Target speaker extraction training system, target speaker extraction training method, and program

Abstract

The present invention realizes a target speaker extraction device which accomplishes target speaker extraction regardless of variations in instances of pre-registered speech. The present invention stores, in a training data set storage unit, teaching voices, mixed sounds including the teaching voices, and a plurality of instances of pre-registered speech that are acquired from the same speaker as a person of the teaching voices. A speaker expression extraction unit generates, from the instances of pre-registered speech, one piece of speaker feature information for each of the instances of pre-registered speech in a neural network 1. A target speaker extraction unit generates one estimated teaching voice from the speaker feature information and the mixed sounds for each of the pieces of speaker feature information in a neural network 2. A data selection unit computes, for each of the estimated teaching voices, the differences 1 between the estimated teaching voices and the teaching voices, and identifies the most inferior estimated teaching voice which has the largest difference 1. The optimization unit computes the difference 2 between the teaching voices and the most inferior estimated teaching voice, adjusts parameters of the neural network 1 and the neural network 2, and minimizes the difference 2.

Description

[Technical Field]

[0001] The present invention relates to an acoustic signal processing technology for extracting only the speech of a target speaker from speeches of multiple people or speech mixed with noise, as well as a learning method for a machine learning model or deep learning model that realizes this technology. [Background technology]

[0002] With the development of deep learning technology, the accuracy of speech recognition by computers has improved, and the scope of application of speech recognition is gradually expanding. However, when the speech of the target speaker is included in a speech that is produced by multiple overlapping voices, it is still difficult for a computer to recognize the target speaker's voice. One approach to dealing with overlapping speech is a method called target speaker extraction. Target speaker extraction is a technology that uses auxiliary information about the target speaker to extract only the desired target speaker's speech from speech from multiple speakers, including the target speaker's speech, or speech from the target speaker superimposed with noise. The most common auxiliary information for the target speaker is utterances that have been registered in advance by that speaker (Non-Patent Document 1).

[0003] Figure 1 shows a target speaker extraction device according to Non-Patent Document 1. Mixed sounds Y, which are sounds of multiple speakers including the target speaker's voice or sounds in which noise has been superimposed on the target speaker's voice, and sounds of only the target speaker (recorded data of about 10 seconds) are input, and the target speaker's voice is extracted from the mixed sounds Y and output. 101 denotes a speaker expression extraction unit, which has a pre-trained neural network and detects speaker features from the target speaker's voice data. 102 denotes a target speaker extraction unit, which has a pre-trained neural network and extracts the target speaker's voice from the mixed sounds Y using the target speaker features acquired from the speaker expression extraction unit 101.

[0004] Figure 2 shows a conventional target speaker extraction training system 20 described in Non-Patent Document 1. The speaker expression extraction unit and target speaker extraction unit are trained using pre-registered utterances, approximately 6 seconds long, recorded of a mixture of sounds Y and the target speaker alone, as training data, and the target speaker's speech to be extracted as training data. A large number of pre-registered utterances are prepared for each mixture of sounds Y, and one utterance is randomly selected from the pre-registered utterances during training. The optimization unit 201 compares the output of the target speaker extraction unit with the training data and repeatedly adjusts the network parameters of the speaker expression extraction unit and target speaker extraction unit until a predetermined stopping condition is met. The speaker expression extraction unit can be configured using machine learning or a deep neural network. The target speaker extraction unit can be configured using a deep neural network. [Prior art documents] [Non-patent literature]

[0005] [Non-Patent Document 1] Zmolikova, Katerina, et al. "SpeakerBeam: Speaker Aware Neural Network for Target Speaker Extraction in Speech Mixtures," IEEE Journal of Selected Topics in Signal Processing, Vol.13, No.4, August 2019, pp.800-814. Summary of the Invention [Problem to be solved by the invention]

[0006] As mentioned above, target speaker extraction technology extracts the speech of a desired speaker from a mixed sound using auxiliary information (pre-registered utterances) that are pre-registered utterances of the target speaker. In other words, speech is extracted using the commonalities between the target speaker's speech to be extracted and the pre-registered utterances. However, because tone and intonation vary even among the same speaker depending on the situation and context in which the utterance was made, even using pre-registered utterances can result in extraction errors, such as not being able to extract the speech of the desired speaker or extracting the speech of a different speaker.

