performance of audio recognition can degrade. This degraded performance can lead to a decrease in the accuracy of speaker recognition.

[0008] Therefore, the present invention provides a recognition device for identifying speakers with higher accuracy.

[0009] The methods used to solve the problem

[0010] One aspect of the present invention relates to a recognition device comprising: an acquisition unit for acquiring sound data; a recognition unit for acquiring a similarity score by means of speaker recognition processing, the similarity being the similarity between the sound data acquired by the acquisition unit and sound data related to the speech of a specified speaker; and a correction unit for performing correction processing on the score and outputting it when it is determined that the sound data acquired by the acquisition unit has characteristics that reduce the recognition performance of the speaker recognition processing performed by the recognition unit, the correction processing being a process that reduces the impact of the reduction in recognition performance on the score.

[0011] Furthermore, these general or specific forms can be implemented by systems, methods, integrated circuits, computer programs, or computer-readable recording media such as CD-ROMs, or by any combination of systems, methods, integrated circuits, computer programs, and recording media.

[0012] Invention Effects

[0013] The recognition device of the present invention can identify speakers with higher accuracy. Attached Figure Description

[0014] Figure 1This is an illustrative diagram showing an example of a method for identifying speakers using fractions.

[0015] Figure 2 This is a block diagram illustrating the functional configuration of the identification device in the embodiment.

[0016] Figure 3 This is a flowchart illustrating the processes performed by the correction unit in the embodiment, such as obtaining representative values.

[0017] Figure 4 This is an explanatory diagram showing the correction score performed by the correction unit in the embodiment.

[0018] Figure 5 This is a flowchart illustrating the identification process performed by the identification device in the embodiment.

[0019] Figure 6 This is an explanatory diagram showing the first evaluation result of the recognition performance performed by the recognition device in the embodiment.

[0020] Figure 7 This is an explanatory diagram showing the second evaluation result of the recognition performance performed by the recognition device in the embodiment. Detailed Implementation

[0021] (Basic knowledge of this invention)

[0022] Regarding the speaker identification techniques described in the "Background Art" section, the inventors of this invention have found the following problems to arise.

[0023] In the past, techniques for speaker identification using audio data have relied on a score showing the similarity between registered audio data and input audio data to identify the speaker. Registered audio data refers to the audio data related to the speech of a specific speaker, while input audio data refers to the audio data related to the speech of an unknown speaker.

[0024] The scores obtained using the above technique are higher when the speaker in the input audio data matches the speaker in the registered audio data, and lower when they don't. Therefore, it's assumed that an appropriate threshold is determined to distinguish the magnitude of the output score. Then, if the score is greater than the threshold, it's determined that the speaker in the input audio data and the speaker in the registered audio data are the same person; otherwise, it's determined that the speaker in the input audio data and the speaker in the registered audio data are different. This is how the speaker identification of the input audio data is performed.

[0025] Figure 1 This is an illustrative diagram showing an example of a method for identifying speakers using fractions.

[0026] Figure 1 A histogram showing the scores for speaker recognition processing is presented. The horizontal axis represents the score for speaker recognition processing, and the vertical axis represents the frequency (number of times) of that score.

[0027] Figure 1 The histogram shown in (a) is a histogram illustrating the similarity scores between two audio data points extracted from multiple audio data sets in various forms. The audio data sets involve short-duration audio, meaning the duration of the audio is relatively short, for example, less than 10 seconds. Furthermore, the multiple audio data sets include audio data from the speeches of multiple speakers, and more specifically, multiple audio data from the individual speeches of multiple speakers. Figure 1 Histogram (1) shown in (a) is a histogram of the scores when the speakers of the two extracted voice data are the same person, and histogram (2) is a histogram of the scores when the speakers of the two extracted voice data are different people.

[0028] Although Figure 1 The histogram shown in (b) is... Figure 1 The histograms shown in (a) are the same, but the durations of the sounds involved in the multiple sound data are different. That is, used to derive... Figure 1 The histogram shown in (b) represents multiple audio data points involving long-duration audio, meaning the audio duration is relatively long, for example, more than 10 seconds.

[0029] Furthermore, regarding fractions, it is assumed that they are used to express the degree of large or small, or to compare sizes, and the numerical value shown as a fraction does not itself have any special meaning.

