Training method and target signal estimation device

Abstract

The present invention aims to improve the accuracy of target signal estimation by a model without increasing the number of parameters of the model. To this end, a target signal estimation device according to the present disclosure defines the parameters of a model as follows. A target signal estimation unit uses the model to recursively determine an nth provisional target signal from an input signal and an (n - 1)th provisional target signal. Among variables, a parameter updating unit includes a correct target signal and an nth provisional target signal with respect to each among n = 1, ..., N, and the parameter updating unit determines gradients regarding parameters of the model on the basis of an evaluation function which evaluates the differences between the correct target signals and the nth provisional target signals. For each parameter of the model, the parameter updating unit determines a cumulative gradient by weighting and adding together N obtained gradients, and the parameter updating unit updates the parameters of the model using the cumulative gradients.

Description

Learning method and target signal estimation device

[0001] The disclosed technology relates to a target signal estimation device and a learning method for the device, which estimates a target signal representing the characteristics of a target sound from an acoustic signal obtained by recording the target sound in an environment with background noise and reverberation.

[0002] Figure 1 shows a conventional target signal estimation device (for example, Complex Spectral Mapping (CSM)-based Speech Enhancement described in Non-Patent Document 1). The target signal estimation device 1 comprises a target signal estimation unit 101, a parameter storage unit 102, and a parameter update unit 103. The operation of the target signal estimation device 1 will be explained using Figure 1.

[0003] <Estimation of target signal (Figure 1(b))> The target signal estimation device 1 receives an acoustic signal with the target signal estimation unit 101 and outputs an estimated target signal using processing parameters read from the parameter storage unit 102. <Learning of target signal estimation unit (Figure 1(a))> The target signal estimation device 1 receives an acoustic signal with the target signal estimation unit 101 and outputs an estimated target signal using processing parameters read from the parameter storage unit 102. The parameter update unit 103 receives the correct target signal and the estimated target signal, reads the processing parameters from the parameter storage unit 102, finds updated processing parameters so that the error between the correct target signal and the estimated target signal is small, and writes them to the parameter storage unit 102.

[0004] T. Nakatani et al., "MULTI-STREAM DIFFUSION MODEL FOR PROBABILISTIC INTEGRATION OF MODEL-BASED AND DATA-DRIVEN SPEECH ENHANCEMENT", 18th International Workshop on Acoustic Signal Enhancement (IWAENC), 2024.

[0005] In the target signal estimation device 1 described above, the accuracy of target signal estimation is determined by the capabilities of the target signal estimation unit, and therefore, it was not possible to achieve higher accuracy. For example, it may be possible to improve the ability to estimate the target signal by making changes such as increasing the number of processing parameters used by the target signal estimation unit 101. However, this processing requires a larger memory area as a computational resource, which presents a problem. In general, in processes that can learn input / output relationships, such as neural networks, the memory area required for processing increases significantly with increasing numbers of parameters. Furthermore, when the number of parameters increases, there is a problem that more training data is required to perform high-precision learning.

[0006] To solve the above problems, the objective signal estimation device related to the disclosed technology uses a model capable of learning input-output relationships from training data. The objective signal estimation device comprises an objective signal estimation unit and a parameter update unit. The input acoustic signal, or acoustic features obtained from the input acoustic signal, are used as the 0th provisional objective signal, and the model parameters are defined as follows. As n=1,...,N, the objective signal estimation unit uses the model to recursively determine the nth provisional objective signal from the input signal and the (n-1)th provisional objective signal. For each of n=1,...,N, the parameter update unit includes the correct objective signal and the nth provisional objective signal as variables and calculates the gradient with respect to the model parameters based on an evaluation function that evaluates the difference between the correct objective signal and the nth provisional objective signal. The parameter update unit obtains a cumulative gradient by weighting and adding the N gradients obtained for each parameter of the model. The parameter update unit updates the model parameters using the cumulative gradient.

[0007] According to the disclosed technology, the overall accuracy of target signal estimation can be improved without increasing the number of processing parameters, and therefore without increasing the memory area required for processing, nor without increasing the training data required for learning to perform high-precision processing.