[0007] The results of the study showed that target speaker extraction may or may not be successful depending on the type of utterance selected as the pre-registered utterance. It is an important practical requirement that target speaker extraction works regardless of variations in pre-registered utterances (speech situation, context, tone, intonation). In Non-Patent Document 1, in order to acquire robustness against variations in pre-registered utterances, a large number of pre-registered utterances are prepared for one type of mixed signal during training, and one utterance is randomly selected from these pre-registered utterances. This is expected to increase the variations in the combinations of mixed signals and pre-registered utterances during training, and to acquire robustness against the selection of pre-registered utterances. However, in reality, there are pre-registered utterances that are easy to extract the target speaker (the extraction result is close to the speech to be extracted) and pre-registered utterances that are difficult to extract (the extraction result is far from the speech to be extracted). This is thought to be because pre-registered utterances are selected with a certain probability during training, and both pre-registered utterances that are easy to extract and those that are difficult to extract are trained in the same way. By placing more emphasis on pre-registered utterances that are difficult to extract in training, it may be possible to further improve robustness against the selection of pre-registered utterances. [Means for solving the problem]

[0008] In order to solve the above problem, one embodiment of the present invention provides a target speaker extraction learning system comprising: a training dataset storage unit that stores teacher speech, a mixed sound containing the teacher speech, and a plurality of pre-registered utterances obtained from the same speaker as the teacher speech; a speaker expression extraction unit that includes a neural network 1 that generates one piece of speaker feature information for each of the pre-registered utterances from the pre-registered utterances; a target speaker extraction unit that includes a neural network 2 that generates one piece of estimated teacher speech for each of the speaker feature information from the speaker feature information and the mixed sound; a data selection unit that calculates a difference 1 between the estimated teacher speech and the teacher speech for each of the estimated teacher speeches and identifies the worst estimated teacher speech that gives the largest difference 1; and an optimization unit that calculates a difference 2 between the teacher speech and the worst estimated teacher speech and adjusts parameters of both neural network 1 and neural network 2 or only neural network 2 to minimize difference 2. [Effects of the Invention]

[0009] According to the target speaker extraction training system and training method of the present invention, it is possible to improve the robustness of the target speaker extraction device with respect to the selection of pre-registered utterances. [Brief explanation of the drawings]

[0010] [Figure 1] FIG. 1 is a diagram illustrating the target speaker extraction device of Non-Patent Document 1. [Figure 2] FIG. 2 is a diagram illustrating the target speaker extraction training system of Non-Patent Document 1. [Figure 3] FIG. 3 is a diagram illustrating the learning dataset construction unit and input / output values. [Figure 4] FIG. 4 is a diagram illustrating the procedure for constructing a learning dataset. [Figure 5] FIG. 5 is a diagram illustrating a target speaker extraction and training system according to the first embodiment. [Figure 6] FIG. 6 is a diagram illustrating a target speaker extraction learning procedure according to the first embodiment. [Figure 7]FIG. 7 is a diagram illustrating a target speaker extraction and training system according to the second embodiment. [Figure 8] FIG. 8 is a diagram illustrating a target speaker extraction learning procedure according to the second embodiment. [Figure 9] FIG. 9 is a diagram illustrating a target speaker extraction and training system according to the third embodiment. [Figure 10] FIG. 10 is a diagram illustrating a target speaker extraction learning procedure according to the third embodiment. [Figure 11] FIG. 11 is a diagram comparing the performance of a target speaker extraction device configured using the conventional method with the performance of a target speaker extraction device configured using the first and second embodiments. [Figure 12] FIG. 12 is a supplementary figure showing the performance evaluation of the target speaker extraction device configured using the conventional method. [Figure 13] FIG. 2 illustrates an example of a functional configuration of a computer. DETAILED DESCRIPTION OF THE INVENTION

[0011] As a result of intensive research, the inventors experimentally demonstrated that the extraction failure rate varies greatly depending on the selection of pre-registered utterances, and made it clear for the first time that robustness to the selection of pre-registered utterances is an important issue. The present invention aims to improve the extraction performance of a target speaker extraction device. Specifically, the present invention aims to improve the worst-case target speaker extraction performance for multiple pre-registered utterance candidates, and proposes a new training system and training method for target speaker extraction that focuses on utterances with particularly poor extraction performance from multiple pre-registered utterances used during training. Hereinafter, embodiments of the present invention will be described in detail. Components having the same functions are given the same numbers, and duplicated explanations will be omitted. [Example]

[0012] The first embodiment of the present invention will be described in order of construction of a learning dataset and execution of learning, using FIGS.