[0030] like Figure 1 As shown in (a), in the results of speaker identification processing for audio data with a short duration, the similarity scores between audio data of different speakers are generally less than -10, while the similarity scores between audio data of the same speaker are generally greater than -10.

[0031] Therefore, for speaker identification processing of audio data with a short duration, a threshold of -10 or a value close to -10 can be used. Then, in speaker identification processing, groups of audio data with scores above the threshold may be identified as groups of audio data from the same speaker, while groups of audio data with scores below the threshold may be identified as groups of audio data from different speakers.

[0032] On the other hand, such as Figure 1As shown in (b), in the results of speaker identification processing for audio data with a longer duration, the similarity scores between audio data of different speakers are generally less than 50, while the similarity scores between audio data of the same speaker are generally greater than 50.

[0033] Therefore, for the results of speaker recognition processing for audio data with a long duration, a value of 50 or close to 50 can be used as a threshold.

[0034] However, when performing speaker recognition processing on audio data where short-duration audio data is mixed with long-duration audio data, determining an appropriate threshold is difficult.

[0035] Specifically, for use Figure 1 In the speaker recognition processing shown in (a) and (b), when short-duration and long-duration audio data are mixed together, using a threshold of -10 or a value close to -10 allows for appropriate and accurate recognition processing for the short-duration audio data, but not for the long-duration audio data; in other words, accuracy decreases. Conversely, using a threshold of 50 or a value close to 50 allows for appropriate and accurate recognition processing for the long-duration audio data, but not for the short-duration audio data; in other words, accuracy decreases. Furthermore, using any value other than those mentioned above as a threshold for the mixed audio data results in inappropriate recognition processing for both long-duration and short-duration audio data; in other words, accuracy decreases.

[0036] Thus, depending on the content of the audio data, there can be a problem of decreased performance in speaker recognition processing of the audio data.

[0037] To address this problem, one aspect of the present invention relates to a recognition device comprising: an acquisition unit for acquiring sound data; a recognition unit for acquiring a score indicating similarity through speaker recognition processing, the similarity being the similarity between the sound data acquired by the acquisition unit and sound data related to the speech of a specified speaker; and a correction unit for performing correction processing on the score and outputting it when it is determined that the sound data acquired by the acquisition unit has characteristics that reduce the recognition performance of the speaker recognition processing performed by the recognition unit, the correction processing being a process that reduces the impact of the reduction in recognition performance on the score.

[0038] In the above-described manner, when the audio data exhibits characteristics that degrade the recognition performance of speaker recognition processing, the recognition device corrects the score and outputs it. Since the correction process reduces the impact of the degraded recognition performance on the score, the score of the audio data with the aforementioned characteristics is corrected to a score assuming a minimal degradation in recognition performance. Because the corrected score serves as a common benchmark with the scores of audio data without the aforementioned characteristics, and because the corrected score indicates whether the audio data belongs to the same speaker or different speakers, the recognition device can identify with higher accuracy whether the acquired audio data pertains to the speech of the specified speaker. Thus, the recognition device can perform speaker recognition with higher precision.

[0039] For example, the correction unit may perform a proximity processing on the scores obtained by the recognition unit in the correction process, wherein the proximity processing is such that the distribution of scores obtained by the recognition unit for two voice data having the same speaker's characteristics is close to the distribution of scores obtained by the recognition unit for two voice data not having the same speaker's characteristics.

[0040] By employing the above-described configuration, the recognition device adjusts the distribution of scores obtained by the recognition unit through correction processing. This allows the scores obtained based on voice data of the same speaker exhibiting characteristics that degrade recognition performance to be closer to the scores obtained based on voice data of the same speaker assuming they do not possess such characteristics. Therefore, the recognition device can more easily and with higher accuracy identify speakers.

[0041] For example, the correction unit may perform scaling processing on the score obtained by the recognition unit in the correction process using a first representative value, a second representative value, and a third representative value. The scaling processing is a process that transforms the range from the third representative value to the second representative value for the score obtained by the recognition unit into a range from the third representative value to the first representative value. The first representative value is a representative value of the score obtained by the recognition unit for two or more first voice data of the same speaker obtained in advance, and the first voice data is also two or more first voice data that do not have the aforementioned feature. The second representative value is a representative value of the score obtained by the recognition unit for two or more second voice data of the same speaker obtained in advance, and the second voice data is also two or more voice data that have the aforementioned feature. The third representative value is a representative value of the score obtained by the recognition unit for two or more third voice data of different speakers obtained in advance.

