Information processing device and information processing method

Abstract

An inference device (18) is configured to be accessible to a storage unit for storing a model that is constructed by machine learning using, as teaching data, information about a predetermined sound and a signal source of a brain wave of a first subject to whom the predetermined sound is presented, the model being configured to output information about sound inferred as would be recognized by the subject on the basis of the inputted signal source of the brain wave of the subject. The inference device (18) acquires a brain wave of a second subject to whom the predetermined sound is presented. The inference device (18) infers a signal source of the brain wave from among a plurality of regions of the brain of the second subject, on the basis of an aspect of the acquired brain wave. The inference device (18) inputs, to the model, information about the inferred signal source, and acquires information about sound that is outputted from the model and that is inferred as would be recognized by the second subject.

Description

[Technical field]

[0001] The present disclosure relates to data processing technology, and more particularly to an information processing device and an information processing method. [Background technology]

[0002] A technology has been proposed that uses the subject's brainwave or brain activity data to perform various analyses on the subject. For example, the following Patent Document 1 proposes a technology that measures the subject's brainwave and determines the subject's language acquisition level based on the measured brainwave. Also, the following Patent Document 2 proposes a technology that identifies the category of a viewed or imagined object from a brain activity signal measured while the subject is viewing or imagining an object image including an object that was not used in the decoder training. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] JP 2019-128533 A [Patent Document 2] JP 2017-076193 A Summary of the Invention [Problem to be solved by the invention]

[0004] In the conventional technology, the subject's cognitive content is determined or selected from multiple options prepared in advance based on the subject's electroencephalogram or brain activity data. Therefore, the inventors considered that it was difficult to determine to what extent a subject (including not only healthy subjects but also subjects in a vegetative state or locked-in syndrome who have difficulty expressing their will) recognized the presented voice (e.g., whether it was recognized as language, etc.) in the conventional technology.

[0005] The present disclosure has been made based on the inventor's recognition of the above problem, and one objective is to provide a technology that assists in determining how a presented voice sounds to a person. [Means for solving the problem]

[0006] In order to solve the above problems, an information processing device of one embodiment of the present disclosure is a model constructed by machine learning using teacher data including information on a predetermined sound and information on the signal source of a signal indicating brain activity of a first subject to which the predetermined sound is presented, and is a device that can access a model memory unit that stores a model that outputs information on the sound that is estimated to be recognized by the subject based on input information on the signal source of the signal indicating the brain activity of the subject, and is equipped with a brain activity acquisition unit that acquires a signal indicating the brain activity of a second subject to which the predetermined sound is presented, a signal source estimation unit that estimates the signal source of the signal indicating brain activity from among multiple regions of the second subject's brain based on the form of the signal indicating brain activity acquired by the brain activity acquisition unit, and a recognized voice acquisition unit that inputs information on the signal source estimated by the signal source estimation unit into the model and acquires information on the recognized voice output from the model, which is the sound that is estimated to be recognized by the second subject.

[0007] Another embodiment of the present disclosure is also an information processing device. The device is a model constructed by machine learning using teacher data including information on a predetermined voice and information on a signal source of a signal indicating brain activity of a first subject to which the predetermined voice is presented, and is capable of accessing a model storage unit that stores a model that outputs information on a voice that is estimated to be recognized by the subject based on the input information on the signal source of the signal indicating the brain activity of the subject, and includes a brain activity acquisition unit that acquires a signal indicating the brain activity of a second subject who recalls an arbitrary voice, a signal source estimation unit that estimates a signal source of the signal indicating brain activity from among multiple regions of the brain of the second subject based on the form of the signal indicating brain activity acquired by the brain activity acquisition unit, and a recognition voice acquisition unit that inputs the information on the signal source estimated by the signal source estimation unit to the model and acquires information on the voice estimated to be recalled by the second subject output from the model.

[0008] Yet another aspect of the present disclosure is an information processing method. This method is a model constructed by machine learning using information on a predetermined voice and information on a signal source of a signal indicating brain activity of a first subject to which the predetermined voice is presented as teacher data, and a computer that can access a model storage unit that stores a model that outputs information on a voice that is estimated to be recognized by the subject based on the input information on the signal source of the signal indicating the brain activity of the subject executes the steps of acquiring a signal indicating brain activity of a second subject to which the predetermined voice is presented, estimating a signal source of the signal indicating brain activity from among multiple regions of the brain of the second subject based on the state of the acquired signal indicating brain activity, and inputting the information on the estimated signal source into the model to acquire information on a recognized voice that is a voice that is estimated to be recognized by the second subject, output from the model.