[0013] In the problem formulation of target speaker extraction, the input observed signal Y is expressed as follows using the target speaker's voice S, the voices of interfering speakers other than the target speaker I, and other noise R. Y=S+I+R The target speaker extraction function TSE(·) extracts auxiliary information about the target speaker, typically a pre-registered utterance C S The target speaker's signal is estimated using and output as ^S. ^S=TSE(Y, C s ) The training of the model for target speaker extraction is performed by minimizing the distance between the target speaker voice S to be extracted and the estimated target speaker voice ^S. Therefore, pair data (S, Y, C) are used for the training. S ) is required.

[0014] [Building a learning dataset] FIG. 3 shows the learning dataset construction unit and its input / output data in the first embodiment of the present invention. The training dataset construction unit 301 is a part that creates a dataset consisting of paired data for training a neural network for target speaker extraction. Figure 3 shows a configuration for constructing a training dataset by creating simulated mixed sounds based on a speech database 302 and a noise database 303. The training data in this invention is a set of N pre-registered utterances C for one target speaker voice S and one mixed sound Y. S ,...,C S N In other words, the training data is created by preparing (S, Y, C S 1,...,C S N )

[0015] Figure 4 shows the training data (S, Y, C S 1,...,C S N First, we select an utterance i from a target speaker P and N utterances j1,...,j from the speech database that are different from i. NThen, an utterance k of another speaker Q is obtained as noise (step S41). Also, noise R is obtained from the noise database (step S42). i, k, and R are mixed with an arbitrary gain (step S43). Next, utterance i of speaker P is output as utterance S (teacher speech) to be extracted (step S44), and j1,...,j N Pre-registered utterance C S 1,...,C S N (step S45), and the mixture of i, k, and R is output as a mixed sound Y (step S46), and these are stored in the learning data set storage unit 202.

[0016] In the above method, the pre-registered utterances were simulated by extracting N different utterances from a speech database, each uttered by the speaker of the speech S to be extracted. However, they may also be obtained by acquiring N utterances of speaker S from a database for pre-registered utterances. Furthermore, the dataset may be constructed on the fly during training, rather than in advance of training; that is, the training data may be different at each time training is performed.

[0017] [Learning implementation] FIG. 5 shows a target speaker extraction training system 50 according to the first embodiment of the present invention. The speaker expression extraction unit 101 converts the speech signal of one pre-registered utterance into a fixed-length vector expression representing speaker-related features. The target speaker extraction unit 102 extracts the target speaker's utterance from the mixed sound Y included in the training data and the fixed-length speaker expressions output from the speaker expression extraction unit 101 . The data selection unit 501 selects data from the extracted speech that is considered to have a large contribution to learning for target speaker extraction. The optimization unit 201 calculates a loss function from both the utterance obtained from the data selection unit 501 and the target speaker speech S to be extracted, and optimizes the parameters included in the model related to target speaker extraction so that the loss function is minimized.

[0018] Figure 6 shows the procedure for optimizing the parameters included in the model related to target speaker extraction. First, a learning model before parameter optimization is set in the speaker expression extraction unit and the target speaker extraction unit. The speaker expression extraction unit 101 can be configured using a deep neural network or a publicly known speaker expression extraction technique such as d-vector or i-vector. The target speaker extraction unit 102 can be configured using a publicly known neural network, and specifically, a publicly known network structure such as ConvTasNet can be adopted. As the learning model before parameter optimization, a network trained in advance using another dataset can be used. The speaker expression extraction unit 101 extracts pre-registered utterance C S 1,...,C S N Each of these is expressed as a fixed-length speaker-related feature vector λ S 1,...,λ S N (step S61). The target speaker extraction unit 102 extracts the mixed sounds Y included in the training data and the fixed-length speaker expressions λ output from the speaker expression extraction unit 101. S 1, ··· ,λ S N That is, the target speaker extraction unit 102 extracts the utterance of the target speaker from each pair (Y, C S 1),...,(Y, C S N ) and extract the target speaker from the extracted speech (estimated teacher speech) ^S1,...,^S N is generated (step S62).