[0042] By employing the above-described configuration, the recognition device performs scaling processing on the score obtained by the recognition unit using pre-obtained first, second, and third representative values. This allows the score obtained based on voice data of the same speaker exhibiting characteristics that degrade recognition performance to be close to the score obtained based on voice data of the same speaker assuming they do not possess such characteristics. Therefore, the recognition device can more easily and with higher accuracy identify speakers.

[0043] For example, the correction unit may, in the scaling process, use S1, which is the score obtained by the recognition unit, V1, which is the first representative value, V2, which is the second representative value, and V3, which is the third representative value, to calculate S2, which is the score after correction by the correction unit, through (Equation A).

[0044] S2=(S1-V3)×(V1-V3) / (V2-V3)+V3(Formula A)

[0045] In the above configuration, the recognition device can easily obtain the corrected score S2 by performing a scaling process (represented by Equation A) on the score S1 obtained by the recognition unit. Therefore, the recognition device can perform speaker recognition with higher accuracy while more easily performing score correction processing.

[0046] For example, features that degrade the recognition performance of the speaker recognition process may include: the duration of the sound data obtained by the acquisition unit being less than a threshold; or the level of noise contained in the sound data obtained by the acquisition unit being above a threshold; or the echo time of the sound data obtained by the acquisition unit being above a threshold.

[0047] Through the above-described configuration, the recognition device performs score correction processing based on the following judgment: features of the sound data acquired by the acquisition unit that are related to the duration of the sound, the level of noise contained in the sound, or the echo time of the sound are determined to be features that degrade recognition performance. Therefore, the recognition device can more easily and with higher accuracy identify the speaker.

[0048] For example, the first representative value may be the average, median, or mode of one or more scores obtained by the recognition unit for the two or more first sound data; the second representative value may be the average, median, or mode of one or more scores obtained by the recognition unit for the two or more second sound data; and the third representative value may be the average, median, or mode of one or more scores obtained by the recognition unit for the two or more third sound data.

[0049] In this manner, the recognition device uses the average, median, or mode of one or more scores as representative values ​​(i.e., a first representative value, a second representative value, or a third representative value) to perform scaling. Therefore, the recognition device can more easily perform score correction while recognizing the speaker with higher accuracy.

[0050] Furthermore, one aspect of the present invention relates to a recognition method in which sound data is obtained, a similarity score is obtained through speaker recognition processing, and when it is determined that the obtained sound data has characteristics that reduce the recognition performance of the speaker recognition processing, the score is corrected and then output. The similarity is the similarity between the obtained sound data and the sound data involved in the speech of a specified speaker, and the correction processing is a process that reduces the impact of the reduction in recognition performance on the score.

[0051] The above-described form can achieve the same effect as the aforementioned identification device.

[0052] Furthermore, one aspect of the present invention relates to a recording medium that records a program that causes a computer to execute the above-described identification method.

[0053] The above-described form can achieve the same effect as the aforementioned identification device.

[0054] Furthermore, these general or specific forms can be implemented by systems, methods, integrated circuits, computer programs, or computer-readable recording media such as CD-ROMs, or by any combination of systems, methods, integrated circuits, computer programs, and recording media.

[0055] The embodiments will now be described in detail with reference to the accompanying drawings.

[0056] Furthermore, the embodiments described below are all general or specific examples. The numerical values, shapes, materials, constituent elements, the arrangement and connection methods of the constituent elements, the steps, and the order of the steps shown in the following embodiments are all examples and are not intended to limit the present invention. Moreover, constituent elements of the independent technical solutions that are not described in the following embodiments will be described as arbitrary constituent elements.

[0057] (Implementation Method)

[0058] The recognition device in this embodiment, which identifies the speaker with higher accuracy, will be described.

[0059] Figure 2 This is a block diagram illustrating the functional configuration of the identification device 10 in this embodiment.

[0060] like Figure 2 As shown, the identification device 10 includes an acquisition unit 11, an identification unit 12, a correction unit 13, and a storage unit 14. The aforementioned functional units of the identification device 10 may be implemented by a processor (CPU (Central Processing Unit) of the identification device 10 executing a predetermined program using memory, or they may be implemented by a storage device.