[0009] Yet another aspect of the present disclosure is also an information processing method. This method is a model constructed by machine learning using teacher data including information on a predetermined sound and information on the signal source of a signal indicating brain activity of a first subject to which the predetermined sound is presented, and a computer that can access a model storage unit that stores a model that outputs information on a sound that is estimated to be recognized by the subject based on the input information on the signal source of the signal indicating the brain activity of the subject executes the steps of acquiring a signal indicating the brain activity of a second subject who recalls an arbitrary sound, estimating a signal source of the signal indicating brain activity from among multiple regions of the brain of the second subject based on the state of the acquired signal indicating brain activity, and inputting the information on the estimated signal source into the model to acquire information on the sound estimated to be recalled by the second subject output from the model.

[0010] Any combination of the above components and any conversion of the expressions of the present disclosure between a system, a program, a recording medium storing a program, etc. are also valid as aspects of the present disclosure. Effect of the Invention

[0011] According to the present disclosure, it is possible to assist in determining how a presented voice sounds to a person, or in determining the voice that a person imagines. [Brief description of the drawings]

[0012] [Figure 1] FIG. 1 is a diagram showing an overview of an estimation system according to an embodiment. [Diagram 2] FIG. 1 is a diagram illustrating a configuration of an estimation system according to an embodiment. [Diagram 3] FIG. 3 is a block diagram showing functional blocks of the model generating device of FIG. 2. [Figure 4] FIG. 1 is a diagram illustrating a network configuration of a speech estimation model. [Diagram 5] FIG. 3 is a block diagram showing functional blocks of the estimation device of FIG. 2. [Figure 6] FIG. 13 is a diagram showing an example of a comparison image. [Figure 7] FIG. 1 is a diagram showing a schematic diagram of a method for generating brain information. [Figure 8] FIG. 13 is a diagram showing an example of brain information. [Figure 9] FIG. 13 is a diagram showing an example of brain information. [Figure 10] 10(a) and 10(b) are graphs showing the experimental results. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0013] Before describing the configuration of the estimation system of the embodiment, an overview will be given. In the embodiment, a technology is proposed that uses a mathematical model constructed by machine learning (a neural network in the embodiment, hereinafter also referred to as a "voice estimation model") to reproduce how a presented voice sounds to a person and assist in distinguishing between the two. In the embodiment, electroencephalograms (scalp electroencephalograms) are used as a signal indicating the subject's brain activity. As will be described in detail later, magnetoencephalograms may also be used as a signal indicating the subject's brain activity, or the results of measurements made by a near-infrared spectroscopy (NIRS) brain measuring device may also be used.

[0014] FIG. 1 is a diagram showing an outline of an estimation system according to an embodiment. In a learning phase, the estimation system according to an embodiment measures the electroencephalogram of a first subject by having the first subject listen to a predetermined sound such as "a" or "i" (hereinafter also referred to as "original sound"), and estimates a signal source of the electroencephalogram. The estimation system receives input of signal source information based on signal source information and original sound information related to the first subject, and generates a voice estimation model that outputs information on a sound (hereinafter also referred to as "recognized sound") that is estimated to be recognized by a person to whom the original sound is presented. The original sound does not have to be a linguistic sound. The original sound may be, for example, an animal's cry, or a meaningless mechanical sound.

[0015] In the estimation phase, the estimation system of the embodiment has the second subject listen to the original voice, measures the second subject's electroencephalogram, and estimates the signal source of the electroencephalogram. The estimation system inputs signal source information about the second subject to a voice estimation model, and obtains information about the voice (recognized voice) that is estimated to be recognized by the second subject when the original voice is presented from the voice estimation model. By playing back the recognized voice, the estimation system can clarify how the original voice sounds to the second subject.

