Title of Invention: Signal Extraction System, Signal Extraction Learning Method and Signal Extraction Learning Program
Technical field
[0001]
　The present invention relates to a signal extraction system, a signal extraction learning method, and a signal extraction learning program for extracting signals belonging to each class.
Background technology
[0002]
　Various techniques are known for extracting signals belonging to each class from observed signals. For example, speaker diarization is a technology that analyzes speech signals with unknown speaker information (such as the number of speakers) and estimates which speaker spoke when. This is a technique for identifying a set from an audio signal.
[0003]
　A common method of speaker diarization is to segment the speech signal and cluster the segmented speech signal to identify a set of segments for each speaker.
[0004]
　FIG. 10 is an illustration showing a method of identifying a speaker's speech signal from a segmented speech signal. In the method shown in FIG. 10, a monaural audio stream 201 is segmented to extract audio portions. FIG. 10 shows an example in which four segments 202-205 are extracted. Here, the segments 202, 203 and 205 and the segment 204 are each judged to be similar signals and clustered. As a result, the former segment is identified as speaker A and the latter segment is identified as speaker B. .
[0005]
　On the other hand, clustering of segments as illustrated in FIG. 10 has the problem that it is susceptible to noise and cannot separate speech signals of temporarily overlapping speakers. In order to deal with such problems, a method of specifying a speech signal by using the speech signal of a target speaker (sometimes referred to as an anchor) has also been proposed.
[0006]
　FIG. 11 is an explanatory diagram showing a method of identifying a speaker's speech signal using anchors. In the method shown in FIG. 11, for a monaural audio stream 201, an anchor 206, which is the audio signal of speaker A to be identified, is prepared. Audio stream 201 and anchor 206 are then compared to identify a set of speaker A segments (segments 207, 208, and 209).
[0007]
　Non-Patent Document 1 describes a method of extracting a speaker's voice signal using deep learning technology. In the method described in Non-Patent Document 1, a mask (segment) for extracting the speech signal (segment) of the target speaker based on the anchor, which is the speech signal of the target speaker, and the mixed speech signal ( reconstruction mask). By applying the learned reconstruction mask to the mixed speech signal, a set of segments for the speaker of interest is identified.
[0008]
　Note that Non-Patent Document 2 describes a method for extracting feature amounts from input speech.
prior art documents
Non-patent literature
[0009]
Non-Patent Document 1: Jun Wang, et al., "Deep Extractor Network for Target Speaker Recovery From Single Channel Speech Mixtures", Interspeech 2018, 2 - 6 September 2018.
Non-Patent Document 2: David Snyder et al., "X- Vectors: Robust DNN Embeddings for Speaker Recognition", ICASSP 2018 - 2018 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), April 2018.
SUMMARY OF THE INVENTION
Problems to be Solved by the Invention
[0010]
　In the method described in Non-Patent Document 1, the reconstruction is based on the speaker's anchor X f,t as expressed in two time-frequency dimensions and the mixed speech signal X f,t ms . A mask M f,t is learned. Then, by applying the learned reconstruction mask M f,t to the mixed speech signal X f,t ms , the spectrogram S^ f,t ms of the speaker is estimated (where S is the showing a superscript hat). Specifically, the speaker's spectrogram S^ f,t ms is calculated based on Equation 1 shown below.
[0011]
[Number 1]

　
[0012]
　During training, the reconstruction mask is learned by optimizing to minimize the loss function of Equation 2 below. S f,t ms in Equation 2 is the speaker's spectrogram. Specifically, the neural network described in Non-Patent Document 1 learns a reconstruction mask that can handle overlapping speech out of irrelevant noise.
[0013]
[Number 2]

　
[0014]
However, the ground truth of the reconstruction mask M f,t and the reconstructed speaker's spectrogram S f,t ms 　contained in Equation 2 above are generally unknown. Therefore, optimization using Equation 2 above has a problem that there is a limit to improving the accuracy of the reconstructed mask.
[0015]
　It is also conceivable to improve the accuracy of the reconstructed mask by artificially generating learning data in which a plurality of (for example, two) audio signals are superimposed. However, it is difficult for artificial data to sufficiently reflect the factors that exist in real data (for example, exchanges of conversations, reverberation, etc.). It is difficult to generate a reconstruction mask that can extract the target speaker's speech signal from the ambient sound.