[0061] The acquisition unit 11 is a functional unit for acquiring sound data. The sound data acquired by the acquisition unit 11 is the sound data of the object being identified by the recognition device 10 for speaker identification. The acquisition unit 11 can also acquire sound data from a device external to the recognition device 10 via communication. Furthermore, if the acquisition unit 11 is a microphone that picks up the sound reaching the recognition device 10 and generates sound data, the acquisition unit 11 can also acquire sound data by generating sound data.

[0062] The recognition unit 12 is a functional unit that performs speaker recognition processing. By performing speaker recognition processing, the recognition unit 12 obtains a score indicating the similarity between the audio data obtained by the acquisition unit 11 and the audio data related to the speaker's speech. The speaker recognition processing is implemented using well-known techniques, such as x-vectors.

[0063] The correction unit 13 is a functional unit that performs correction processing on the score obtained by the recognition unit 12. The correction unit 13 determines whether the audio data obtained by the acquisition unit 11 has characteristics that would decrease the recognition performance of the speaker recognition processing performed by the recognition unit 12. Then, if it is determined that the audio data obtained by the acquisition unit 11 has the aforementioned characteristics, the correction unit 13 performs correction processing on the score obtained by the recognition unit 12 to suppress the decrease in recognition performance, and outputs the corrected score. Conversely, if it is determined that the audio data obtained by the acquisition unit 11 does not have the aforementioned characteristics, the correction unit 13 does not perform the aforementioned correction processing, and outputs the score obtained by the recognition unit 12.

[0064] As an example, in the correction process, the correction unit 13 performs a proximity processing on the scores obtained by the recognition unit 12, which means that the distribution of scores obtained by the recognition unit 12 for two voice data of the same speaker that have the above-mentioned characteristics is close to the distribution of scores obtained by the recognition unit 12 for two voice data of the same speaker that do not have the above-mentioned characteristics.

[0065] More specifically, in the correction process, the correction unit 13 uses three pre-obtained representative values—a first representative value, a second representative value, and a third representative value—to perform a scaling process (generally also called a scale transformation process) on the score obtained by the recognition unit 12. The scaling process transforms the scale of the score obtained by the recognition unit 12 from the range of the third representative value to the second representative value to the range of the third representative value to the first representative value. Furthermore, the scaling process applies to the entire range of the score obtained by the recognition unit 12, and is not limited to the range from the third representative value to the second representative value.

[0066] Here, the first representative value is a representative score obtained by the recognition unit 12 for two or more pre-acquired voice data of the same speaker, and this voice data also consists of two or more voice data that do not have the aforementioned characteristics (also referred to as first voice data). The second representative value is a representative score obtained by the recognition unit 12 for two or more pre-acquired voice data of the same speaker, and this voice data also consists of two or more voice data that have the aforementioned characteristics (also referred to as second voice data). The third representative value is a representative score obtained by the recognition unit 12 for two or more pre-acquired voice data of different speakers (also referred to as third voice data). In addition, two or more first voice data, two or more second voice data, and two or more third voice data are included in the voice data 16 stored in the storage unit 14, and can be obtained by reading the voice data 16.

[0067] The first representative value is a value obtained by statistically processing one or more scores obtained by the recognition unit 12 for the two or more first sound data points, specifically, the average, median, or mode. The second representative value is a value obtained by statistically processing one or more scores obtained by the recognition unit 12 for the two or more second sound data points, specifically, the average, median, or mode. The third representative value is a value obtained by statistically processing one or more scores obtained by the recognition unit 12 for the two or more third sound data points, specifically, the average, median, or mode.

[0068] The above scaling process is represented by the following (Equation 1). That is, using S1, which is the score obtained by the recognition unit 12, V1, which is the first representative value, V2, which is the second representative value, and V3, which is the third representative value, S2, which is the score after being corrected by the correction unit 13, is represented by the following (Equation 1).

[0069] S2=(S1-V3)×(V1-V3) / (V2-V3)+V3 (Formula 1)

[0070] Furthermore, features that degrade the recognition performance of the speaker recognition processing performed by the recognition unit 12 include, for example, the duration of the sound data acquired by the acquisition unit 11 being less than a threshold; or the noise level contained in the sound data being above a threshold; or the echo time of the sound data being above a threshold. Hereinafter, we will describe the case where the duration of the sound data is less than a threshold. As an example, the threshold for the duration of the sound data can be set to 10 seconds, but it is not limited to this and can be set in the range of approximately 1 second to 20 seconds. Additionally, the threshold for the noise level can be set, for example, to 12 dB as an expression of the signal-to-noise ratio (SN ratio), but it is not limited to this and can be set in the range of approximately 20 dB to 0 dB. Furthermore, as an example, the threshold for the echo time can be set to 500 milliseconds, but it is not limited to this and can be set in the range of approximately 300 milliseconds to 800 milliseconds.