[0008] A prior art target signal estimation device. A diagram illustrating the tensor quantity of a complex signal. A functional block diagram of the target signal estimation device according to the first embodiment. A flowchart illustrating the estimation process according to the first embodiment. A flowchart illustrating the overall learning processes according to the first, second, and third embodiments. A flowchart illustrating the details of the parameter update process according to the first embodiment. A functional block diagram of the target signal estimation device according to the second embodiment. A detailed functional block diagram of the target signal estimation unit according to the second embodiment. A flowchart illustrating the estimation process according to the second embodiment. A flowchart illustrating the details of the parameter update process according to the second embodiment. A functional block diagram of the target signal estimation device according to the third embodiment. A detailed functional block diagram of the target signal estimation unit according to the third embodiment. A flowchart illustrating the estimation process according to the third embodiment. A flowchart illustrating the details of the parameter update process according to the third embodiment. A flowchart illustrating the details of the initialization process and recursive calculation process according to the third embodiment. A diagram assisting in the explanation of the estimation process and learning process according to the first embodiment. A diagram showing an example of the functional configuration of a computer.

[0009] The target signal estimation method related to the disclosed technology consists of a process (such as a neural network) capable of learning input-output relationships from training data, and the estimated target signal is defined by the training signal used during training. The target signal may be defined as the target sound itself, or as another signal representing the characteristics of the target sound. Embodiments of the disclosed technology will be described in detail below. Components with the same function will be given the same number, and redundant explanations will be omitted.

[0010] [Definition of Symbols] First, the symbols used in this specification will be explained. Equation (1) represents the entire set of three-dimensional complex signals whose elements are complex numbers with width W, height H, and number of channels C. Equation (2) shows that a is an element of the set shown in equation (1). Figure 2 shows the shape of a. W × H together is also called the signal size. Equation (3) represents the universal set of three-dimensional real number signals whose elements are real numbers with width T, height F, and number of channels C. In equation (4), y represents the recorded acoustic signal. y is the tensor quantity of the complex signal obtained by frequency-dividing the acoustic signal (e.g., using a short-time Fourier transform) and arranging the results in the channel direction. T is the number of time frames, and F is the number of frequency bins. x in equation (5) n This represents the nth provisional target signal. In equation (6), s and s^ represent the correct target signal and the estimated target signal.

[0011] When each processing unit handles real-valued signals rather than complex signals, the complex signal y (and other complex signals) is converted to an equivalent real-valued signal y (equation (7)) obtained by decomposing the elements of each signal into real and imaginary parts and arranging them in the channel direction. In the following text, complex signals and real signals will not be distinguished.

[0012] [First Embodiment] In the first embodiment, the target sound itself is estimated as the target signal. Figure 3 is a functional block diagram showing an example of the configuration of the target signal estimation device according to the first embodiment. The target signal estimation device 3 comprises a target signal estimation unit 301, a parameter storage unit 302, and a parameter update unit 303.

[0013] <Target signal estimation unit 301> Acoustic signal y and provisional target signal x n-1 Upon receiving the provisional target signal x n It consists of a neural network that outputs y and x. n-1 Received x n Any neural network capable of generating the above is acceptable. In the first embodiment, as an example, mNCSN++ described in Non-Patent Document 1 is used. In this specification, neural networks may be abbreviated as NN.

[0014] <Parameter Storage Unit 302> Stores all processing parameters of the neural network, such as weight coefficients.

[0015] <Parameter update unit 303> Acoustic signal y i and the correct target signal s iFor each of a number of training data (where i is the number of the training data, i = 1, 2, …, I) consisting of the set, the gradient of the error function defined below is obtained. The error function D(s i , x1 i , …, x N i ) is defined as the weighted sum with α n as the weight coefficient, with respect to the error D(s i , x n i ) for each tentative target signal x ) is defined as the weighted sum with α n as the weight coefficient, with respect to the error D(s i , x n i ) for each tentative target signal x n i of the error D(s i , x n i ). The error D(s i , x n i ) can use various ones, such as the mean squared error and the error used in the learning of the CSM-based SE of non-patent literature. For each application of the target signal estimation unit, a set of gradients g ) can use various ones, such as the mean squared error and the error used in the learning of the CSM-based SE of non-patent literature. For each application of the target signal estimation unit, a set of gradients g n i regarding the processing parameters of D(s i , x n i ) is obtained using the error backpropagation method or the like, and a set of gradients g i regarding the processing parameters of the above error function is obtained as the weighted sum of g n i . Using the gradient g i obtained for each training data, the processing parameters are repeatedly updated using a method such as the stochastic gradient method.