[0016]
　SUMMARY OF THE INVENTION Accordingly, it is an object of the present invention to provide a signal extraction system, a signal extraction learning method, and a signal extraction learning program capable of accurately extracting signals belonging to each class from observed signals.
Means to solve problems
[0017]
　A signal extraction system according to the present invention estimates a layer for inputting a mixed signal containing an anchor signal belonging to a predetermined class and a target signal belonging to the class, and a reconstruction mask indicating a time-frequency domain in which the target signal exists in the mixed signal. a first network having an output layer as a result, a layer receiving a signal of interest extracted by applying the mixed signal to the reconstruction mask, and a result of classifying the input signal of interest into predetermined classes a neural network input that inputs a neural network coupled with a second network having a layer to output; and applying the anchor signal and the mixed signal to the first network to reconstruct the class to which the anchor signal belongs. a reconstructed mask estimator for estimating a mask; and a signal classifier for applying a mixed signal to the estimated reconstructed mask to extract a target signal and applying the extracted target signal to a second network to classify it into classes. a loss calculation unit that calculates a loss function between a class into which the extracted target signal is classified and a true class; It is characterized by comprising a parameter updating section for updating parameters of the second network and an output section for outputting the updated first network.
[0018]
　The signal extraction learning method according to the present invention includes a layer for inputting a mixed signal containing an anchor signal belonging to a predetermined class and a target signal belonging to the class, and a reconstruction mask indicating a time-frequency domain in which the target signal exists in the mixed signal. a layer for outputting an estimation result; a layer for inputting a target signal extracted by applying a mixed signal to a reconstruction mask; and a result of classifying the input target signal into predetermined classes. and apply the anchor signal and the mixed signal to the first network to estimate the reconstruction mask for the class to which the anchor signal belongs. , applying the mixture signal to the estimated reconstruction mask to extract the signal of interest, applying the extracted signal of interest to a second network to classify the extracted signal of interest into classes, the extracted signal of interest being the classified classes and , calculating the loss function with the true class, updating the parameters of the first network and the parameters of the second network in the neural network based on the loss function calculation result, and the updated first network It is characterized by outputting.
[0019]
　A signal extraction learning program according to the present invention provides a computer with a layer for inputting a mixed signal including an anchor signal belonging to a predetermined class and a target signal belonging to the class, and a layer representing a time-frequency domain in which the target signal exists in the mixed signal. A first network having a layer for outputting a construction mask as an estimation result, a layer for inputting a target signal extracted by applying a mixed signal to the reconstruction mask, and inputting the input target signal into a predetermined class. A neural network input processing that inputs a neural network coupled with a second network having a layer that outputs a classified result, applying the anchor signal and the mixed signal to the first network to determine the class to which the anchor signal belongs a reconstruction mask estimation process for estimating a reconstruction mask of , applying a mixture signal to the estimated reconstruction mask to extract a signal of interest, and applying the extracted signal of interest to a second network to classify it into classes Signal classification processing, loss calculation processing for calculating a loss function between the class into which the extracted target signal is classified and the true class, parameters of the first network in the neural network and the first network based on the calculation result of the loss function A parameter update process for updating the parameters of the second network and an output process for outputting the updated first network are executed.
Effect of the invention
[0020]
　According to the present invention, signals belonging to each class can be accurately extracted from observed signals.
Brief description of the drawing
[0021]
1 is a block diagram showing a configuration example of a first embodiment of a signal extraction system according to the present invention; FIG.
2 is an explanatory diagram showing an example of an input neural network; FIG.
3 is a flowchart showing an operation example of the signal extraction system of the first embodiment; FIG.
4 is a flowchart showing another operation example of the signal extraction system of the first embodiment; FIG.
5 is an explanatory diagram showing an example of signals to be extracted in conversation; FIG.
6 is a block diagram showing a configuration example of a second embodiment of a signal extraction system according to the present invention; FIG.
7 is a flowchart showing an operation example of the signal extraction system of the second embodiment; FIG.
8 is a block diagram showing an overview of a signal extraction system according to the present invention; FIG.
9 is a schematic block diagram showing the configuration of a computer according to at least one embodiment; FIG.
10 is an explanatory diagram showing a method of identifying a speaker's speech signal from a segmented speech signal; FIG.
11 is an explanatory diagram showing a method of identifying a speaker's voice signal using anchors; FIG.