[0016] In the embodiment, the first subject and the second subject are the same person. For example, the first subject and the second subject may be a single healthy person (a person who can understand speech and express their will). The first subject and the second subject may also be a person who has difficulty expressing their will (which can also be called communication), such as a person with hearing impairment, a person in a vegetative state, or a person with locked-in syndrome. As will be described later, in a variant, the first subject and the second subject may be different people. The "subject" in the embodiment may also be a "participant" in the experiment.

[0017] In the estimation phase, the estimation system analyzes the data in the speech estimation model to which the signal source information about the second subject has been input, and visualizes information processing in the brain. Specifically, the estimation system generates brain information that indicates the degree of influence of each of the multiple brain regions of the second subject on the recognized voice. This makes it possible to visualize which areas of the brain are being used at what timing for each individual person.

[0018] FIG. 2 shows a configuration of an estimation system 10 according to an embodiment. The estimation system 10 includes an electroencephalograph 12, an fMRI (functional Magnetic Resonance Imaging) device 14, a model generating device 16, The information processing system includes an estimation device 18. In the embodiment, the devices in Fig. 2 are connected via a communication network such as a LAN, and data is transmitted and received online. As a modified example, data may be exchanged offline via a recording medium such as a USB storage.

[0019] The electroencephalograph 12 detects signals indicative of the subject's brain waves (hereinafter referred to as "brain wave signals") via a plurality of electrodes (in other words, sensors) arranged on the subject's scalp. The number of electrodes can be changed as appropriate, but in this embodiment, there are 30 electrodes. That is, the electroencephalograph 12 detects 30 channels of brain wave signals. The electroencephalograph 12 outputs data indicative of the detected 30 channels of brain wave signals to the model generation device 16 in the learning phase, and outputs the data to the estimation device 18 in the estimation phase.

[0020] The data representing the electroencephalogram signal may be, for example, data in which time and amplitude are associated with each other. It may also be data in which frequency and power spectrum density are associated with each other, i.e., data indicative of frequency characteristics. The electroencephalograph 12 may amplify the electroencephalogram signal and remove noise from the electroencephalogram signal by a known method.

[0021] The fMRI device 14 is a device that uses MRI (Magnetic Resonance Imaging) to visualize hemodynamic responses related to brain activity. The fMRI device 14 outputs brain activity data, which is data indicating active brain regions in the subject's brain, to a model generation device 16 in the learning phase, and outputs the data to an estimation device 18 in the estimation phase. The brain activity data can also be said to be data indicating the signal source of electroencephalograms based on actual measurements.

[0022] The model generation device 16 is an information processing device (in other words, a computer device) that generates a speech estimation model. The estimation device 18 is an information processing device that estimates the recognized speech of the subject by using the speech estimation model generated by the model generation device 16. The detailed configurations of these devices will be described later.

[0023] There is no limit to the number of housings for each device in Fig. 1. For example, at least one device shown in Fig. 1 may be realized by a plurality of information processing devices working together. Furthermore, the functions of the plurality of devices shown in Fig. 1 may be realized by a single information processing device. For example, the function of the model generating device 16 and the function of the estimation device 18 may be implemented in a single information processing device.

[0024] Fig. 3 is a block diagram showing functional blocks of the model generating device 16 in Fig. 2. The model generating device 16 includes an fMRI result acquiring unit 20, a signal source estimation function generating unit 22, a signal source estimation function storage unit 24, an electroencephalogram acquiring unit 26, a signal source estimation unit 28, a voice information acquiring unit 30, a learning unit 32, and a model output unit 34.

[0025] Each block shown in the block diagrams in this specification can be realized in hardware terms by elements and mechanical devices such as a computer's CPU and memory, and in software terms by a computer program, etc., but here, functional blocks realized by the cooperation of these are depicted. Those skilled in the art will understand that these functional blocks can be realized in various forms by combining hardware and software.

[0026] 3 may be stored in a predetermined recording medium and installed in the storage of the model generation device 16 via the recording medium. Alternatively, the computer program may be downloaded from a server via a communication network and installed in the storage of the model generation device 16. The CPU of the model generation device 16 may load the computer program into a main memory and execute it to fulfill the functions of the multiple functional blocks shown in FIG.

[0027] The fMRI result acquisition unit 20 acquires brain activity data of the subject (the above-mentioned first subject) input from the fMRI device 14. In the embodiment, "acquiring data" includes receiving data transmitted from an external source, and also includes storing the received data in a memory or storage.

