JP6553015B2 - Speaker attribute estimation system, learning device, estimation device, speaker attribute estimation method, and program - Google Patents

Description

  The present invention relates to a technique for estimating speaker attributes from speech.

  There is a need for a technology for identifying speaker attributes (e.g., gender, age, etc.) from speech in order to collect marketing information in a speech interactive robot or a call center. Conventional techniques for identifying speaker attributes include converting the voice quality of the input speech into an i-vector and identifying the i-vector using a support vector machine (SVM), or a mixed Gaussian model. There is a method of identifying using (GMM: Gaussian Mixture Model) (for example, see Non-Patent Document 1).

Shoko Miyamori et al., “Study on Child User Discrimination by Speech Recognition in a Single Word”, Proceedings of Information Science and Technology Forum, vol. 9 (3), pp. 469-472, 2010

  In the conventional speaker attribute estimation technique, improvement of the identification rate is a problem. The identification rate of speaker attribute estimation according to the prior art is about 80 to 90% in the case of three classes of adult male, adult female, and child (male and female). In particular, it is necessary to prevent misidentification when a feature (for example, noise, clipping of speech, etc.) that is not related to the attribute included in the learning data is included in the speech to be identified.

  In the prior art, it is also a problem that the accuracy of the identification result cannot be obtained. For example, if a response is returned based on an incorrect identification result by a voice interactive robot or the like, it may cause discomfort to the user. If the identification result is not certain, it is necessary to return the response as a neutral attribute. is there. In addition, if attributes can be identified during speech input and certainty can be sequentially obtained, it is possible to return a quick response based on the identification result.

  An object of this invention is to provide the speaker attribute estimation technique which can estimate a speaker attribute with higher precision than before in view of the above points.

  In order to solve the above problems, a speaker attribute estimation system of the present invention includes a learning device and an estimation device. The learning device includes an attribute label generation unit that generates an attribute label sequence representing a speaker attribute of each frame of learning speech from attribute information representing speaker attributes of each learning speech, an acoustic feature amount sequence of each frame of learning speech, and A deep learning unit that learns a deep neural network model using the attribute label series. The estimation device includes a posterior probability calculation unit that calculates a posterior probability sequence for each frame using a deep neural network model from an acoustic feature amount sequence for each frame of the input speech, and a logarithmic sum of the posterior probability sequences obtained for each speaker attribute. And an identification unit for identifying speaker attributes based on

  According to the present invention, it is possible to perform robust estimation with respect to spectral characteristics (for example, clip sound) other than noise and attributes, and it is possible to estimate speaker attributes (for example, gender and age) with high accuracy. In addition, since the reliability of the identification result can be obtained, for example, a voice interaction robot or the like can respond quickly.

FIG. 1 is a diagram illustrating a functional configuration of the learning device according to the first embodiment. FIG. 2 is a diagram illustrating a functional configuration of the estimation apparatus according to the first embodiment. FIG. 3 is a diagram illustrating a processing procedure of the speaker attribute estimation method according to the first embodiment. FIG. 4 is a diagram for explaining a method of creating a clip sound. FIG. 5 is a diagram illustrating a functional configuration of the learning device according to the second embodiment. FIG. 6 is a diagram illustrating a processing procedure of the speaker attribute estimation method according to the second embodiment. FIG. 7 is a diagram illustrating a functional configuration of the learning device according to the third embodiment. FIG. 8 is a diagram illustrating a processing procedure of the speaker attribute estimation method according to the third embodiment. FIG. 9 is a diagram illustrating a functional configuration of the learning device according to the fourth embodiment. FIG. 10 is a diagram illustrating a functional configuration of the estimation device according to the fourth embodiment. FIG. 11 is a diagram illustrating a processing procedure of the speaker attribute estimation method according to the fourth embodiment.

  Hereinafter, embodiments of the present invention will be described in detail. In addition, the same number is attached | subjected to the component which has the same function in drawing, and duplication description is abbreviate | omitted.

First Embodiment
In the first embodiment, the sequential posterior probability of the input speech is determined using the deep learning model, and the speaker attribute is estimated using the logarithmic posterior probability obtained by summing the posterior probabilities. This makes it possible to estimate speaker attributes with higher accuracy than in the past.