MODE FOR CARRYING OUT THE INVENTION
[0022]
　BEST MODE FOR CARRYING OUT THE INVENTION Hereinafter, embodiments of the present invention will be described with reference to the drawings. In the following description, as a specific example of extracting signals belonging to each class from observed signals, a method of extracting speech signals (segments) of each speaker from a speech stream will be described. However, signals to be extracted by the present invention are not limited to audio signals.
[0023]
Embodiment 1.
　FIG. 1 is a block diagram showing a configuration example of a first embodiment of a signal extraction system according to the present invention. A signal extraction system 100 of this embodiment includes a neural network input unit 10, an anchor signal input unit 20, a mixed signal input unit 30, a learning unit 40, an output unit 50, and an extraction unit 60.
[0024]
　A neural network input unit 10 inputs a neural network for extracting signals belonging to a certain class. A class in this embodiment means a set of signals having a specified property. In the case of audio signals, the classes are specifically individual speakers, gender, age, language, emotion, and the like. For example, when speaker A is defined as a class, a signal indicating an utterance of speaker A is a signal belonging to speaker A's class.
[0025]
　Also, the neural network input in this embodiment is configured as a neural network in which two types of networks are combined. The first network includes a layer for inputting a mixed signal including an anchor signal belonging to a predetermined class and a signal belonging to that class (hereinafter referred to as a target signal), and a layer for inputting a target signal in the input mixed signal. and a layer that outputs a mask indicating the time-frequency domain (hereinafter referred to as a reconstruction mask) as an estimation result. The time-frequency domain refers to the region that can be identified from a signal based on time and frequency. For example, if speaker A is defined as a class, the target signal is a signal indicating speaker A's utterance.
[0026]
　A specific example of the first network is a CNN (Convolutional Neural Network). In particular, when assuming an audio stream as a mixed signal, it is conceivable that the length of the signal will be variable. Therefore, it is preferable to use TDNN (Time Delay Neural Network), which is a one-dimensional convolutional neural network (1D CNN), for the first network. It should be noted that the first network may be a network that inputs a mixed signal obtained by dividing the mixed signal into predetermined lengths (for example, 4 seconds).
[0027]
　The second network has a layer for inputting a target signal extracted by applying the mixed signal to the reconstruction mask, and a layer for outputting the result of classifying the input target signal into predetermined classes. Therefore, it can be said that the neural network that is input to this embodiment is a neural network that receives the anchor signal and the mixed signal as inputs and outputs classes into which the target signal to be extracted is classified.
[0028]
　Specifically, the second network corresponds to outputs corresponding to the number of classes assumed to be extracted, that is, all or part of the classes included in the learning data referred to by the learning unit 40 described later. Has a layer with output set. As an exception handling, it is assumed that the mixed signal does not contain signals of any of the expected classes. may have This added output is the output for detecting exception handling.
[0029]
　FIG. 2 is an explanatory diagram showing an example of an input neural network. The first network N1 exemplified in FIG. 2 receives the speaker's utterance X f,tas as an anchor signal and the mixed speech X f,t ms including the speaker's utterance, and uses the reconstruction mask M f , t as outputs. In addition, the second network N2 receives as input the signal S^ f,t ms representing the speaker's utterance extracted by applying the mixed speech X f,t m to the reconstruction mask M f,t , It is a network that outputs the results of classifying signals into predetermined classes. These two types of networks are combined to form one neural network N3.
[0030]
　An anchor signal input unit 20 inputs an anchor signal to be input to the neural network. Specifically, the anchor signal input unit 20 inputs an anchor signal belonging to a class to be extracted using a reconstruction mask. In other words, the learning unit 40, which will be described later, learns a reconstruction mask for extracting the class to which the input anchor signal belongs. In the example shown in FIG. 2, the utterance X f, tas corresponds to the anchor signal.
[0031]
　The mixed signal input unit 30 inputs a signal (that is, a mixed signal) containing a target signal to be extracted. In the example shown in FIG. 2, the mixed speech X f,t ms corresponds to the mixed signal.
[0032]
　The learning unit 40 learns the entire neural network including two types of networks. Since the target reconstruction mask is unknown, the learning unit 40 of the present embodiment performs learning using weak labels with the label of the class to be classified as the target. The learning unit 40 includes a reconstruction mask estimator 42 , a signal classifier 44 , a loss calculator 46 and a parameter updater 48 .
[0033]
　A reconstructed mask estimator 42 applies the input anchor signal and mixed signal to the first network to estimate a reconstructed mask for the class to which the anchor signal belongs. Specifically, the reconstruction mask estimation unit 42 estimates the output of the first network in the neural network as the reconstruction mask.