[0071] Storage unit 14 is a storage device that stores sound data 16 for correction processing by correction unit 13. The sound data 16 stored in storage unit 14 includes two or more first sound data, two or more second sound data, and two or more third sound data. Storage unit 14 can be a volatile storage device such as RAM (Read Only Memory), or a non-volatile storage device such as HDD (Hard Disk Drive) or SSD (Solid State Drive).

[0072] Figure 3 This is a flowchart showing the processing steps performed by the correction unit 13 in this embodiment to obtain representative values, etc.

[0073] Here, sound data that has characteristics that reduce the recognition performance of speaker recognition processing performed by recognition unit 12 is called "sound data under adverse conditions", and sound data that does not have the above characteristics is called "sound data under good conditions".

[0074] In step S11, the correction unit 13 extracts feature values ​​from two or more voice data points of the same speaker under good conditions, and uses the extracted feature values ​​to obtain a score indicating the similarity between the two or more voice data points. When there are two voice data points, the correction unit 13 obtains a score indicating the similarity. More generally, when there are N voice data points, the correction unit 13 obtains a score by combining the voice data points in turn. N C2 represents the similarity scores. Furthermore, it's not necessary to utilize all combinations; in this case, the number of scores obtained can be less than [a certain number]. N C2 items.

[0075] In step S12, the correction unit 13 obtains a representative value V1 of the scores of two or more voice data of the same speaker under good conditions obtained in step S11. If only one score is obtained in step S11, this score is used as the representative value V1. If multiple scores are obtained in step S11, the representative value V1 is obtained by statistically processing these multiple scores.

[0076] In step S13, the correction unit 13 extracts feature values ​​from two voice data points of the same speaker under adverse conditions, and uses the extracted feature values ​​to obtain a score indicating the similarity between the two voice data points. The number of scores obtained is the same as that described in step S11.

[0077] In step S14, the correction unit 13 obtains a representative value V2 of the scores of the two voice data of the same speaker under adverse conditions obtained in step S13. The representative value V2 is obtained using the same method as in step S12 for the scores obtained in step S13.

[0078] In step S15, the correction unit 13 extracts feature values ​​from two voice data points of different speakers and uses the extracted feature values ​​to obtain a score indicating the similarity between the two voice data points. The number of scores obtained is the same as that described in step S11.

[0079] In step S16, the correction unit 13 obtains a representative value V3 for the scores of two different speakers' voice data obtained in step S15. The representative value V3 is obtained using the same method as in step S12 for the scores obtained in step S15.

[0080] Furthermore, although steps S11 and S12 must be executed in this order, they can also be executed after the other steps S13 to S16. Similarly, although steps S13 and S14 must be executed in this order, they can also be executed before or after the other steps S11 to S12 and S15 to S16. Although steps S15 and S16 must be executed in this order, they can also be executed before the other steps S11 to S14.

[0081] Figure 4 This is an explanatory diagram showing the correction score performed by the correction unit 13 in this embodiment.

[0082] Figure 4 (a) shows the score obtained by the identification unit 12 (that is, the score before correction by the correction unit 13), which is also the score from (1) to (4) below.

[0083] (1) is a score that shows the similarity between two or more audio data from the same speaker involved in a short period of sound (equivalent to...). Figure 1 (a) of (1))

[0084] (2) is a score showing the similarity between two or more audio data from different speakers involved in a short period of sound (equivalent to...). Figure 1 (a) of (2))

[0085] (3) is a score that shows the similarity between two or more audio data from the same speaker involved in a long period of sound (equivalent to...). Figure 1 (b) of (1))

[0086] (4) is a score showing the similarity between two or more voice data from different speakers involved in a long-duration audio recording (equivalent to...). Figure 1 (b) of (2))

[0087] and, Figure 4 (a) shows representative values ​​V1, V2, and V3. Representative value V1 is the value calculated by the correction unit 13 based on the sound data in (3) above. Representative value V2 is the value calculated by the correction unit 13 based on the sound data in (1) above. Representative value V3 is the value calculated by the correction unit 13 based on the sound data in (2) and (4) above.