  The speaker attribute estimation system according to the first embodiment includes, for example, a learning device for learning a Deep Neural Network (DNN) model (hereinafter referred to as DNN model) from learning data, and an input speech using the learned DNN model. And an estimation device for estimating the speaker attribute. As shown in FIG. 1, the learning apparatus according to the first embodiment includes a learning data storage unit 10, a feature extraction unit 11, an attribute label creation unit 12, a deep learning unit 13, and a DNN model storage unit 20. As illustrated in FIG. 2, the estimation apparatus according to the first embodiment includes a DNN model storage unit 20, a feature amount extraction unit 21, a posterior probability calculation unit 22, and an identification unit 23. The learning device and the estimation device perform the process of each step shown in FIG. 3 to realize the speaker attribute estimation method of the first embodiment.

  The learning device and the estimation device are configured, for example, by loading a special program into a known or dedicated computer having a central processing unit (CPU), a main memory (RAM), and the like. It is a special device. For example, the learning device and the estimation device execute each process under the control of the central processing unit. The data input to the learning device and estimation device and the data obtained by each process are stored, for example, in the main storage device, and the data stored in the main storage device is read out as needed to perform other processing. It is used. Further, at least a part of each processing unit of the learning device and the estimation device may be configured by hardware such as an integrated circuit. Each storage unit included in the learning device and the estimation device is, for example, a main storage device such as a random access memory (RAM), an auxiliary storage device configured by a semiconductor memory element such as a hard disk, an optical disk, or a flash memory. Alternatively, it can be configured by middleware such as a relational database or key-value store. Each storage unit included in the learning device and the estimation device only needs to be logically divided, and may be stored in one physical storage device.

The learning data storage unit 10 stores learning data used to learn the DNN model. The learning data includes learning speech s (k, t) and attribute information L (k). k (= 0, 1,..., K) is a learning speech number. t (= 0, 1, ..., T _k -1) is the sample time. T _k is the time length of the k-th learning speech. s (k, t) is the amplitude of the k-th learning speech at sample time t when the sampling frequency is f _s [Hz]. L (k) is a numerical value indicating the speaker attribute of the kth learning speech. For example, L (k) = 0 is expressed as “adult male”, L (k) = 1 is expressed as “adult female”, and L (k) = 2 is expressed as “child”.

  With reference to FIG. 3, the processing procedure of the speaker attribute estimation method executed by the learning device and the estimation device of the first embodiment will be described.

In step S1, the feature quantity extraction unit 11 of the learning apparatus reads the learning speech s (k, t) stored in the learning data storage unit 10, and the mel frequency cepstrum coefficient (MFCC) is read from the learning speech s (k, t). : Mel-Frequency Cepstrum Coefficient) acoustic feature sequence c (k, i, j) is extracted and output. i (= 0, 1, 2,…, I _k -1) is the frame number, I _k is the number of frames of the kth learning speech, and j (= 0, 1, 2,…, J _k -1) is the sound The number indicating the dimension of the feature, J _k is the number of dimensions of the acoustic feature of the k-th learning speech. The mel frequency cepstrum coefficient may be extracted by a known method. For example, 12 dimensions and their Δ feature amount and Δ power feature amount may be used. The analysis frame width should be about 10 milliseconds. The extracted acoustic feature quantity sequence c (k, i, j) is sent to the attribute label creating unit 12 and the deep learning unit 13.

In step S2, the attribute label creating unit 12 of the learning device reads out the attribute information L (k) stored in the learning data storage unit 10, and obtains the attribute label sequence l (k, i) of the number I _k of frames of learning speech. create. Specifically, l (k, i) = L (k) is set for all frames (i = 0, 1,..., I _k −1). The created attribute label sequence l (k, i) is sent to the deep learning unit 13.

  In step S 3, the deep learning unit 13 of the learning device receives the acoustic feature quantity sequence c (k, i, j) of the learning speech s (k, t) received from the feature quantity extraction unit 11 and the attribute label creating unit 12 By using the attribute label sequence l (k, i) of the received learning speech s (k, t), the DNN model λ represented by Equation (1) is learned.