[0028] In the embodiment, a plurality of sites (also called "areas") are defined by dividing the surface of the brain (e.g., the cerebral cortex) into predetermined sizes, and 100 sites are defined in the embodiment. These multiple sites may include well-known sites such as the amygdala, insular cortex, anterior cingulate gyrus, etc., or may include sites obtained by further dividing these well-known sites. The signal source estimation function generating unit 22 generates a signal source estimation function for estimating the signal source of the electroencephalogram from the electroencephalogram data.

[0029] The signal source estimation function in the embodiment is a function that receives 30-channel electroencephalogram data as input and outputs data indicating the presence or absence of activity of each of 100 brain regions. In other words, it is a function that outputs data indicating whether each brain region is a signal source or not. The signal source estimation function may be a 30×100 matrix that weights each of the 100 brain regions as a signal source from the 30-channel electroencephalogram data.

[0030] The above processing by the signal source estimation function generating unit 22 is realized by VBMEG (Variational Bayesian Multimodal EncephaloGraphy), which is a known software provided by Advanced Telecommunications Research Institute International. VBMEG estimates the cortical current of the brain based on electroencephalogram data. The electroencephalogram data may be data showing the waveform of an electroencephalogram, data showing a transition of amplitude over time, or data showing frequency characteristics of an electroencephalogram.

[0031] Specifically, the signal source estimation function generating unit 22 inputs, into a predetermined API (Application Programming Interface) provided by the VBMEG, (1) a function indicating the structure of the brain imaged by fMRI. By inputting image data, (2) brain activity data measured by fMRI, (3) data indicating the placement positions of the electrodes on the scalp, and (4) EEG signal data acquired by the EEG acquisition unit 26, the VBMEG generates a signal source estimation function.

[0032] The signal source estimating function generating unit 22 stores the generated signal source estimating function in a signal source estimating function storage unit 24. The signal source estimating function storage unit 24 is a storage area for storing the signal source estimating function generated by the signal source estimating function generating unit 22.

[0033] The electroencephalogram acquiring unit 26 may also be called a brain activity acquiring unit. The electroencephalogram acquiring unit 26 acquires electroencephalogram signal data input from the electroencephalograph 12 as a signal indicating the subject's brain activity. The electroencephalogram acquiring unit 26 outputs the acquired electroencephalogram signal data to the signal source estimation function generating unit 22 and the signal source estimation unit 28.

[0034] The signal source estimation unit 28 estimates one or more signal sources of the electroencephalogram from among multiple parts of the brain of the subject (here, the first subject) based on the form of the electroencephalogram as a signal indicating the brain activity of the subject acquired by the electroencephalogram acquisition unit 26. The signal source estimation unit 28 may estimate one or more signal sources of the electroencephalogram based on spatiotemporal information regarding the electroencephalogram of the first subject (for example, the shape of the electroencephalogram waveform, the position on the scalp where the electroencephalogram was measured, the positions of multiple signal sources, the shape of the unevenness of the cerebral cortex (so-called brain wrinkles), the conductivity of the tissue between the scalp and the brain, etc.).

[0035] In the embodiment, the signal source estimation unit 28 acquires data relating to one or more signal sources (hereinafter also referred to as "signal source data") as an output of the signal source estimation function by inputting 30 channels of electroencephalogram data to the signal source estimation function stored in the signal source estimation function storage unit 24. The signal source estimation unit 28 passes the acquired signal source data to the learning unit 32 as an estimation result.

[0036] The signal source data output by the signal source estimation unit 28 is data indicating the time series transition of the signal strength (magnitude of current) of the electroencephalogram output from each of a plurality of (candidate) signal sources. Specifically, the signal source data is data indicating the signal strength at 77 points within 0.3 seconds for the electroencephalogram output from each of 100 predetermined signal sources. The same applies to the signal source data output by the signal source estimation unit 50 described later.

[0037] The voice information acquisition unit 30 acquires data of original voice presented to a subject whose brainwaves are measured by the electroencephalograph 12 and whose brain activity is measured by the fMRI device 14 from an external storage device or the like. The voice information acquisition unit 30 performs a known voice analysis (e.g., Mel-cepstral analysis) on the original voice data acquired from outside to generate original voice information that is information indicating the time series transition of multiple features (acoustic features) related to the original voice. As shown in Fig. 1, the original voice information in the embodiment is time series data of five features.