[0034]
　The signal classification unit 44 applies the mixed signal to the estimated reconstruction mask to extract the target signal, and applies the extracted target signal to the second network to classify it into classes. Specifically, the signal classification unit 44 acquires the output of the second network in the neural network as a class into which the target signal is classified. For example, when the mixed signal is an audio stream representing the utterance of a speaker, the signal classification unit 44 extracts the spectrogram of the speaker as the target signal, applies the extracted spectrogram to the second network, and identifies the speaker. Classify.
[0035]
　A loss calculator 46 calculates a loss function between the class into which the extracted target signal is classified and the true class. A true class is a class to which the input anchor signal belongs. The loss calculator 46 may, for example, calculate the loss function using the cross entropy exemplified in Equation 3 below.
[0036]
[Number 3]

　
[0037]
　In Equation 3, c i is the true label information of the anchor signal and takes a value of 1 when the anchor signal belongs to the i-th class and 0 otherwise. Also, c^ i is label information of the classified class, and is the output value of each element in the output layer of the second network. This output value is preferably normalized in the second network, such as with a softmax activation function. The label information is added by the signal classification unit 44 and preset to the anchor signal.
[0038]
　The parameter updating unit 48 updates the parameters of the first network and the parameters of the second network in the neural network based on the calculation result of the loss function. Specifically, the parameter updating unit 48 updates each parameter in the neural network so as to minimize the loss function. The parameter updating unit 48 may update the parameters by, for example, the error backpropagation method. However, the parameter update method is not limited to the error backpropagation method, and the parameter update unit 48 may update the parameters using a generally known method.
[0039]
　The output unit 50 outputs the updated first network. That is, the output unit 50 outputs a neural network obtained by removing the network for classifying the target signal into classes (that is, the second network) from the input neural network.
[0040]
　The extraction unit 60 applies the anchor signal and the mixed signal to the output first network and extracts the signal (target signal) of the class to which the anchor signal belongs. The extracted signal can be used, for example, for speaker identification.
[0041]
　For example, in the method described in Non-Patent Document 1, a process of optimizing the loss function exemplified in Equation 2 above is performed. However, as mentioned above, the true values ​​of the reconstruction mask M f,t and the reconstructed speaker's spectrogram S f,t ms are generally unknown, so to improve the accuracy of the reconstruction mask is limited. On the other hand, in this embodiment, the learning unit 40 learns the neural network so as to optimize the loss function of Equation 3 (that is, the loss function between classes) shown above. Therefore, it becomes possible to learn a reconstruction mask that can accurately extract signals belonging to each class from observed signals.
[0042]
　Neural network input unit 10, anchor signal input unit 20, mixed signal input unit 30, learning unit 40 (more specifically, reconstruction mask estimation unit 42, signal classification unit 44, loss calculation unit 46, parameter The update unit 48), the output unit 50, and the extraction unit 60 are realized by a computer processor (e.g., CPU (Central Processing Unit), GPU (Graphics Processing Unit)) that operates according to a program (signal extraction learning program). be.
[0043]
　For example, the program is stored in a storage unit (not shown) included in the signal extraction system 100, the processor reads the program, and according to the program, neural network input unit 10, anchor signal input unit 20, mixed signal input unit 30 , the learning unit 40 (more specifically, the reconstructed mask estimator 42, the signal classifier 44, the loss calculator 46, and the parameter updater 48), the output unit 50, and the extractor 60. . Also, the functions of the signal extraction system 100 may be provided in a SaaS (Software as a Service) format.
[0044]
　Neural network input unit 10, anchor signal input unit 20, mixed signal input unit 30, learning unit 40 (more specifically, reconstruction mask estimation unit 42, signal classification unit 44, loss calculation unit 46, parameter The update unit 48), the output unit 50, and the extraction unit 60 may each be realized by dedicated hardware. Also, part or all of each component of each device may be implemented by general-purpose or dedicated circuitry, processors, etc., or combinations thereof. These may be composed of a single chip, or may be composed of multiple chips connected via a bus. A part or all of each component of each device may be implemented by a combination of the above-described circuits and the like and programs.
[0045]
　Further, when part or all of each component of the signal extraction system 100 is realized by a plurality of information processing devices, circuits, etc., the plurality of information processing devices, circuits, etc. may be centrally arranged, They may be distributed. For example, the information processing device, circuits, and the like may be realized as a form in which each is connected via a communication network, such as a client-server system, a cloud computing system, or the like.