The DNN model is used in image recognition and speech recognition, and can learn detailed features. p (m | λ, c (k, i, j)) is a posterior probability that the feature quantity c (k, i, j) belongs to the attribute m (= 0, 1,..., M). For example, m = 0 is “adult male”, m = 1 is “adult female”, m = 2 is “child”, and the like. The learning of the DNN model is based on the attribute labels l (k, i) for all frames (i = 0, 1, 2,..., I _k -1) of all speech (k = 0, 1, 2,..., K). Do it using). The learned DNN model λ is stored in the DNN model storage unit 20.

  In step S4, the feature amount extraction unit 21 of the estimation device calculates the acoustic feature amount sequence c ′ (i, j) (i = 0, 1,..., I) of the mel frequency cepstrum coefficient from the input speech s ′ (t). −1, j = 0, 1,..., J−1, I is the number of frames of the input speech, and J is the number of dimensions of the acoustic feature quantity of the input speech. The extracted acoustic feature quantity sequence c ′ (i, j) is sent to the posterior probability calculation unit 22.

  In step S 5, the posterior probability calculation unit 22 of the estimation device stores the acoustic feature quantity sequence c ′ (i, j) of the input speech s ′ (t) received from the feature quantity extraction unit 21 in the DNN model storage unit 20. A posteriori probability series q (i, m) = p (m |?, C '(i, j)) (i = 0, 1, ..., I-1, m = 0, 1,…, M). The calculated posterior probability sequence q (i, m) is sent to the identification unit 23.

  In step S6, the identification unit 23 of the estimation device identifies and outputs the speaker attribute L ′ from the posterior probability series q (i, m) received from the posterior probability calculation unit 22. The speaker attribute is identified by calculating the logarithmic sum of the posterior probabilities of all frames by using the expression (2), and outputting the speaker attribute having the highest value as the identification result.

Second Embodiment
In a voice dialogue robot or the like, a clip sound with a completely swung amplitude may be input because the voice is input too close to the microphone. If a learning voice including a clip sound is present in part of the learning data, the same clip sound may be identified as an attribute having a feature of the clip sound unlike the original attribute when the same clip sound is input. Therefore, in the second embodiment, as shown in FIG. 4, by creating a clip sound clipped from learning data and adding it to the learning data, it is possible to identify the original attribute without being dragged by the features of the clip sound. To be possible.

  The learning apparatus according to the second embodiment, as shown in FIG. 5, includes a learning data storage unit 10, a feature quantity extraction unit 11, an attribute label creation unit 12, a deep learning unit 13, and a DNN model storage unit 20 according to the first embodiment. The clip sound synthesizer 14 is further provided. The learning device and the estimation device of the first embodiment perform the process of each step shown in FIG. 6 to realize the speaker attribute estimation method of the second embodiment.

  With reference to FIG. 6, the processing procedure of the speaker attribute estimation method of the second embodiment will be described. Below, it demonstrates centering on difference with the above-mentioned 1st embodiment.

  In step S7, the clip sound synthesis unit 14 of the learning device reads the learning speech s (k, t) stored in the learning data storage unit 10, amplifies the amplitude of the learning speech s (k, t), The clip sound S (k, t) is synthesized by rounding the amplitude exceeding the threshold value to the threshold value. The synthesized clip sound S (k, t) is stored in the learning data storage unit 10.

  Specifically, the clip sound synthesizer 14 synthesizes the clip sound as follows.

1. Speech S (k, t) = a * s (k, t) (k = 0, 1,..., K, t = 0, 1,..., T obtained by multiplying the amplitude of learning speech s (k, t) by a Create _k -1).

2. In order to round out values of speech S (k, t) exceeding a predetermined threshold value ± h (h> 0), all samples (t = 0, 1) of all learning speech (k = 0, 1,..., K) ,…, T _k −1) are set as follows.

(A) If S (k, t)> h, set S (k, t) = h. (B) If S (k, t) <-h, set S (k, t) =-h. .

  It is advisable to set multiple values for a. For example, a = 1, 3, 6, etc. The clip sound synthesized in this way shows a waveform as shown in FIG.

  In the subsequent processing of the learning device, the added clip sound S (k, t) is used in the same manner as the learning sound s (k, t). As a result, it is possible to correctly estimate the speaker attributes even if the input speech s ′ (t) of the estimation device is a clipped speech.