[0038] The learning unit 32 can also be called a model generation unit, and generates a voice estimation model by a known machine learning method (deep learning in this embodiment) using the original voice information acquired by the voice information acquisition unit 30 and the signal source data estimated by the signal source estimation unit 28 as teacher data. The voice estimation model is a convolutional neural network that receives as input signal source data of the electroencephalogram of a subject to which a voice is presented, and outputs information on the voice (recognized voice) that is estimated to be recognized by the subject. The learning unit 32 may generate the voice estimation model using a known library or framework such as Keras.

[0039] As a modified example, the machine learning process such as deep learning (e.g., generation of a speech estimation model) executed by the learning unit 32 may be executed on a computer on the cloud (a cloud computer). In this case, the model generation device 16 may pass the teacher data to the cloud computer via a communication network, obtain a learning result (e.g., a speech estimation model) by the cloud computer, and provide it to the estimation device 18. The estimation device 18 may estimate the recognized speech of the subject using the learning result by the cloud computer.

[0040] Fig. 4 shows the network configuration of the speech estimation model. The speech estimation model includes an input layer 100, multiple convolution layers 102, a max pooling layer 104, a fully connected layer 106, and an output layer 108. Time series data of signal strength (signal strength at 77 points in time) for each of 100 signal sources indicated by the signal source data is input to the input layer 100. Recognized speech information is output from the output layer 108, which is time series data of multiple feature amounts related to the recognized speech, and indicates values ​​at 60 points in time for five feature amounts in the example of Fig. 4.

[0041] Returning to FIG. 3, the model output unit 34 transmits the data of the speech estimation model generated by the learning unit 32 to the estimation device 18, and causes the model storage unit 40 of the estimation device 18 to store the data of the speech estimation model.

[0042] Fig. 5 is a block diagram showing functional blocks of the estimation device 18 in Fig. 2. The estimation device 18 includes a model storage unit 40, an fMRI result acquisition unit 42, a signal source estimation function generation unit 44, a signal source estimation function storage unit 46, an EEG acquisition unit 48, a signal source estimation unit 50, a recognized voice estimation unit 52, a recognized voice storage unit 54, an output unit 56, a brain information generation unit 62, and a brain information storage unit 64.

[0043] A computer program in which the functions of at least some of the functional blocks shown in Fig. 5 are implemented may be stored in a predetermined recording medium and installed in the storage of the estimating device 18 via the recording medium. Alternatively, the computer program may be downloaded from a server via a communication network and installed in the storage of the estimating device 18. The CPU of the estimating device 18 may perform the functions of the functional blocks shown in Fig. 5 by reading the computer program into a main memory and executing it.

[0044] The model storage unit 40 stores data of the speech estimation model transmitted from the model generation device 16. As a modified example, the model generation device 16 may be configured to include a storage unit that stores the speech estimation model, and in this case, the estimation device 18 may refer to the speech estimation model stored in the model generation device 16 via a communication network. That is, it is sufficient for the estimation device 18 to be able to access a local or remote storage unit that stores the speech estimation model, in other words, it is sufficient for the estimation device 18 to be able to refer to the speech estimation model stored in a local or remote storage unit.

[0045] The fMRI result acquisition unit 42, the signal source estimation function generation unit 44, the signal source estimation function storage unit 46, and the electroencephalogram acquisition unit 48 correspond to the fMRI result acquisition unit 20, the signal source estimation function generation unit 22, the signal source estimation function storage unit 24, and the electroencephalogram acquisition unit 26 of the already-described model generation device 16. Therefore, for the fMRI result acquisition unit 42, the signal source estimation function generation unit 44, the signal source estimation function storage unit 46, and the electroencephalogram acquisition unit 48, descriptions of contents in common with the corresponding functional blocks will be omitted as appropriate, and differences from the corresponding functional blocks will mainly be described.