[0016] FIG. 4 is a flowchart for explaining the estimation process of the target signal. FIG. 5 is a flowchart for explaining the learning of the target signal estimation device. FIG. 6 is a detailed flowchart of the "parameter update process" in FIG. 5.

[0017] <Estimation of the target signal> This will be described using FIGS. 3(b) and FIG. 4. The target signal estimation unit 301 acquires the processing parameters of the neural network (step S401). The target signal estimation unit 301 acquires the acoustic signal y (step S402) and sets the tentative target signal x0 = y (step S403).

[0018] Next, x N nThe calculation is repeated from n=1 to N. Therefore, the target signal estimation device 3 initializes the value of n to 1 (step S404). The target signal estimation unit 301 calculates y and x n-1 from x n The target signal estimation device 3 calculates (step S405). The target signal estimation device 3 increments the value of n by 1 (step S408) and determines if n exceeds N. If n > N (No in step S409), it returns to step S405. If n > N (Yes in step S409), the target signal estimation device 3 calculates x N This is output as the estimated target signal (step S410).

[0019] The above is a description of the "target signal estimation process" according to the first embodiment. To aid in understanding the estimation process of the first embodiment, a provisional target signal x n Figure 16(b) shows how this is calculated recursively.

[0020] <Neural Network Learning> This will be explained using Figures 3(a), 5, and 6. The target signal estimation device 3 repeats the following process until learning converges.

[0021] The target signal estimation device 3 uses the numbers Φ={1,2,...,I} of all the training data to determine several mini-batches Φ1, Φ2,..., Φ B Divided into (step S501). The target signal estimation device 3 measures each minibatch Φ b For (b=1,2,…,B), the following is performed: The target signal estimation unit 301 obtains the neural network processing parameters from the parameter storage unit 302 (step S502). The parameter update unit 303 obtains the neural network processing parameters from the parameter storage unit 302 (step S503). The parameter update unit 303 performs an update process for the processing parameters (step S504). Details of the process in step S504 will be described later. The parameter update unit 303 records the updated processing parameters in the parameter storage unit 302 (step S505).

[0022] The target signal estimation device 3 determines the convergence of learning (step S506). For example, the following criteria may be used to determine the convergence of learning: (1) The number of learning iterations has reached a certain number. (2) An error function is calculated for the development data, which consists of a pair of acoustic signals not included in the learning data and the correct target signal, and the decrease in error has fallen below a threshold. (3) The estimated target signal obtained from the above development data is evaluated using some signal quality criterion (e.g., signal-to-distortion ratio), and the improvement in the evaluation value has fallen below a threshold.

[0023] <Parameter Update Process> The details of the parameter update process will be explained using Figures 3(a) and 6. The target signal estimation device 3 uses the learning data i∈Φ b In response, the following will be implemented:

[0024] The parameter update unit 303 updates the correct target signal s i The target signal estimation unit 301 obtains the acoustic signal y i Obtain (step S602), provisional target signal x0 i =y i (Step S603)

[0025] Next, x n i The calculation is repeated from n=1 to N. For this reason, the target signal estimation device 3 initializes the value of n to 1 (step S604). The target signal estimation unit 301 calculates y i and x n-1 i from x n i The parameter update unit 303 calculates x n i The parameter update unit 303 obtains the error function D(s i ,x n i ) A set of gradients g n iThe parameter update unit 303 calculates (step S609). The target signal estimation device 3 increments the value of n by 1 (step S610) and determines whether n has exceeded N. If n > N (No in step S611), the process returns to step S605. If n > N (Yes in step S611), the parameter update unit 303 calculates g n i Weighted sum of g i Calculate (step S612).

[0026] For all i∈Φ b After performing the above processing on g, the parameter update unit 303 will i We calculate g by adding up to all the training data in the minibatch (step S613, equation (10)). The parameter update unit 303 updates the processing parameters using g (step S614).

[0027] The above is a description of the learning process according to the first embodiment. To aid in understanding the learning process of the first embodiment, the provisional target signal x n i Figure 16(a) shows how the recursive calculation is performed and used for parameter updates.

[0028] The above is a description of the first embodiment.

[0029] [Second Embodiment] In the first embodiment, the target sound itself was estimated as the target signal. In the second embodiment, the distribution parameters of a Gaussian mixture (GMM) representing the distribution of the target sound are estimated as the target signal.