Third Embodiment
The speaker attribute is likely to misidentify due to the influence of the unvoiced sound, since the feature is not likely to appear in the unvoiced sound. In addition, when a section where speech is not uttered is included, there may be erroneous identification by including noise around the corresponding portion in the learning data. Therefore, attribute identification is preferably performed only for voiced sounds. Therefore, in the third embodiment, it is possible to increase the identification rate by giving label data to the unvoiced sound or silent part of the learning data and excluding from the target of identification when the probability of unvoiced sound or silent is high. Do.

  The learning apparatus according to the third embodiment, as shown in FIG. 7, includes a learning data storage unit 10, a feature quantity extraction unit 11, an attribute label creation unit 12, a deep learning unit 13, a clip sound synthesis unit 14, and a DNN model storage unit. 20 as in the second embodiment, and further includes a voiced / unvoiced determination unit 15. The learning device and the estimation device of the first embodiment perform the process of each step shown in FIG. 8 to realize the speaker attribute estimation method of the third embodiment.

  With reference to FIG. 8, the processing procedure of the speaker attribute estimation method of the third embodiment will be described. Below, it demonstrates centering on difference with the above-mentioned 2nd embodiment.

  In step S8, the voiced / unvoiced determination unit 15 of the learning device reads the learning voice s (k, t) stored in the learning data storage unit 10, and determines the voiced / unvoiced section of the learning voice s (k, t). Voiced and unvoiced information v (k, i) is generated. The generated voiced / unvoiced information v (k, i) is stored in the learning data storage unit 10 in association with the clip sound S (k, t) synthesized by the clip sound synthesis unit 14. The voiced unvoiced information v (k, i) is, for example, v (k, i) = 1 if the i-th frame of the k-th learning speech s (k, t) is voiced and v (k (i) if unvoiced). k, i) = 0. The determination of voiced / unvoiced may be performed with the same frame width as that of the feature amount extraction unit 11 and may be performed by a general method of basic frequency extraction.

In step S2, the attribute label creating unit 12 of the learning device reads the attribute information L (k) and the voiced voiceless information v (k, i) stored in the learning data storage unit 10, and the attributes for the number of frames of learning speech Create a label series l (k, i). Specifically, for all frames (i = 0, 1,..., I _k −1), the voiced part (in the case of v (k, i) = 1) gives l (k, i) = L (k) Set and unvoiced part (when v (k, i) = 0) set l (k, i) = -1.

  In step S3, the deep learning unit 13 of the learning device learns the DNN model λ as in the first embodiment. In the DNN model λ of the third embodiment, the unvoiced part (l (k, i) = − 1) is an attribute m = −1, and the feature amount c (k, i, j) is an attribute m (= −1, 0). , 1,…, M).

  In step S5, the a posteriori probability calculating unit 22 of the estimating device calculates the posterior probability sequence q (i, m) (i = 0, 1,..., I-1, m = -1, as in the first embodiment. 0, 1,…, M).

  In step S6, the identification unit 23 of the estimation device identifies and outputs the speaker attribute L ′ from the posterior probability series q (i, m) received from the posterior probability calculation unit 22. Since the identification unit 23 of the third embodiment learns the unvoiced part as the attribute m = −1, the identification is limited to only the voiced part using the function f (i, m) shown in Expression (3). Do.

Fourth Embodiment
In the fourth embodiment, the reliability indicating the probability of the identification result is obtained from the distribution of the posterior probability sequence at the time of learning and the posterior probability sequence at the time of estimation. The reliability is a numerical value of 0 or more and 1 or less, and it can be said that the closer to 1, the more reliable the identification result L ′. By using the reliability, for example, a voice dialogue robot or the like can select an appropriate response according to the reliability.