[0088] Figure 4 (b) shows the score output by the correction unit 13, and also shows the similarity scores between the respective sound data in (1) to (4) above.

[0089] Here, Figure 4 The scores in (1) and (2) shown in (b) are scores obtained by the recognition unit 12 after correction processing by the correction unit 13. By performing the correction processing (specifically, scaling processing), the scale from the representative value V3 to the representative value V2 is transformed to the scale from the representative value V3 to the representative value V1. Furthermore, Figure 4 The scores in (3) and (4) shown in (b) are the scores obtained by the recognition unit 12 itself.

[0090] Before and after the correction process performed by the correction unit 13, the representative value V3 remains unchanged. Furthermore, the representative value V2 after the correction process performed by the correction unit 13 is the same as the representative value V1.

[0091] In the scores after such correction, for both short and long audio data, the similarity scores between audio data from different speakers are generally less than 50, while the similarity scores between audio data from the same speaker are generally greater than 50. Therefore, for audio data involving both short and long audio, by using a value of 50 or close to 50 as a threshold, it is possible to distinguish between the same speaker and different speakers. The same applies to audio data where short and long audio are mixed together.

[0092] Furthermore, the threshold (equivalent to 50 mentioned above) used to distinguish between scores indicating similarity between voice data of the same speaker and voice data of different speakers can be calculated either through calculation based on the distribution of scores or determined by human judgment.

[0093] The following describes the processing (also called the identification method) performed by the identification device 10 configured as described above.

[0094] Figure 5 This is a flowchart illustrating the identification process performed by the identification device 10 in this embodiment.

[0095] In step S21, the acquisition unit 11 acquires sound data.

[0096] In step S22, the recognition unit 12 extracts the feature data of the sound data obtained in step S21 and obtains a score.

[0097] In step S23, the correction unit 13 determines whether the sound data obtained in step S21 has features that degrade recognition performance. If the sound data is determined to have the aforementioned features ("Yes" in step S23), the process proceeds to step S24; otherwise ("No" in step S23), the process proceeds to step S25.

[0098] In step S24, the correction unit 13 performs a correction process on the score obtained in step S22.

[0099] In step S25, the correction unit 13 outputs a score. If it is determined that the sound data obtained in step S21 has characteristics that reduce recognition performance ("Yes" in step S23), the score output by the correction unit 13 is the score after correction processing. If it is determined that the data does not have the aforementioned characteristics ("No" in step S23), the score output by the correction unit 13 is the score after no correction processing, which is the score obtained in step S22.

[0100] pass Figure 5 Through the series of processes shown, the recognition device 10 is able to identify the speaker with higher accuracy.

[0101] The following is an example of the evaluation results of the performance of the identification device 10.

[0102] Figure 6 This is an explanatory diagram showing the first evaluation result of the recognition performance performed by the recognition device 10 in this embodiment.

[0103] As for the error rate, Figure 6 The error rate for each of the four audio data sets (data sets 1 through 4) is shown, indicating whether the data belongs to the same speaker or different speakers, given a threshold that is considered appropriate.

[0104] Data type 1 refers to audio data involving long-duration sounds, such as audio data involving sounds with a duration of 10 seconds or more.

[0105] Data type 2 is short-duration audio data, such as data involving audio with a duration of less than 10 seconds.

[0106] Data No. 3 is a mixture of audio data from Data No. 1 and Data No. 2.

[0107] Data No. 4 is a mixture of data No. 1 and data No. 2. However, the correction unit 13 performs correction processing on the short-duration sound data included in the above sound data.

[0108] like Figure 6 As shown, the error rate for data number 3 is 2.15%. That is, without correction processing by the correction unit 13, errors occur at a rate of 2.15% regarding whether the audio data belongs to the same speaker or to different speakers.

[0109] On the other hand, regarding data number 4, the error rate shown is 0.78%. That is, through correction processing by the correction unit 13, the error rate regarding whether the data is from the same speaker or different speakers is improved from 2.15% to 0.78%.

[0110] Figure 7 This is an explanatory diagram showing the second evaluation result of the recognition performance performed by the recognition device 10 in this embodiment. Figure 7 The error rate shown means the same as Figure 6 The situation is the same as in [the previous sentence].