[0030] <GMM> GMM is defined as shown in formula (11) below. Here, This is the target sound. is an element of s that corresponds to the time frame t. This is the provisional target signal (GMM distribution parameters) output by the target signal estimation unit on the nth iteration. is, Θ n These are the elements corresponding to the time frame t. K and k are the GMM mixture number and the number of each element. is, Θ n,tThis is the parameter of the k-th Gaussian distribution included in [the specified variable]. is, Θ n,t,k These are the mixing ratios, mean, and covariance matrix included in the formula. This is a complex mixture Gaussian distribution defined by the following equation (19). Here t,n,k Let be a diagonal matrix, σ t,f,n,k Θ is the f-th diagonal element, where f represents the frequency number. (Provisional target value Θ) 0,t For example, consider an acoustic signal y with a time frame t and a frequency f element y. t,f Set the settings as follows:

[0031] Next, the configuration of the target signal estimation device 7 will be described. Figure 7 is a functional block diagram showing an example of the configuration of the target signal estimation device according to the second embodiment. The target signal estimation device 7 comprises an initialization unit 704, a target signal estimation unit 701, a parameter storage unit 702, and a parameter update unit 703. Figure 8 is a detailed functional block diagram of the target signal estimation unit 701. The target signal estimation unit 701 comprises a sampling unit 711, a neural network 712, and a GMM parameter configuration unit 713.

[0032] <Target Signal Estimation Unit 701> The target signal estimation unit 701 takes the acoustic signal y and the provisional target value Θ n-1 Upon receiving the provisional target value Θ n The output is Θ. The sampling unit 711 outputs Θ n-1 Received, Θ n-1 From the mixture Gaussian distribution determined by this, J signal realizations The neural network 712 samples the acoustic signal y and J signal realizations X. n,j The system receives the signal and outputs 3,000 time-series signals according to equation (22) below. The neural network 712 uses y and X n,1 ,X n,2 ,…,X n,J Received, m n ,v n ,w nAny neural network may be used as long as it can generate the following. In the second embodiment, as an example, mNCSN++ described in Non-Patent Document 1 is used.

[0033] The GMM parameter configuration unit 713 receives the above time-series signal and converts it into, for example, parameters of GMM and outputs them as follows. Here, is the element of the time frame t of m n,k , w t,f,n,k and v t,f,n,k are w n and v n of the time frame t and frequency f, and ε = 10 -5 is a positive flooring constant.

[0034] <Parameter storage unit 702> Stores all processing parameters of the above neural network, such as weight coefficients.

[0035] <Parameter update unit 703> For each of a number of learning data (i is the number of learning data, i = 1,..., I) consisting of a pair of an acoustic signal y i and a correct target signal s i , the gradient of the negative log-likelihood function defined below is obtained. The negative log-likelihood function L(s i , Θ1 i ,..., Θ N i ) is defined as a weighted sum with α n as a weight coefficient for the negative log-likelihood L(s i , Θ n i ) of each provisional target signal Θ n i . Here, p(s i ; Θ n i ) is a probability distribution according to GMM. For each application of the target signal estimation unit, a set of gradients g i n i regarding the processing parameters of L(s i , Θ n i ) is obtained using the error backpropagation method or the like, and a set of gradients g i regarding the processing parameters of the above error function is set to gn i The weighted sum is obtained (Equation (9)). The gradient g is obtained for each training data. i Using this method, the processing parameters are repeatedly updated using methods such as stochastic gradient descent.

[0036] Next, the operation of the target signal estimation device 7 will be explained. Figure 9 is a flowchart illustrating the target signal estimation process according to the second embodiment. Figure 5 is a flowchart illustrating the learning process of the target signal estimation device according to the second embodiment. Figure 10 is a detailed flowchart of the "parameter update process" according to the second embodiment.

[0037] <Estimation of the target signal> This will be explained using Figures 7(b) and 9. Steps S401 and S402 are the same as in the first embodiment. Subsequently, the target signal estimation device 7 generates a provisional target signal Θ0 from the acoustic signal y in the initialization unit 704 (step S903).