  The learning apparatus according to the fourth embodiment includes, as shown in FIG. 9, a learning data storage unit 10, a feature quantity extraction unit 11, an attribute label creation unit 12, a deep learning unit 13, and a DNN model storage unit 20 according to the first embodiment. The learning data posterior probability calculation unit 16, the reliability parameter learning unit 17, and the reliability parameter storage unit 30 are further provided. The estimation apparatus according to the fourth embodiment includes, as shown in FIG. 10, a DNN model storage unit 20, a feature quantity extraction unit 21, a posterior probability calculation unit 22, and an identification unit 23, as in the first embodiment. A calculation unit 24 and a reliability parameter storage unit 30 are further provided. The learning device and the estimation device perform the process of each step shown in FIG. 11 to realize the speaker attribute estimation method of the fourth embodiment.

  Although FIG. 9 shows a configuration in which the concept of the fourth embodiment is applied to the learning device of the first embodiment, the concept of the fourth embodiment can also be applied to the second embodiment and the third embodiment. . That is, the learning device according to the fourth embodiment may include one or both of the clip sound synthesis unit 14 and the voiced / unvoiced determination unit 15.

  With reference to FIG. 11, the processing procedure of the speaker attribute estimation method of the fourth embodiment will be described. Below, it demonstrates centering on difference with the above-mentioned 1st embodiment.

In step S9, the learning data posterior probability calculating unit 16 of the learning device performs a deep learning unit from the acoustic feature amount sequence c (k, i, j) extracted from the learning speech s (k, t) by the feature amount extracting unit 11 The posterior probability series q ′ (k, i, m) = p (m | λ, c (k, i, j)) (k = 0, 1,..., K, using the DNN model λ learned by 13) Calculate i = 0, 1,..., I _k -1, m = 0, 1,. The calculated posterior probability series q ′ (k, i, m) is sent to the reliability parameter learning unit 17.

  In step S10, the reliability parameter learning unit 17 of the learning apparatus receives the posterior probability sequence q ′ (k, i, m) received from the learning data posterior probability calculation unit 16 and the attribute label sequence l created by the attribute label creation unit 12. From (k, i), the average μ (m), standard deviation σ (m), and number of frames n (m) of the posterior probability series for obtaining the reliability are calculated. Hereinafter, these are collectively referred to as reliability parameters. The calculated reliability parameters μ (m), σ (m), and n (m) are stored in the reliability parameter storage unit 30.

  Specifically, the reliability parameter learning unit 17 calculates the reliability parameters μ (m), σ (m), and n (m) as follows.

  1. The number of frames n (m) and the total posterior probability s (m) are obtained from the equation (4).

  2. Average μ (m) = s (m) / n (m) is obtained for all attributes (m = 0, 1,..., M).

  3. For all the attributes (m = 0, 1,..., M), the difference total value d (m) from the average is obtained by equation (5).

  4. For all the attributes (m = 0, 1,..., M), the standard deviation σ (m) is obtained by equation (6).

  In step S <b> 11, the reliability calculation unit 24 of the estimation device calculates a reliability parameter storage unit from the posterior probability sequence q (i, m) output from the posterior probability calculation unit 22 and the speaker attribute L ′ output from the identification unit 23. Using the reliability parameters μ (m), σ (m), and n (m) stored in 30, the reliability r is obtained. The obtained reliability r is output together with the identification result L ′. The reliability r is the mean μ ′, standard deviation σ ′, and frame number n ′ of the posterior probability sequence q (i, L ′) when the attribute m = L ′ for the posterior probability sequence q (i, m). The distribution can be obtained based on the obtained distribution and the distribution based on the reliability parameters μ (m), σ (m), and n (m) obtained in advance.

  Specifically, the reliability calculation unit 24 calculates the reliability r as follows.

  1. The number n ′ of frames and the total posterior probability value s ′ are obtained by the equation (7).

  2. Find the average μ '= s' / n'.

  3. The difference total value d ′ from the average is obtained by Expression (8).

  4. The standard deviation σ ′ is obtained from equation (9).

  5. The statistic t is obtained from equation (10).

  6. In the t distribution T (x) with t> 0 and n '+ n (L')-2 degrees of freedom, the reliability r of the equation (11) is determined using the statistic t determined in 5 above.

  The reliability calculation unit 24 can also obtain the reliability r as follows without using the reliability parameters μ (m), σ (m), n (m). The reliability at this time is a value for determining whether the average value of the attribute to be obtained is significantly high from the average / variance value of the posterior probability series of each attribute. In this case, the learning device may not include the learning data posterior probability calculation unit 16, the reliability parameter learning unit 17, and the reliability parameter storage unit 30. The estimation device may not include the reliability parameter storage unit 30.