[0046]
　Next, the operation of the signal extraction system 100 of this embodiment will be described. FIG. 3 is a flow chart showing an operation example of the signal extraction system 100 of this embodiment. The neural network input unit 10 inputs a neural network in which the first network and the second network are combined (step S11).
[0047]
　The anchor signal input unit 20 inputs the anchor signal (step S12), and the mixed signal input unit 30 inputs the mixed signal (step S13). The learning unit 40 (more specifically, the reconstructed mask estimating unit 42) applies the input anchor signal and mixed signal to the first network to estimate the reconstructed mask of the class to which the anchor signal belongs ( step S14).
[0048]
　The learning unit 40 (more specifically, the signal classification unit 44) applies the mixed signal to the estimated reconstruction mask to extract the target signal, and applies the extracted target signal to the second network. Classify into classes (step S15). The learning unit 40 (more specifically, the loss calculation unit 46) calculates a loss function between the class into which the extracted target signal is classified and the true class (step S16).
[0049]
　The learning unit 40 (more specifically, the signal classification unit 44) updates the parameters of the first network and the parameters of the second network in the neural network based on the calculation result of the loss function (step S17). The output unit 50 then outputs the updated first network (step S18).
[0050]
　FIG. 4 is a flowchart showing another operation example of the signal extraction system 100 of this embodiment. The extraction unit 60 receives the output first network (step S21). Also, the anchor signal input unit 20 inputs the anchor signal (step S22), and the mixed signal input unit 30 inputs the mixed signal (step S23). Then, the extraction unit 60 applies the input anchor signal and mixed signal to the first network to estimate a reconstruction mask (step S24), and applies the estimated reconstruction mask to the mixed signal to obtain the target signal is extracted (step S25).
[0051]
　As described above, in this embodiment, the neural network input unit 10 inputs the neural network in which the first network and the second network are combined, and the reconstruction mask estimation unit 42 inputs the anchor signal and the mixed signal to the first network to estimate the reconstruction mask for the class to which the anchor signal belongs. Further, the signal classification unit 44 applies the mixed signal to the estimated reconstruction mask to extract the target signal, and applies the extracted target signal to the second network to classify it into classes. Then, the loss calculation unit 46 calculates a loss function between the class into which the extracted target signal is classified and the true class, and the parameter updating unit 48 calculates the first class in the neural network based on the calculation result of the loss function. Update the parameters of the first network and the parameters of the second network. The output unit 50 then outputs the updated first network.
[0052]
　Such a configuration can improve the accuracy of the reconstructed mask estimated by the first network. As a result, signals belonging to each class can be accurately extracted from the observed mixed signals.
[0053]
　Also, the signal extraction system of the present embodiment can be implemented as a system for extracting signals of any class as exemplified below, for example.
　・A system that extracts signals of specific speaker, gender, age, language, and emotion classes from mixed utterances
　・A system that extracts signals of specific instrument classes from mixed music
　・Specific sounds from mixed speech Systems for extracting signals of the class of events, e.g. explosions and gunshots Systems for extracting signals of
　specific electrical equipment classes from mixed electrical flows Systems for extracting signals of
　specific communication equipment classes from mixed radio waves system to
[0054]
Embodiment 2.
　Next, a second embodiment of the signal extraction system according to the invention will be described. By using the reconstruction mask estimated according to the first embodiment, signals belonging to each class can be extracted with high accuracy from the mixed signal. Furthermore, in this embodiment, a method for extracting the target signal of each speaker from the speech signal with higher accuracy will be described.
[0055]
　In the process of extracting a target signal from a speech signal, utterances (segments) of individual speakers are generally estimated independently. In normal conversation, each speaker generally speaks alternately and exclusively.
[0056]
　FIG. 5 is an explanatory diagram showing an example of signals to be extracted in conversation. Signals of a plurality of speakers (speaker A and speaker B) are generally alternately and exclusively observed as illustrated in FIG. 5(a), and simultaneously observed as illustrated in FIG. 5(b). is rarely done. In this embodiment, we will focus on the characteristics of this conversation and describe a method of correcting the reconstruction mask in relation to other signals.