[0038] Next, Θ n The calculation is repeated from n=1 to N. Therefore, the target signal estimation device 7 initializes the value of n to 1 (step S404). The sampling unit 711 then... n-1 From X n,j (Step S905) generates (j=1,2,,J). The neural network 712 then processes y and X. n,1 ,X n,2 ,…,X n,J From m n ,v n ,w n The GMM parameter component 713 calculates m n ,v n ,w n From Θ n The system generates (step S907). The target signal estimation device 7 increments the value of n by 1 (step S408) and determines whether n has exceeded N. If n > N (No in step S409), the system returns to step S905. If n > N (Yes in step S409), the target signal estimation device 7 generates Θ N This is output as the estimated target signal (step S910).

[0039] The above is a description of the "target signal estimation process" according to the second embodiment.

[0040] <Neural Network Learning> This will be explained using Figures 7(a), 5, 8, and 10. Similar to the first embodiment, the target signal estimation device 7 repeats the process in Figure 5 until the learning converges.

[0041] <Parameter Update Process> The details of the parameter update process according to the second embodiment will be explained using Figures 7(a) and 10. The target signal estimation device 7 uses the learning data i∈Φ b In response, the following will be implemented:

[0042] Steps S601 and S602 are the same as in the first embodiment. The target signal estimation device 7 uses the initialization unit 704 to perform y i From the provisional objective signal Θ0 i Generate (step S1003).

[0043] Next, Θ n i The calculation is repeated from n=1 to N. Therefore, the target signal estimation device 7 initializes the value of n to 1 (step S604). The sampling unit 711 of the target signal estimation unit 701 is Θ n-1 i From X n,j (j=1,2,,J) is generated (step S1005). The neural network 712 then generates y i TotoX n,1 ,X n,2 ,…,X n,J From m n ,v n ,w n The GMM parameter component 713 calculates m n ,v n ,w n From Θ n i The parameter update unit 703 generates Θ (step S1007). n i Obtain the negative log-likelihood function L(s) (step S1008), i ,Θ n i ) A set of gradients g ni The calculation is performed (step S1009). The target signal estimation device 7 increments the value of n by 1 (step S610) and determines whether n has exceeded N. If n > N (No in step S611), the process returns to step S1005. If n > N (Yes in step S611), the process proceeds to step S612. Steps S612, and steps S613 and S614 after mini-batch learning, are the same as in the first embodiment.

[0044] The above is a description of the "learning process" according to the second embodiment.

[0045] The above describes the second embodiment. Note that the estimated target signal s obtained in the second embodiment is t ^=Θ N Using a Gaussian mixture distribution determined by the above, the estimated value of the target sound s is taken as the mean value of that distribution. t ~ can be calculated as follows.

[0046] [Third Embodiment] In the third embodiment, both the target signal of the first embodiment and the target signal of the second embodiment are estimated using an acoustic signal y as input.

[0047] Figure 11 is a functional block diagram showing an example configuration of a target signal estimation device according to the third embodiment. The target signal estimation device 11 comprises an initialization unit 1104, a target signal estimation unit 1101, a parameter storage unit 1102, and a parameter update unit 1103. Figure 12 is a detailed functional block diagram of the target signal estimation unit 1101. The target signal estimation unit 1101 comprises a sampling unit 1111, a neural network 1112, and a GMM parameter configuration unit 1113.

[0048] <Target Signal Estimation Unit 1101> The target signal estimation unit 1101 takes the acoustic signal y and the provisional target value x n-1 and Θ n-1 Received, provisional target value x n and Θ n The output is the same as in the second embodiment. The function of the sampling unit 1111 is the same as in the second embodiment. The neural network 1112, like the neural network in the first embodiment, uses the acoustic signal y and the provisional target signal x n-1 Upon receiving the provisional target signal x nIt outputs the following: Also, similar to the neural network 712 of the second embodiment, it outputs an acoustic signal y and J signal realization values ​​X. n,j Received, m n ,v n ,w n It outputs the following. The function of the GMM parameter component 1113 is the same as in the second embodiment.

[0049] <Parameter Storage Unit 1102> Stores all processing parameters of the neural network, such as weight coefficients.