  Specifically, the reliability calculation unit 24 calculates the reliability r as follows.

  1. The number of frames n ′ and the total posterior probability value s ′ (m) of each attribute are obtained by Expression (12).

  2. Average μ ′ (m) = s ′ (m) / n ′ of each attribute is obtained.

  3. The difference total value d ′ (m) from the average of each attribute is obtained by Expression (13).

  4. The standard deviation σ ′ (m) of each attribute is obtained by Expression (14).

  5. A statistic t (m) between the identified speaker attribute L ′ and the other speaker attributes is obtained by Expression (15).

6. In the t (m) distribution T _m (x) with t (m)> 0 and degrees of freedom 2n′−2, the reliability of the equation (16) is obtained using the statistic t (m) obtained in 5 above. Find the average value of r.

  1-r represents the probability that the posterior probability in the t-test is an average μ. For example, in the case of 1-r <0.05, it can be said that the posterior probability series μ is significantly higher than the average posterior probability μ (m) in the speaker attributes other than the attribute m obtained in advance at the significance level of 5%. A value obtained by subtracting the probability of occurrence from 1 can be used as the reliability and the probability of the identification result L ′.

  As described above, the embodiments of the present invention have been described, but the specific configuration is not limited to these embodiments, and even if there is a design change or the like as appropriate without departing from the spirit of the present invention, Needless to say, it is included in the present invention. The various processes described in the embodiment are not only executed chronologically according to the order described, but may be executed in parallel or individually depending on the processing capability of the apparatus executing the process or the necessity.

[Program, recording medium]
When various processing functions in each device described in the above embodiment are realized by a computer, the processing contents of the functions that each device should have are described by a program. Then, by executing this program on a computer, various processing functions in each of the above devices are realized on the computer.

  The program describing the processing content can be recorded in a computer readable recording medium. As the computer readable recording medium, any medium such as a magnetic recording device, an optical disc, a magneto-optical recording medium, a semiconductor memory, etc. may be used.

  This program is distributed by selling, transferring, or lending a portable recording medium such as a DVD or CD-ROM in which the program is recorded. Furthermore, this 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, at the time of execution of the process, the computer reads the program stored in its own recording medium and executes the process according to the read program. Further, as another execution form of this program, the computer may read the program directly from the portable recording medium and execute processing according to the program, and further, the program is transferred from the server computer to this computer Each time, processing according to the received program may be executed sequentially. In addition, a configuration in which the above-described processing is executed by a so-called ASP (Application Service Provider) type service that realizes processing functions only by executing instructions and acquiring results from the server computer without transferring the program to the computer It may be Note that the program in the present embodiment includes information provided for processing by a computer that conforms to the program (such as data that is not a direct command to the computer but has a property that defines the processing of the computer).

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

10 learning data storage unit 11 feature amount extraction unit 12 attribute label generation unit 13 deep learning unit 14 clip sound synthesis unit 15 voiced voiceless determination unit 16 learning data posterior probability calculation unit 17 reliability parameter learning unit 20 DNN model storage unit 21 feature amount Extraction unit 22 Posterior probability calculation unit 23 Identification unit 24 Reliability calculation unit 30 Reliability parameter storage unit

Claims (13)