[0057]
　FIG. 6 is a block diagram showing a configuration example of a second embodiment of the signal extraction system according to the present invention. The signal extraction system 200 of this embodiment includes a neural network input unit 10, an anchor signal input unit 20, a mixed signal input unit 30, a learning unit 40, an output unit 50, a reconstruction mask conversion unit 52, an extraction a portion 60;
[0058]
　That is, the signal extraction system 200 of this embodiment differs from the signal extraction system 100 of the first embodiment in that it further includes a reconstruction mask conversion section 52 . Other configurations are the same as those of the first embodiment.
[0059]
　In this embodiment, the signal extraction system 200 uses the reconstruction masks of multiple speakers and modifies at least one of the reconstruction masks. Therefore, the anchor signal input unit 20 inputs anchor signals of a plurality of speakers. In the following description, the case of using the reconstructed masks of two speakers will be described, but the same applies even if there are three or more speakers. That is, the anchor signal input unit 20 inputs the anchor signals of the two speakers respectively.
[0060]
　The mixed signal input section 30 inputs a mixed signal.
[0061]
　The learning unit 40 estimates the first network for each speaker based on the input anchor signals and mixed signals, and the output unit 50 outputs the generated first networks.
[0062]
　The reconstructed mask transforming unit 52 inputs the generated first networks, applies the anchor signal and the mixed signal of each speaker to the first network corresponding to each speaker, and reconstructs the reconstructed mask. Estimate each. Then, the reconstructed mask transforming unit 52 transforms at least one of the estimated reconstructed masks based on the degree of similarity with other reconstructed masks. Specifically, the reconstructed mask transforming unit 52 transforms the reconstructed mask such that the higher the similarity to the frequency of another reconstructed mask, the lower the reliability of that frequency.
[0063]
　The conversion by the reconstruction mask conversion unit 52 means conversion so that the target reconstruction mask is not used so much as to be similar to other reconstruction masks. A reconstruction mask being similar to other reconstruction masks means trying to extract signals of similar frequencies with reconstruction masks of different speakers. However, since such signals are rare in speech, the reliability of such reconstructed masks is reduced in an attempt to improve accuracy.
[0064]
　Any method may be used by the reconstruction mask conversion unit 52 to calculate the degree of similarity. Let Sim be a function for calculating the degree of similarity, let M f,t A be a set of reconstructed masks for speaker A, and M f,t B be a set of reconstructed masks for speaker B. FIG. At this time, the frequency similarity sf is represented by Equation 4 exemplified below.
[0065]
[Number 4]

　
[0066]
　The reconstruction mask conversion unit 52 may calculate, for example, a cosine similarity as the similarity. In this case, the similarity sf is calculated by Equation 5 exemplified below.
[0067]
[Number 5]

　
[0068]
　The reconstructed mask transforming unit 52 transforms the reconstructed mask such that the higher the calculated similarity, the lower the reliability. For example, if the reconstruction mask of an arbitrary speaker is M f,t * , the reconstruction mask conversion unit 52 may convert the reconstruction mask using Equation 6 exemplified below.
[0069]
[Number 6]

　
[0070]
　In Equation 6 above, α is a normalization coefficient, which is calculated by Equation 7 exemplified below.
[0071]
[Number 7]

　
[0072]
　The extractor 60 extracts the target signal using the transformed reconstruction mask.
[0073]
　Neural network input unit 10, anchor signal input unit 20, mixed signal input unit 30, learning unit 40 (more specifically, reconstruction mask estimation unit 42, signal classification unit 44, loss calculation unit 46, parameter The update unit 48), the output unit 50, the reconstruction mask conversion unit 52, and the extraction unit 60 are implemented by a computer processor that operates according to a program (signal extraction learning program).
[0074]
　Next, the operation of the signal extraction system 200 of this embodiment will be described. FIG. 7 is a flowchart showing an operation example of the signal extraction system 200 of this embodiment. Here, it is assumed that the processing of the flowchart illustrated in FIG. 3 has been executed and a reconstruction mask for each speaker has been generated.
[0075]
　The reconstructed mask transforming unit 52 transforms at least one of the estimated reconstructed masks based on the degree of similarity with other reconstructed masks (step S31). The extraction unit 60 extracts the target signal using the transformed reconstruction mask (step S32).
[0076]
　As described above, in the present embodiment, the reconstructed mask transforming unit 52 transforms at least one of the estimated reconstructed masks based on the degree of similarity with other reconstructed masks, and the extracting unit 60 , extract the target signal using the transformed reconstruction mask. Therefore, in addition to the effects of the first embodiment, it is possible to extract the utterances of each speaker in consideration of the nature of the conversation.