[0111] Data type 1 is sound data with low noise levels, such as sound data with a signal-to-noise ratio (SN) of 12 dB or higher.

[0112] Data point number 2 represents sound data with high noise levels, such as sound data with a signal-to-noise ratio (SN) of less than 12 dB.

[0113] Data No. 3 is a mixture of audio data from Data No. 1 and Data No. 2.

[0114] Data No. 4 is a mixture of data No. 1 and data No. 2. However, the correction unit 13 performs correction processing on the scores of the above audio data that contain audio data with high noise levels.

[0115] like Figure 7As shown, the error rate for data number 3 is 5.81%. That is, without correction processing by the correction unit 13, an error occurs at a rate of 5.81% regarding whether the voice data is from the same speaker or different speakers.

[0116] On the other hand, regarding data number 4, the error rate shown is 4.95%. That is, through correction processing by the correction unit 13, the error rate regarding whether the voice data is from the same speaker or different speakers is improved from 5.81% to 4.95%.

[0117] As described above, the recognition device of this embodiment corrects and outputs scores when the sound data has characteristics that degrade the recognition performance of speaker recognition processing. Since the correction process reduces the impact of the degradation in speaker recognition performance on the score, the scores of sound data with the aforementioned characteristics are corrected to scores assuming a minimal degradation in recognition performance. Because the corrected scores are based on a common benchmark with scores of sound data without the aforementioned characteristics, and because the corrected scores indicate whether the sound data belongs to the same speaker or different speakers, it is possible to identify with higher accuracy whether the acquired sound data pertains to the speech of the specified speaker. Thus, the recognition device can perform speaker recognition with higher accuracy.

[0118] Furthermore, by adjusting the distribution of scores obtained by the recognition unit through correction processing, the recognition device can make the scores obtained based on the voice data of the same speaker with characteristics that degrade recognition performance closer to the scores obtained based on the voice data of the same speaker assuming that such characteristics are not present. Therefore, the recognition device can more easily and with higher accuracy identify speakers.

[0119] Furthermore, by performing scaling processing on the score obtained by the recognition unit using pre-obtained first, second, and third representative values, the recognition device can make the score obtained based on the voice data of the same speaker with features that degrade recognition performance close to the score obtained based on the voice data of the same speaker assuming that such features are not present. Therefore, the recognition device can more easily and with higher accuracy identify the speaker.

[0120] Furthermore, by performing scaling processing (Equation 1) on the score S1 obtained by the recognition unit, the recognition device can easily obtain the corrected score S2. Therefore, the recognition device can perform speaker recognition with higher accuracy while performing score correction processing more easily.

[0121] Furthermore, the recognition device performs score correction processing based on the following judgment: features of the sound data acquired by the acquisition unit that are related to the duration of the sound, the level of noise contained in the sound, or the echo time of the sound are judged to be features that degrade recognition performance. Therefore, the recognition device can more easily and with higher accuracy identify the speaker.

[0122] Furthermore, the recognition device uses the average, median, or mode of more than one score as a representative value (i.e., a first representative value, a second representative value, or a third representative value) to perform scaling. Therefore, the recognition device can more easily perform score correction while recognizing the speaker with higher accuracy.

[0123] Furthermore, in the above embodiments, each component can be constructed by dedicated hardware, or it can be implemented by executing software programs suitable for each component. Each component can also be implemented by a program execution unit such as a CPU or processor reading and executing software programs recorded on a recording medium such as a hard disk or semiconductor memory. Here, the software implementing the identification device, etc., of the above embodiments is the following program.

[0124] That is, the program is a program that enables a computer to perform the following recognition method, in which sound data is obtained, a similarity score is obtained through speaker recognition processing, and if it is determined that the obtained sound data has characteristics that reduce the recognition performance of the speaker recognition processing, the score is corrected and then output, wherein the similarity is the similarity between the obtained sound data and the sound data involved in the speech of a specified speaker, and the correction processing is a process that reduces the impact of the reduction in recognition performance on the score.

[0125] While the identification device and the like in one or more embodiments have been described above according to the embodiments, the present invention is not limited to these embodiments. Various modifications conceivable to those skilled in the art can be performed on these embodiments to obtain forms, as well as forms constructed by combining constituent elements from different embodiments, may also be included within the scope of the one or more embodiments without departing from the spirit of the present invention.