A speaker attribute estimation system including a learning device and an estimation device,
The above learning device
An attribute label creating unit for creating an attribute label series representing a speaker attribute for each frame of learning speech from attribute information representing a speaker attribute for each learning speech;
A clip sound synthesis unit that amplifies the amplitude of the learning speech and rounds the amplitude above a predetermined threshold to that threshold to synthesize clip sound;
A deep learning unit that learns a deep neural network model using the acoustic feature amount sequence for each frame of the learning speech and the clip sound and the attribute label sequence;
Including
The above estimation device is
A posterior probability calculation unit that calculates a posterior probability sequence for each frame from the acoustic feature amount sequence for each frame of the input speech using the deep neural network model;
An identification unit for identifying a speaker attribute based on a logarithmic sum of the above-mentioned posterior probability series obtained for each speaker attribute;
Speaker attribute estimation system including. The speaker attribute estimation system according to claim 1 ,
The above learning device
Further including a voiced voiceless determination unit generating voiced voiceless information indicating whether voiceed or unvoiced for each frame of the learning voice;
The attribute label creating unit sets the value of the attribute information for the voiced frame and sets the value indicating that the voiceless frame is unvoiced based on the voiced voiceless information, and creates the attribute label sequence. It is
Speaker attribute estimation system. A speaker attribute estimation system including a learning device and an estimation device,
The above learning device
An attribute label creating unit for creating an attribute label series representing a speaker attribute for each frame of learning speech from attribute information representing a speaker attribute for each learning speech;
A deep learning unit that learns a deep neural network model using the acoustic feature amount sequence for each frame of the learning speech and the attribute label sequence;
A learning data posterior probability calculation unit that calculates a posterior probability sequence for each frame from the acoustic feature amount sequence for each frame of the learning speech using the deep neural network model;
A reliability parameter learning unit that calculates a reliability parameter that represents the distribution of the posterior probability series of the learning speech;
Including
The above estimation device is
A posterior probability calculation unit that calculates a posterior probability sequence for each frame from the acoustic feature amount sequence for each frame of the input speech using the deep neural network model;
An identification unit for identifying a speaker attribute based on a logarithmic sum of the above-mentioned posterior probability series obtained for each speaker attribute;
A reliability calculation unit that calculates the reliability based on the distribution represented by the reliability parameter and the distribution of the posterior probability series of the input speech;
Speaker attribute estimation system including. A speaker attribute estimation system including a learning device and an estimation device,
The above learning device
An attribute label creating unit for creating an attribute label series representing a speaker attribute for each frame of learning speech from attribute information representing a speaker attribute for each learning speech;
A deep learning unit that learns a deep neural network model using the acoustic feature amount sequence for each frame of the learning speech and the attribute label sequence;
Including
The above estimation device is
A posterior probability calculation unit that calculates a posterior probability sequence for each frame from the acoustic feature amount sequence for each frame of the input speech using the deep neural network model;
An identification unit for identifying a speaker attribute based on a logarithmic sum of the above-mentioned posterior probability series obtained for each speaker attribute;
A reliability calculation unit that calculates a reliability based on the distribution of the posterior probability series of the input speech regarding the identified speaker attribute and the distribution of the posterior probability series of the input speech regarding the other speaker attributes;
Speaker attribute estimation system including. An attribute label creating unit for creating an attribute label series representing a speaker attribute for each frame of learning speech from attribute information representing a speaker attribute for each learning speech;
A clip sound synthesis unit that amplifies the amplitude of the learning speech and rounds the amplitude above a predetermined threshold to that threshold to synthesize clip sound;
A deep learning unit that learns a deep neural network model using the acoustic feature amount sequence for each frame of the learning speech and the clip sound and the attribute label sequence;
A learning device including An attribute label creating unit for creating an attribute label series representing a speaker attribute for each frame of learning speech from attribute information representing a speaker attribute for each learning speech;
A deep learning unit that learns a deep neural network model using the acoustic feature amount sequence for each frame of the learning speech and the attribute label sequence;
A learning data posterior probability calculation unit that calculates a posterior probability sequence for each frame from the acoustic feature amount sequence for each frame of the learning speech using the deep neural network model;
A reliability parameter learning unit that calculates a reliability parameter that represents the distribution of the posterior probability series of the learning speech;
A learning device including A posterior probability calculation unit that calculates a posterior probability sequence for each frame using a deep neural network model learned in advance from an acoustic feature amount sequence for each frame of the input speech;
An identification unit for identifying a speaker attribute based on a logarithmic sum of the above-mentioned posterior probability series obtained for each speaker attribute;
Including