[0050] <Parameter update unit 1103> Acoustic signal y i and the correct target signal s i For each of the numerous training data sets consisting of pairs (where i is the training data number, i=1,2,…,I), we find the gradient of the error function defined below. Optimization function O(s i ,x1 i ,…,x N i ,Θ1 i ,…,Θ N i ) to each provisional target signal x n i Error D(s) i ,x n i ) and each provisional objective signal Θ n i The negative log-likelihood L(s) i ,Θ n i ) concerning α n It is defined as a weighted sum with weight coefficients (Equation (27)). Each time the target signal estimation unit is applied, D(s i ,x n i ) + L(s i ,Θ n i A set of gradients g related to the processing parameters of ) n i This is obtained using methods such as backpropagation, and the set of gradients g with respect to the processing parameters of the above error function is obtained. i to, g n i The weighted sum is obtained (Equation (9)). The gradient g is obtained for each training data.i Using this method, the processing parameters are repeatedly updated using methods such as stochastic gradient descent.

[0051] Next, the operation of the target signal estimation device 11 will be explained. Figure 13 is a flowchart illustrating the target signal estimation process according to the third embodiment. Figure 5 is a flowchart illustrating the learning process of the target signal estimation device according to the third embodiment. Figures 14 and 15 are detailed flowcharts of the "parameter update process" according to the third embodiment.

[0052] <Estimation of the target signal> This will be explained using Figures 11(b) and 13. Steps S401 and S402 are the same as in the first embodiment. Subsequently, the target signal estimation device 11 performs step S403 of the first embodiment and step S903 of the second embodiment.

[0053] Next, the target signal estimation unit 1101 performs x n and Θ n Repeat the calculation from n=1 to N. n The calculation is as in the first embodiment and Θ n The calculation is the same as in the second embodiment. Finally, the target signal estimation device 11 determines x N and Θ N This is output as the estimated target signal (step S1310).

[0054] The above is a description of the "target signal estimation process" according to the third embodiment.

[0055] <Neural Network Learning> This will be explained using Figures 11(a), 5, 12, 14, and 15. Similar to the first embodiment, the target signal estimation device 11 repeats the process in Figure 5 until the learning converges.

[0056] <Parameter Update Process> The details of the parameter update process according to the third embodiment will be explained using Figures 11(a), 12, 14, and 15. The target signal estimation device 11 uses the learning data i∈Φ b In response, the following will be implemented:

[0057] Steps S601 and S602 are the same as in the first embodiment. The target signal estimation device 11 performs an initialization process (step S1303).i The generation of the first embodiment and Θ0 i The generation is the same as in the second embodiment (Figure 15(a)).

[0058] Next, the target signal estimation unit 1101 performs x n i and Θ n i The calculation is repeated from n=1 to N. Therefore, the target signal estimation device 11 initializes the value of n to 1 (step S604). Next, the target signal estimation device 11 performs recursive calculation processing (step S1305). x n i The calculation is as in the first embodiment and Θ n i The calculation is the same as in the second embodiment (Figure 15(b)). Next, the parameter update unit 1103 calculates D(s i ,x n i ) + L(s i ,Θ n i ) A set of gradients g n i The calculation is performed (step S1409). The target signal estimation device 11 increments the value of n by 1 (step S610) and determines whether n has exceeded N. If n > N (No in step S611), the process returns to step S1305. If n > N (Yes in step S611), the process proceeds to step S612. Steps S612, and steps S613 and S614 after mini-batch learning, are the same as in the first embodiment.

[0059] The above is a description of the "learning process" according to the third embodiment.

[0060] The above is a description of the third embodiment.

[0061] [Confirmation of the effects of the present invention through experiments] Noise suppression and reverberation suppression were performed on acoustic signals recorded in an environment with noise and reverberation. In the "conventional method," speech enhancement was performed using the complex spectral mapping method CSM with NCSN++. In the "third embodiment," the third embodiment was implemented using NCSN++ and speech enhancement was performed. From Table 1, it can be confirmed that the signal-to-distortion ratio (SDR) was improved in the third embodiment compared to the conventional method. The parameter size of the neural network (NCSN++) used in the experiment was 65.6 MB for the conventional method and 65.7 MB for the third embodiment. By using the third embodiment, a significant improvement in SDR was obtained without significantly increasing the parameter size of the neural network.

[0062] [Supplement] In each embodiment, a neural network was used in the target signal estimation unit. However, any model capable of learning input / output relationships from training data can be used instead of a neural network.

[0063] [Programs, Recording Media] The functions realized by the components described herein may be implemented in a circuit or processing circuitry, including a general-purpose processor, an application-specific processor, an integrated circuit, an ASIC (Application Specific Integrated Circuit), a CPU (a Central Processing Unit), conventional circuits, and / or a combination thereof, programmed to realize the functions described herein. A processor is considered to be a circuit or processing circuitry, including transistors and other circuits. A processor may be a programmed processor that executes a program stored in memory.