[0077]
　Next, an outline of the present invention will be described. FIG. 8 is a block diagram showing an overview of a signal extraction system according to the invention. A signal extraction system 80 (eg, signal extraction system 100) according to the present invention includes an anchor signal belonging to a predetermined class (eg, a speaker) and a target signal (a speaker's actual speech) belonging to that class. A first network (e.g., TDNN, A first network N1) illustrated in FIG. 2, a layer for inputting a target signal extracted by applying a mixed signal to a reconstruction mask, and outputting a result of classifying the input target signal into predetermined classes. A neural network input unit 81 for inputting a neural network (for example, the neural network N3 illustrated in FIG. 2) coupled with a second network (for example, the second network N2 illustrated in FIG. 2) having a layer to (e.g. neural network input 10) and a reconstruction mask estimator 82 (e.g. reconstruction a mask estimator 42), and a signal classifier 83 (for example, , signal classification unit 44), and a loss calculation unit 84 (for example, loss a calculation unit 46); a parameter update unit 85 (for example, a parameter update unit 48) that updates the parameters of the first network and the parameters of the second network in the neural network based on the calculation result of the loss function; and an output unit 86 (for example, the output unit 50) that outputs the first network.
[0078]
　With such a configuration, signals belonging to each class can be accurately extracted from observed signals.
[0079]
　In addition, the signal extraction system 80 (for example, the signal extraction system 200) includes a reconstruction mask conversion unit that converts at least one of the plurality of estimated reconstruction masks based on the degree of similarity with other reconstruction masks. (for example, the reconstruction mask conversion unit 52) ​​and an extraction unit (for example, the extraction unit 60) that extracts the target signal using the converted reconstruction mask.
[0080]
　Specifically, the reconstructed mask transforming unit may transform the reconstructed mask such that the higher the similarity to the frequency of another reconstructed mask, the lower the reliability of the frequency.
[0081]
　Also, the parameter updating unit 85 may update the parameters of the first network and the parameters of the second network in the neural network so as to reduce the loss calculated by the loss function.
[0082]
　Further, the neural network input unit 81 may input a neural network in which a second network having a layer in which an output corresponding to the number of classes assumed to be extracted is set is coupled.
[0083]
　For example, in the context of extracting the speech of a speaker, the reconstructed mask estimator 82 applies the anchor signal representing the speaker's utterance and the audio stream to the first network to estimate the reconstructed mask of the speaker. You may The signal classifier 83 may then apply the mixed signal to the estimated reconstruction mask to extract the spectrogram of the speaker, and apply the extracted spectrogram to a second network to classify the speaker. .
[0084]
　FIG. 9 is a schematic block diagram showing the configuration of a computer according to at least one embodiment. A computer 1000 comprises a processor 1001 , a main storage device 1002 , an auxiliary storage device 1003 and an interface 1004 .
[0085]
　The signal extraction system described above is implemented in computer 1000 . The operation of each processing unit described above is stored in the auxiliary storage device 1003 in the form of a program (signal extraction learning program). The processor 1001 reads out the program from the auxiliary storage device 1003, develops it in the main storage device 1002, and executes the above processing according to the program.
[0086]
　It should be noted that, in at least one embodiment, secondary storage device 1003 is an example of non-transitory tangible media. Other examples of non-transitory tangible media include magnetic disks, magneto-optical disks, CD-ROMs (Compact Disc Read-only memory), DVD-ROMs (Read-only memory), which are connected via the interface 1004. A semiconductor memory etc. are mentioned. Further, when this program is distributed to the computer 1000 via a communication line, the computer 1000 receiving the distribution may develop the program in the main storage device 1002 and execute the above process.
[0087]
　Also, the program may be for realizing part of the functions described above. Furthermore, the program may be a so-called difference file (difference program) that implements the above-described functions in combination with another program already stored in the auxiliary storage device 1003 .