The deep neural network creates an attribute label sequence representing speaker attributes for each frame of learning speech from attribute information representing speaker attributes for each learning speech, amplifies the amplitude of the learning speech, and exceeds a predetermined threshold. The amplitude is rounded to the threshold value to synthesize clip sound, and learning is performed using the learning voice and the acoustic feature amount series for each frame of the clip sound and the attribute label series.
Estimator. A posterior probability calculation unit that calculates a posterior probability sequence for each frame using a deep neural network model learned in advance from an acoustic feature amount sequence for each frame of the input speech;
An identification unit for identifying a speaker attribute based on a logarithmic sum of the above-mentioned posterior probability series obtained for each speaker attribute;
A reliability calculation unit that calculates the reliability based on the distribution represented by the reliability parameter calculated in advance and the distribution of the posterior probability sequence of the input speech;
Including
The deep neural network creates an attribute label sequence representing a speaker attribute of each frame of learning speech from attribute information representing a speaker attribute of each learning speech, and an acoustic feature amount sequence of each frame of the learning speech and the attribute all SANYO learned by using the label series,
The reliability parameter, Ru der represents the distribution of the posterior probability sequence for each frame calculated from acoustic features sequence of each frame of the training speech using the deep neural network model,
Estimator. A posterior probability calculation unit that calculates a posterior probability sequence for each frame using a deep neural network model learned in advance from an acoustic feature amount sequence for each frame of the input speech;
An identification unit for identifying a speaker attribute based on a logarithmic sum of the above-mentioned posterior probability series obtained for each speaker attribute;
A reliability calculation unit that calculates a reliability based on the distribution of the posterior probability series of the input speech regarding the identified speaker attribute and the distribution of the posterior probability series of the input speech regarding the other speaker attributes;
Including
The deep neural network creates an attribute label sequence representing a speaker attribute of each frame of learning speech from attribute information representing a speaker attribute of each learning speech, and an acoustic feature amount sequence of each frame of the learning speech and the attribute It is learned using the label sequence and
Estimator. The attribute label creation unit creates an attribute label series representing the speaker attribute for each frame of the learning speech from the attribute information representing the speaker attribute for each learning speech,
The clip sound synthesis unit amplifies the amplitude of the learning speech, and the amplitude exceeding a predetermined threshold is rounded to the threshold to synthesize the clip sound,
The deep learning unit learns the deep neural network model using the acoustic feature amount sequence for each frame of the learning speech and the clip sound and the attribute label sequence,
The posterior probability calculation unit calculates the posterior probability sequence for each frame using the deep neural network model from the acoustic feature amount sequence for each frame of the input speech,
The identification unit identifies the speaker attribute based on the logarithmic sum of the a posteriori probability series obtained for each speaker attribute.
Speaker attribute estimation method. The attribute label creation unit creates an attribute label series representing the speaker attribute for each frame of the learning speech from the attribute information representing the speaker attribute for each learning speech,
The deep learning unit learns the deep neural network model using the acoustic feature amount sequence for each frame of the learning speech and the attribute label sequence,
The learning data posterior probability calculation unit calculates a posterior probability sequence for each frame using the deep neural network model from the acoustic feature amount sequence for each frame of the learning speech,
The reliability parameter learning unit calculates a reliability parameter representing the distribution of the posterior probability series of the learning speech,
The posterior probability calculation unit calculates the posterior probability sequence for each frame using the deep neural network model from the acoustic feature amount sequence for each frame of the input speech,
The identification unit identifies the speaker attribute based on the logarithmic sum of the a posteriori probability series obtained for each speaker attribute ,
A reliability calculation unit calculates the reliability based on the distribution represented by the reliability parameter and the distribution of the posterior probability series of the input speech ;
Speaker attribute estimation method. The attribute label creation unit creates an attribute label series representing the speaker attribute for each frame of the learning speech from the attribute information representing the speaker attribute for each learning speech,
The deep learning unit learns the deep neural network model using the acoustic feature amount sequence for each frame of the learning speech and the attribute label sequence,
The posterior probability calculation unit calculates the posterior probability sequence for each frame using the deep neural network model from the acoustic feature amount sequence for each frame of the input speech,
The identification unit identifies the speaker attribute based on the logarithmic sum of the a posteriori probability series obtained for each speaker attribute ,
A confidence calculation unit calculates the confidence based on the distribution of the posterior probability sequence of the input speech regarding the identified speaker attributes and the distribution of the posterior probability sequence of the input speech regarding the other speaker attributes .
Speaker attribute estimation method. A program for causing a computer to function as the learning device according to claim 5 or 6 , or the estimation device according to any one of claims 7 to 9 .