[0046] The fMRI result acquisition unit 42 acquires brain activity data of the subject who is the target of estimating the recognized voice (i.e., the second subject to whom the original voice was presented) input from the fMRI device 14. The brain wave acquisition unit 48 can also be said to be a brain activity acquisition unit. The brain wave acquisition unit 48 acquires data of the brain wave signal of the second subject to whom the original voice was presented as a signal indicating the subject's brain activity.

[0047] The signal source estimation function generating unit 44 generates a signal source estimation function for estimating the signal source of the electroencephalogram from the electroencephalogram signal data of the second subject. The signal source estimation function storage unit 46 stores the signal source estimation function for the second subject. In the embodiment, since the first subject and the second subject are the same person, the estimation device 18 may use the signal source estimation function generated by the model generating device 16 (in other words, generated in the learning phase), and the signal source estimation function storage unit 46 may store the signal source estimation function generated by the model generating device 16.

[0048] The signal source estimation unit 50 estimates one or more signal sources of the electroencephalogram from among multiple parts of the brain of the subject (here, the second subject) based on the form of the electroencephalogram as a signal indicating the brain activity of the subject acquired by the electroencephalogram acquisition unit 48. The signal source estimation unit 50 may estimate one or more signal sources of the electroencephalogram based on spatiotemporal information regarding the electroencephalogram of the second subject (for example, the shape of the electroencephalogram waveform, the position on the scalp where the electroencephalogram was measured, the positions of multiple signal sources, the shape of the unevenness of the cerebral cortex (so-called brain wrinkles), the conductivity of the tissue between the scalp and the brain, etc.).

[0049] In the embodiment, the signal source estimation unit 50 acquires signal source data relating to one or more signal sources as an output of the signal source estimation function by inputting 30 channels of electroencephalogram data to the signal source estimation function stored in the signal source estimation function storage unit 46. The signal source estimation unit 50 passes the acquired signal source data to the recognized speech estimation unit 52 as an estimation result.

[0050] The recognized speech estimation unit 52, which can also be referred to as a recognized speech acquisition unit, reads data of the speech estimation model stored in the model storage unit 40 into the main memory, and inputs signal source data estimated by the signal source estimation unit 50 to the input layer of the speech estimation model. The recognized speech estimation unit 52 acquires time-series data (the above-mentioned recognized speech information) relating to the features of the recognized speech estimated to be recognized by the second subject, which is output from the output layer of the speech estimation model.

[0051] The recognized speech storage unit 54 stores the recognized speech information acquired by the recognized speech estimation unit 52. The recognized speech estimation unit 52 may store the recognized speech information in the recognized speech storage unit 54, or the recognized speech storage unit 54 may acquire the recognized speech information from the recognized speech estimation unit 52 and store it. Furthermore, the recognized speech storage unit 54 may be a volatile storage area or a non-volatile storage area.

[0052] The output unit 56 outputs the recognized voice information acquired by the recognized voice estimation unit 52 to the outside, and in this embodiment, outputs the recognized voice information stored in the recognized voice storage unit 54 to the outside. The output unit 56 includes a reproduction unit 58 and an image generation unit 60. The reproduction unit 58 executes a known voice synthesis process on the recognized voice information acquired by the recognized voice estimation unit 52 and stored in the recognized voice storage unit 54 in this embodiment, thereby reproducing the voice indicated by the recognized voice information and outputting the reproduced voice from a speaker (not shown).

[0053] The image generating unit 60 acquires data of the voice (i.e., the original voice) presented to the second subject from an external storage device (not shown). The image generating unit 60 performs a known Mel-Cepstrum analysis on the original voice to generate time series data ("original voice information") indicating the transition of multiple feature quantities of the original voice. The image generating unit 60 also reads the recognized voice information stored in the recognized voice storage unit 54, i.e., the time series data indicating the transition of multiple feature quantities of the recognized voice. The image generating unit 60 generates an image (hereinafter also referred to as a "comparison image") indicating both the waveform of the original voice and the waveform of the recognized voice based on the original voice information and the recognized voice information.

[0054] Fig. 6 shows an example of a comparison image. In the figure, the waveform of the original voice is indicated by a dashed line, and the waveform of the recognized voice is indicated by a solid line, for each of "a", "i", and noise. In the example of Fig. 6, the image generating unit 60 generates a comparison image in which the features are superimposed for each of "a", "i", and noise.