[0019] The data selection unit 501 selects the extracted voices ^S1,...,^S N From these, data that is thought to have a large contribution to learning for target speaker extraction is selected. In conventional learning methods, there is a problem that the performance of target speaker extraction varies depending on the pre-registered utterance selected. S 1,...,C S NFrom the results of target speaker extraction using each of the above, learning is performed on those that are not well extracted, thereby improving the robustness of the selection of pre-registered utterances. For this purpose, the data selection unit 501 selects the speech to be extracted (teacher speech) S and each of ^S1,...,^S N The difficulty index of extraction is calculated using the above (step S63), and the most difficult to extract (the worst estimated teacher speech, ^S m ) and outputs it to the optimization unit 201 (step S64). As an index showing the difficulty of extraction, for example, the extracted speech ^S1,...,^S N The distance between the target speech S and the target speech S can be used. The distance measure can be a known source-to-distortion ratio (SDR), etc.

[0020] The optimization unit 201 uses the ^S obtained from the data selection unit 501 m The loss function is calculated from both the target speech S and the correct speech S to be extracted, and the parameters included in the model for target speaker extraction are optimized by minimizing this loss function. As a loss function, for example, a well-known scale dependent signal-to-noise ratio (sd-SNR) can be used, and as an optimization method, a well-known method such as Adam can be used. When optimizing the parameters, the parameters of both the speaker expression extraction unit 101 and the target speaker extraction unit 102 may be updated simultaneously, or the parameters of the speaker expression extraction unit 101 may be fixed and only the parameters of the target speaker extraction unit 102 may be updated. Once the parameter optimization is complete, the training is complete. [Example]

[0021] In the first embodiment, a configuration is shown for selecting one utterance from N utterances for which target speaker extraction is most difficult. This method is called hard selection. Incidentally, if we utilize not only the most difficult utterance but also the second most difficult utterance, we can make full use of data that is likely to be useful. Also, by incorporating some of the easy utterances into the learning process rather than ignoring them altogether, we can mitigate the phenomenon (which tends to occur in hardware) where the system adapts only to difficult utterances and performance for easy utterances declines. Therefore, we can consider a method of assigning weights to the N utterances according to the difficulty of extraction. This method is called soft selection. In Example 1, extraction was the most difficult. m In Example 2, the parameters of the learning model were optimized using only ^S1,...,^S N All of these are used to optimize the parameters. Second Embodiment FIG. 7 shows a target speaker extraction training system 70 according to a second embodiment of the present invention, and FIG. 8 shows a procedure for optimizing parameters included in a model related to target speaker extraction. 5 and 7, the data selection unit 501 is replaced with a weight calculation unit 701, and in S64 and S65, ^S m In S84 and S85, one ^S1,...,^S N The difference is that a weight is calculated and used for each of the

[0022] [Building a learning dataset] The construction of the learning dataset is the same as in the first embodiment.

[0023] [Learning implementation] Steps S61 to S63 are the same as in the first embodiment, so the explanation will start from step S84. The weight calculation unit 701 calculates the weights ^S1,...,^S according to the difficulty of extraction. N (step S84), and N The weights are calculated by, for example, calculating the extracted speech ^S1,...,^S NThe distance between the target speech S and the target speech S can be calculated by applying a temperature-added softmax function. The distance measure can be a known source-to-distortion ratio (SDR), etc.

[0024] The optimization unit 201 calculates the weights ^S1,...,^S obtained from the weight calculation unit 703. N For each of the above, a loss function for the speech to be extracted (teacher speech) S is calculated, and the loss sum is calculated by weighting and adding the loss function with the weight information output from the weight calculation unit 703. By minimizing this loss sum, the parameters included in the model for target speaker extraction are optimized. As in Example 1, the loss function can be, for example, the well-known scale dependent signal-to-noise ratio (sd-SNR), the optimization method can be the well-known Adam method, and when optimizing the parameters, the parameters of both the speaker expression extraction unit and the target speaker extraction unit can be updated simultaneously, or the parameters of the speaker expression extraction unit can be fixed and only the parameters of the target speaker extraction unit can be updated. Once the parameter optimization is complete, the training is complete. [Example]