[0064] In this specification, circuitry, unit, and means are hardware programmed to perform or execute the functions described herein. Such hardware may be any hardware disclosed herein, or any hardware known to be programmed to perform or execute the functions described herein.

[0065] If the hardware is a processor that is considered to be a type of circuitry, then the circuitry, means, or unit is a combination of hardware and software used to constitute the hardware and / or processor.

[0066] The various processes described above can be carried out by loading a program that executes each step of the above method into the recording unit 2020 of the computer 2000 shown in Figure 17, and then causing the control unit 2010, input unit 2030, output unit 2040, display unit 2050, etc. to operate.

[0067] The program describing this process can be recorded on a computer-readable recording medium. Any computer-readable recording medium can be used, such as a magnetic recording device, optical disc, magneto-optical recording medium, or semiconductor memory.

[0068] Furthermore, this program may be distributed, for example, by selling, transferring, or lending portable recording media such as DVDs or CD-ROMs on which the program is recorded. Alternatively, the program may be stored in the storage device of a server computer and distributed by transferring the program from the server computer to other computers via a network.

[0069] A computer executing such a program may, for example, first store the program, either recorded on a portable storage medium or transferred from a server computer, in its own memory. Then, when processing is to be executed, the computer reads the program stored in its memory and executes the processing according to the read program. Alternatively, the computer may directly read the program from the portable storage medium and execute the processing according to that program, or it may sequentially execute the processing according to the received program each time a program is transferred to it from a server computer. Furthermore, the processing may be executed using a so-called ASP (Application Service Provider) type service, where the processing function is realized only by issuing execution instructions and obtaining results, without transferring the program from the server computer to this computer.In addition, the processing may be executed using a so-called SaaS (Software as a Service) type service, where a part of the server computer is made available to the user along with the program. Furthermore, the aforementioned programs include information used for processing by electronic computers that is equivalent to a program (data, etc., that is not a direct instruction to the computer but has the property of defining the computer's processing).

[0070] Furthermore, although the above explanation assumes that the device is configured by executing a predetermined program on a computer, at least a part of these processes may be implemented in hardware.

Claims

1. A model for estimating a target signal, and a method for learning a model capable of learning input-output relationships from training data, wherein the model takes a signal derived from an acoustic signal and a first provisional target signal as inputs and outputs a second provisional target signal, takes a signal derived from the acoustic signal and the second provisional target signal as inputs and outputs a third provisional target signal, and uses a function that evaluates the difference between the correct target signal and the provisional target signals as an evaluation function, and updates the parameters of the model using the gradient of the evaluation function obtained from the second provisional signal and the gradient of the evaluation function obtained from the third provisional signal.

2. A model for estimating a target signal, and a learning method for a model capable of learning input-output relationships from training data, wherein the input acoustic signal, or acoustic features obtained from the input acoustic signal, is the 0th provisional target signal, and n=1,...,N, the model is used to recursively obtain the nth provisional target signal from the input signal and the (n-1)th provisional target signal, for each of n=1,...,N, the gradient with respect to the model parameters is obtained based on an evaluation function that includes the correct target signal and the nth provisional target signal as variables and evaluates the difference between the correct target signal and the nth provisional target signal, the N gradients obtained for each parameter of the model are weighted and added together to obtain an accumulated gradient, and the model parameters are updated using the accumulated gradient.

3. A target signal estimation device that uses a model for estimating a target signal, and is capable of learning input / output relationships from training data, comprising: a target signal estimation unit and a parameter update unit, wherein the input acoustic signal, or acoustic features obtained from the input acoustic signal, is used as the 0th provisional target signal, and the parameters of the model are n=1,...,N, the target signal estimation unit recursively obtains the nth provisional target signal from the input signal and the (n-1)th provisional target signal using the model, the parameter update unit obtains a gradient with respect to the parameters of the model for each of n=1,...,N, based on an evaluation function that includes the correct target signal and the nth provisional target signal as variables and evaluates the difference between the correct target signal and the nth provisional target signal, the parameter update unit obtains an accumulated gradient by weighting and adding the N gradients obtained for each parameter of the model, and the parameter update unit updates the parameters of the model using the accumulated gradient.

Images