[0055] The image generating unit 60 may store the generated data of the comparison image in a local or remote storage unit, or may output the generated data of the comparison image to a display device (not shown) and display the comparison image on the display device.

[0056] The brain information generation unit 62 generates brain information indicating the degree of influence of each of the multiple regions of the brain of the second subject on the recognized voice, by referring to the information recorded in the voice estimation model to which the signal source data has been input by the recognized voice estimation unit 52. The brain information generation unit 62 stores the generated brain information in the brain information storage unit 64. The brain information storage unit 64 is a storage area for storing the brain information generated by the brain information generation unit 62. The output unit 56 outputs the brain information stored in the brain information storage unit 64 to a local or remote storage device for storage, or outputs the brain information to a local or remote display device for display.

[0057] FIG. 7 shows a schematic diagram of a method for generating brain information. As mentioned above, the speech estimation model includes an input layer 100, multiple convolution layers 102, a maximum pooling layer 104, a fully connected layer 106, and an output layer 108. The multiple convolution layers 102 extract signal sources that have a large influence on the recognized voice by multiple filtering processes without a pooling layer in between. The convolution layer 110 is the last layer among the multiple consecutive convolution layers 102, and is the layer in which information (which can also be called weights) regarding the influence of each of the multiple brain regions (i.e., multiple signal sources) on the recognized voice is most clearly recorded.

[0058] The brain information generation unit 62 receives the signal source data from the recognized speech estimation unit 52, references the speech estimation model which outputs the recognized speech information, reads out the information recorded in the convolution layer 110, in other words, the information (which can also be called weight information) output from the convolution layer 110, and generates an array 70. In Fig. 7, the array 70 is illustrated as a one-dimensional array of 100 signal sources x 32 channels x 67 time points.

[0059] The brain information generating unit 62 prestores the correspondence between each of a plurality of brain regions (e.g., MOG, IOG, FFG, etc. shown in FIG. 7) and one or more signal sources. Note that the left brain and the right brain are treated as separate regions even if they have the same name. For example, L in FFG in FIG. 7 is FFG of the left brain, and R in FFG in FIG. 7 is FFG of the right brain. The brain information generating unit 62 generates brain information indicating the degree of influence each brain region had on the recognized voice, based on information about the corresponding one or more signal sources for each of a plurality of brain regions.

[0060] Specifically, for each region in the brain, the brain information generating unit 62 calculates the average value of 32×67 numerical values ​​(values ​​indicating the magnitude of the influence on the recognized voice) for each signal source as information on one or more corresponding signal sources. The brain information generating unit 62 records the average value for each region in the brain in the brain information as a value indicating the degree of influence of each region in the brain on the recognized voice.

[0061] In brain information 71 of Figure 7, when the original voice is "a," the influence of each brain area on the recognized voice is indicated by the length of index 72. In addition, in brain information 71, when the original voice is "i," the influence of each brain area on the recognized voice is indicated by the length of index 74. In brain information 71, when the original voice is noise (white noise), the influence of each brain area on the recognized voice is indicated by the length of index 76. The longer index 72, index 74, and index 76 indicate that the corresponding voice is being processed more actively.

[0062] 8 and 9 show examples of brain information. FIG. 8 shows brain information 71 generated when the second subject listens to original sounds such as "a" and "i", i.e., brain information 71 showing the brain area processing the sound when listening to the original sound. On the other hand, FIG. 9 shows brain information 71 generated when the second subject remembers an original sound heard in the past, i.e., brain information 71 showing the brain area processing the sound when remembering an original sound heard in the past. Indicators 72, 74, and 76 in FIGS. 8 and 9 correspond to indicators 72, 74, and 76 in FIG. 7.

[0063] Comparing the brain information 71 in Figure 8 with the brain information 71 in Figure 9, the difference in brain processing when listening to a sound and when remembering a sound becomes clear. For example, areas 80, 82, and 84 tend to be used both when listening to a sound and when remembering a sound. On the other hand, areas 86, 88, 90, and 92 are thought to be areas that are differently needed when listening to a sound and when remembering a sound.