[0025] As an indicator of the difficulty of target speaker extraction, we used pre-registered utterance C. S 1,...,C S N It is also possible to use the accuracy of speaker identification performed by a computer. A pre-registered utterance that can be correctly identified using a speaker identification model is considered to be typical of that speaker's speech and to be relatively easy to extract the target speaker. In other words, since pre-registered utterances that are difficult to identify also make target speaker extraction difficult, it is considered effective to train with an emphasis on such pre-registered utterances in order to obtain robustness against pre-registered utterances. 9 shows a target speaker extraction training system 90 according to a third embodiment of the present invention, and FIG. 10 shows a procedure for optimizing parameters included in a model related to target speaker extraction. The difference from the first embodiment is that the utterances to be trained, extracted utterances ^S1,...,^S N Instead of selecting from pre-registered utterance C S 1,...,C S N 9, in the third embodiment, a pre-registered utterance selection unit 901 is arranged before the speaker expression extraction unit 101.

[0026] [Building a learning dataset] The construction of the learning dataset is the same as in the first embodiment.

[0027] [Learning implementation] The pre-registered utterance selection unit 901 selects C S 1,...,C S N Among these, utterance C is the most difficult to identify. S n (Step S101). In this case, for example, a method can be used in which the utterance with the lowest posterior probability of the correct speaker class is selected from among the N pre-registered utterances. The speaker expression extraction unit 101 is S n From speaker feature information λ S n is generated (step S102). The target speaker extraction unit 102 extracts Y and λ S n Extracted speech (estimated teacher speech)^S n is generated (step S94). The optimization unit 201 n The loss function is calculated from the speech to be extracted (teacher speech) S, and the parameters included in the model for target speaker extraction are optimized by minimizing it. As in Example 1, the loss function can be, for example, the well-known scale dependent signal-to-noise ratio (sd-SNR), the optimization method can be the well-known Adam method, and when optimizing the parameters, the parameters of both the speaker expression extraction unit and the target speaker extraction unit can be updated simultaneously, or the parameters of the speaker expression extraction unit can be fixed and only the parameters of the target speaker extraction unit can be updated. Once the parameter optimization is complete, the training is complete.

[0028] [Speaker extraction using the target speaker extraction device] The parameters learned as described above are set in the speaker expression extraction unit and target speaker extraction unit of the target speaker extraction device shown in FIG. Auxiliary data consisting of the target speaker's voice alone (approximately 10 seconds long, does not need to match the speech in the mixed sound) is input to the target speaker extraction device 100, along with the mixed sound of multiple speakers including the target speaker's voice, which is the subject of analysis. The speaker expression extraction unit extracts the target speaker's features from the auxiliary data, and the target speaker extraction unit extracts and outputs the target speaker's speech from the target speaker's feature data and the mixed sound.

[0029] [Performance evaluation] FIG. 11 shows the performance evaluation results of a target speaker extraction device trained according to the present invention. The results were compared between a conventional training method, the hard selection method of the present invention (Example 1), and the soft selection method of the present invention (Example 2). The evaluation measures used were SDR improvement (SDRi), character error rate (CER), and extraction failure rate. SDRi is the difference between the SDR of the mixture Y and the SDR of the extracted speech ^S, and the extraction failure rate was the proportion of data with an SDRi of less than 5 dB. To evaluate the robustness of the selection of pre-registered utterances, 10 pre-registered utterances (auxiliary data) were prepared for each mixture Y in the evaluation dataset, and the average, worst, and second-worst performance values ​​for the 10 utterances were averaged across the entire evaluation dataset. Our method improved the minimum performance for selecting pre-registered utterances by up to 0.8 dB in SDRi and up to 1.5% in CER. We also significantly improved the target speaker extraction failure rate from an average of 2.0% to 1.3%.