Code explanation
[0088]
　10 neural network input unit
　20 anchor signal input unit
　30 mixed signal input unit
　40 learning unit
　42 reconstruction mask estimation unit
　44 signal classification unit
　46 loss calculation unit
　48 parameter update unit
　50 output unit
　52 reconstruction mask conversion unit
　60 extraction unit
　100, 200 signal extraction system
The scope of the claims
[Claim 1]
　A layer for inputting a mixed signal containing an anchor signal belonging to a predetermined class and a target signal belonging to the class, and a layer for outputting, as an estimation result, a reconstruction mask indicating a time-frequency region in which the target signal exists in the mixed signal. a layer for inputting the target signal extracted by applying the mixed signal to the reconstruction mask; and outputting a result of classifying the input target signal into a predetermined class. and
　applying the anchor signal and the mixed signal to the first network to reconstruct the class to which the anchor signal belongs. a reconstructed mask estimator for estimating a mask
　; applying the mixed signal to the estimated reconstructed mask to extract a target signal; and applying the extracted target signal to the second network for classifying into classes. a signal classification unit;
　a loss calculation unit that calculates a loss function between a class into which the extracted target signal is classified and a true class
　; 　A signal extraction system , comprising : a parameter updating unit for updating parameters of a network and parameters of a second network; and
　an output unit for outputting the updated first network .
[Claim 2]
　A reconstruction mask conversion unit that converts at least one of the estimated plurality of reconstruction masks based on similarity to other reconstruction masks,
　and extracts a target signal using the converted reconstruction mask. 2.
　The signal extraction system of claim 1, comprising an extractor.
[Claim 3]

　3. The signal extraction system according to claim 2 　, wherein the reconstruction mask transforming unit transforms the reconstruction mask such that the reliability of the frequency decreases as the frequency similarity of another reconstruction mask increases .
[Claim 4]

　4. Any one of claims 1 to 3, 　wherein the parameter updating unit updates the parameters of the first network and the parameters of the second network in the neural network so as to reduce the loss calculated by the loss function. A signal extraction system according to clause.
[Claim 5]

　5. Any one of claims 1 to 4, 　wherein the neural network input unit inputs a neural network coupled with a second network having layers set with outputs corresponding to the number of classes assumed to be extracted. A signal extraction system according to clause.
[Claim 6]
　A reconstructed mask estimator applies the anchor signal and the audio stream indicative of the speaker's utterance to the first network to estimate a reconstructed mask for the speaker, and a
　signal classifier applies the estimated reconstructed mask. to extract the spectrogram of the speaker, and apply the extracted spectrogram to a second network to classify the speaker
　. Signal extraction system.
[Claim 7]
　A layer for inputting a mixed signal containing an anchor signal belonging to a predetermined class and a target signal belonging to the class, and a layer for outputting, as an estimation result, a reconstruction mask indicating a time-frequency region in which the target signal exists in the mixed signal. a layer for inputting the target signal extracted by applying the mixed signal to the reconstruction mask; and outputting a result of classifying the input target signal into predetermined classes. applying an
　anchor signal and the mixed signal to the first network to estimate a reconstruction mask for the class to which the anchor signal belongs;
　applying the mixed signal to the estimated reconstruction mask to extract a target signal; applying the extracted target signal to the second network to classify into classes; classifying the
　extracted target signal into classes ; and calculating a loss function with the true class,
　updating the parameters of the first network and the parameters of the second network in the neural network based on the calculation result of the loss function, and updating the
　updated first
　A signal extraction learning method characterized by outputting one network .
[Claim 8]
　8. The signal of claim 7 , wherein 　at least one of the plurality of estimated reconstruction masks is transformed based on similarity with other reconstruction masks, and the
　target signal is extracted using the transformed reconstruction mask.
Extractive learning method.
[Claim 9]
　Outputting to a computer
　a layer for inputting a mixed signal containing an anchor signal belonging to a predetermined class and a target signal belonging to the class and a reconstruction mask indicating a time-frequency domain in which the target signal exists in the mixed signal as an estimation result. a layer for inputting the target signal extracted by applying the mixed signal to the reconstruction mask, and a result of classifying the input target signal into a predetermined class. applying the
　anchor signal and the mixed signal to the first network to obtain the class to which the anchor signal belongs A reconstruction mask estimation process for
　estimating a reconstruction mask, applying the mixed signal to the estimated reconstruction mask to extract a target signal, and applying the extracted target signal to the second network for classification into classes. loss calculation processing for calculating a loss function between the class into which the
　extracted target signal is classified and the true class; and
　the first network in the neural network based on the calculation result of the loss function and a parameter update process for updating the parameters of the second network, and
　an output process for outputting the updated first network
　.
[Claim 10]
　a reconstruction mask transformation process for transforming at least one of a plurality 　of
　estimated reconstruction masks based on a degree of similarity with other reconstruction masks; 　10. The signal extraction learning program according to claim 9
, which executes an extraction process for extracting a signal.