[0126] Industrial applicability

[0127] The present invention provides a recognition device capable of identifying the speaker of a sound in sound data.

[0128] Symbol Explanation

[0129] 10 Identification Devices

[0130] 11. Obtained by the department

[0131] 12 Identification Department

[0132] 13 Correction Department

[0133] 14 Storage Department

[0134] 16. Audio Data

Claims

1. An identification device comprising: The acquisition department obtains sound data; The identification unit obtains a similarity score through speaker identification processing, the similarity being the similarity between the sound data obtained by the acquisition unit and the sound data related to the speech of a specified speaker; and The correction unit, upon determining that the sound data acquired by the acquisition unit has characteristics that degrade the recognition performance of the speaker recognition processing performed by the recognition unit, performs correction processing on the score and outputs it. The correction processing reduces the impact of the degradation in recognition performance on the score. The correction unit is mentioned in the correction process. The score obtained by the recognition unit is scaled using a first representative value, a second representative value, and a third representative value. This scaling process transforms the score obtained by the recognition unit from a range from the third representative value to the second representative value into a range from the third representative value to the first representative value. The first representative value is a representative value of the score obtained by the recognition unit for two or more first voice data of the same speaker obtained in advance, and the first voice data are also two or more voice data that do not have the aforementioned feature. The second representative value is a representative value of the score obtained by the recognition unit for two or more pre-acquired second voice data of the same speaker, and the second voice data also consists of two or more voice data having the aforementioned characteristics. The third representative value is a representative value of the score obtained by the recognition unit for two or more pre-acquired third voice data from mutually different speakers. The correction unit is used in the scaling process. Using S1, which is the score obtained by the recognition unit, V1, which is the first representative value, V2, which is the second representative value, and V3, which is the third representative value, S2, which is the score after correction by the correction unit, is calculated by formula A, namely S2=(S1-V3)×(V1-V3) / (V2-V3)+V3.

2. The identification device as described in claim 1, The correction unit is mentioned in the correction process. The scores obtained by the recognition unit are subjected to a proximity processing, which means that the distribution of scores obtained by the recognition unit for two voice data of the same speaker who have the aforementioned features is close to the distribution of scores obtained by the recognition unit for two voice data of the same speaker who do not have the aforementioned features.

3. The identification device as described in claim 1 or 2, Features that degrade the recognition performance of the speaker recognition process include: The sound data obtained by the acquiring unit has the characteristic that the duration of the sound is less than a threshold; or The sound data obtained by the acquiring unit contains a noise level above a threshold; or The sound data obtained by the acquiring unit has the characteristic that the echo time of the sound is above a threshold.

4. The identification device as described in claim 1, The first representative value is the average, median, or mode of one or more scores obtained by the recognition unit for the two or more first sound data. The second representative value is the average, median, or mode of one or more scores obtained by the recognition unit for the two or more second sound data. The third representative value is the average, median, or mode of one or more scores obtained by the recognition unit for the two or more third sound data.

5. An identification method, In the identification method, Obtain sound data, A similarity score is obtained through speaker recognition processing, whereby the similarity is the degree of similarity between the obtained audio data and the audio data related to the speaker's speech. If it is determined that the obtained audio data has characteristics that degrade the recognition performance of the speaker recognition process, the score is corrected and then output. This correction process reduces the impact of the reduced recognition performance on the score. In the correction process, The obtained score is scaled using a first representative value, a second representative value, and a third representative value. This scaling process transforms the obtained score from the range of the third representative value to the second representative value into the range of the third representative value to the first representative value. The first representative value is a representative value of the score obtained for two or more first voice data points of the same speaker obtained in advance, and the first voice data points are also two or more voice data points that do not have the aforementioned feature. The second representative value is a representative value of the score obtained for two or more pre-acquired second voice data of the same speaker, where the second voice data also consists of two or more voice data having the aforementioned characteristics. The third representative value is a representative value of the score obtained for two or more pre-obtained third voice data from different speakers. In the scaling process, Using S1 as the obtained score, V1 as the first representative value, V2 as the second representative value, and V3 as the third representative value, S2 as the corrected score is calculated by formula A, namely S2=(S1-V3)×(V1-V3) / (V2-V3)+V3.

6. A recording medium having a program that causes a computer to execute the identification method of claim 5.

Images