[0030] [Supplementary information on performance evaluation] Figure 12 shows the evaluation of the performance of the target speaker extraction device in Non-Patent Document 1. For each observed signal Y in the evaluation set, N=10 pre-registered utterances C S 1,...,C S N Prepare ^S1,...,^S N and obtain ^S1,...,^S N The data are sorted in descending order of extraction performance, and the distribution of SDRi is shown for the worst, second-worst, ..., best extraction performance. Each box and whisker represents the 5th, 25th, 50th, 75th, and 95th percentiles. For example, the 50th percentile line for "Worst" indicates the position where the worst extraction performance in each evaluation set is ranked at 50% when sorted by SDRi value. In relation to the table in Figure 11, the SDRi-worst result of the conventional method (13.3 dB) corresponds to the average value of the leftmost box and whiskers, and the 2nd-worst result (14.3 dB) corresponds to the average value of the second box and whiskers. Note that the 50th percentile line is the median, so it does not coincide with the average value. The bottom part of the figure shows the failure rate for each extraction performance, i.e., the percentage of pre-registered utterances for which the SDRi was less than 5 dB. When the most difficult pre-registered utterance was selected, the failure rate was 7.2%, while when the easiest pre-registered utterance was selected, the failure rate was 0.2%, showing a large difference. Because extraction failures impair usability, it is more important in practice to avoid failures than to improve average performance. The graph shows that extraction failures occur intensively when pre-registered utterances that were difficult to extract are selected, demonstrating the effectiveness of the present invention's training, which emphasizes pre-registered utterances from which target speakers are difficult to extract.

[0031] [Programs, recording media] The various processes described above can be implemented by loading a program that executes each step of the above method into the recording unit 2020 of the computer 2000 shown in Figure 13 and operating the control unit 2010, input unit 2030, output unit 2040, etc.

[0032] The program describing the processing contents can be recorded on a computer-readable recording medium, which may be, for example, a magnetic recording device, an optical disk, a magneto-optical recording medium, a semiconductor memory, or any other suitable recording medium.

[0033] The program may be distributed, for example, by selling, transferring, lending, etc. a portable recording medium such as a DVD or CD-ROM on which the program is recorded. Furthermore, the program may be stored in a storage device of a server computer, and then transferred from the server computer to another computer via a network, thereby distributing the program.

[0034] A computer that executes such a program may first temporarily store the program recorded on a portable recording medium or transferred from a server computer in its own storage device. Then, when executing a process, the computer reads the program stored on its own recording medium and executes the process in accordance with the read program. Alternatively, the computer may read the program directly from a portable recording medium and execute the process in accordance with the program. Furthermore, the computer may execute the process in accordance with the received program each time a program is transferred from a server computer to the computer. Alternatively, the server computer may not transfer the program to the computer, but may execute the process through a so-called ASP (Application Service Provider) service, which realizes the processing function by issuing an execution instruction and obtaining the results. In this embodiment, the program includes information used for processing by a computer that is equivalent to a program (such as data that is not a direct instruction to the computer but has properties that define computer processing).

[0035] Furthermore, in this embodiment, the device is configured by executing a predetermined program on a computer, but at least a part of the processing contents may be realized by hardware.

Claims

1. 1. A system for training a neural network for target speaker extraction, comprising: a learning dataset storage unit that stores a teacher voice, a mixed sound including the teacher voice, and a plurality of pre-registered utterances obtained from the same speaker as the teacher voice; a speaker expression extraction unit including a neural network 1 that generates one piece of speaker feature information for each pre-registered utterance from the pre-registered utterance; a target speaker extraction unit including a neural network 2 that generates one estimated teacher speech for each of the speaker feature information from the speaker feature information and the mixed sound; a data selection unit that calculates a difference 1 between the estimated teacher speech and the teacher speech for each of the estimated teacher speeches and identifies a worst-case estimated teacher speech that gives the largest difference 1; an optimization unit that calculates a difference 2 between the teacher speech and the best estimated teacher speech, and adjusts parameters of both the neural network 1 and the neural network 2, or only the neural network 2, to minimize the difference 2; A target speaker extraction learning system equipped with:

2. 1. A method of training a neural network for target speaker extraction, comprising: storing a teacher voice, a mixed sound including the teacher voice, and a plurality of pre-registered utterances obtained from the same speaker as the teacher voice in a training dataset storage unit; generating one piece of speaker feature information for each of the pre-registered utterances from the pre-registered utterances using a neural network; generating an estimated training speech for each of the speaker feature information from the speaker feature information and the mixed sound by a neural network (2); calculating a difference 1 between the estimated teacher speech and the teacher speech for each of the estimated teacher speeches, and identifying a worst-case estimated teacher speech that gives the largest difference 1; Calculating a difference 2 between the teacher speech and the best estimated teacher speech, and adjusting parameters of both the neural network 1 and the neural network 2, or only the neural network 2, to minimize the difference 2; Target speaker extraction learning method.

3. A program for causing a computer to function as the target speaker extraction and training system according to claim 1.

Images