[0064] The brain information 71 makes it possible to visualize in real time which areas of the brain are active when listening to a sound and when remembering a sound. This makes it possible to visualize changes in brain activity due to differences in the subject's awareness when listening to a sound. For example, by visualizing the brain activity of a person who has difficulty hearing a sound using the brain information 71, it is possible to understand which areas of the brain are weakly active. In addition, by visualizing the brain activity of the subject in real time using the brain information 71 while presenting a sound to the subject, it is possible to obtain information useful for improving hearing function.

[0065] The operation of the estimation system 10 having the above configuration will now be described. First, the operation of the learning phase mainly performed by the model generating device 16 will be described. The fMRI device 14 measures the brain activity of the first subject and outputs the measured brain activity data to the model generating device 16. The fMRI result acquiring unit 20 of the model generating device 16 acquires the brain activity data, and the signal source estimation function generating unit 22 starts VBMEG to generate a signal source estimation function. Specifically, the signal source estimation function generating unit 22 generates a signal source estimation function for the first subject by calling a known function provided by VBMEG using the brain structure data, electrode positions, and brain activity data of the first subject as parameters. The signal source estimation function generating unit 22 stores the signal source estimation function for the first subject in the signal source estimation function storage unit 24.

[0066] The electroencephalograph 12 measures the electroencephalogram of the first subject via electrodes placed on the scalp of the first subject to which an original sound (e.g., "a," "i," or noise) is presented. The electroencephalograph 12 outputs the electroencephalogram data of the first subject to the model generation device 16. The electroencephalogram acquisition unit 26 of the model generation device 16 acquires the electroencephalogram data of the first subject, and the signal source estimation unit 28 estimates the signal source of the electroencephalogram of the first subject according to the electroencephalogram data of the first subject and the signal source estimation function stored in the signal source estimation function memory unit 24.

[0067] The learning unit 32 generates teacher data that associates the original voice presented to the first subject with the signal source data of the electroencephalogram of the first subject, and executes machine learning based on the teacher data. The learning unit 32 receives the signal source data as an input, and generates a voice estimation model that outputs recognized voice information that is assumed to be recognized by the subject presented with the original voice. The model output unit 34 transmits data of the voice estimation model to the estimation device 18, and stores it in the model storage unit 40 of the estimation device 18.

[0068] Next, the operation of the estimation phase, which is mainly performed by the estimation device 18, will be described. The fMRI device 14 measures the brain activity of the second subject and outputs the measured brain activity data to the estimation device 18. The fMRI result acquisition unit 42 of the estimation device 18 acquires the brain activity data, and the signal source estimation function generation unit 44 starts VBMEG to generate a signal source estimation function. Specifically, the signal source estimation function generation unit 44 generates a signal source estimation function for the second subject by calling a known function provided by VBMEG using the brain structure data, electrode positions, and brain activity data of the second subject as parameters. The signal source estimation function generation unit 44 stores the signal source estimation function for the second subject in the signal source estimation function storage unit 46.

[0069] The electroencephalograph 12 measures the electroencephalogram of the second subject via electrodes placed on the scalp of the second subject to which the original sound is presented. The electroencephalograph 12 outputs the electroencephalogram data of the second subject to the estimation device 18. The electroencephalogram acquisition unit 48 of the estimation device 18 acquires the electroencephalogram data of the second subject, and the signal source estimation unit 50 estimates the signal source of the electroencephalogram of the second subject according to the electroencephalogram data of the second subject and the signal source estimation function stored in the signal source estimation function storage unit 46.

[0070] The recognized speech estimation unit 52 reads the speech estimation model stored in the model storage unit 40. The recognized speech estimation unit 52 inputs signal source data of the electroencephalogram of the second subject to the speech estimation model, and acquires the recognized speech information regarding the second subject output from the speech estimation model. The recognized speech estimation unit 52 stores the recognized speech information regarding the second subject in the recognized speech storage unit 54. The playback unit 58 plays back the speech indicated by the recognized speech information stored in the recognized speech storage unit 54. The image generation unit 60 generates a comparison image that shows the waveform of the original speech and the waveform of the recognized speech indicated by the recognized speech information stored in the recognized speech storage unit 54 side by side. The image generation unit 60 causes the generated comparison image to be displayed on a local or remote display device.

[0071] The brain information generating unit 62 generates brain information indicating the degree of influence of each of the second subject's multiple regions on the recognized voice, with reference to information recorded in the voice estimation model to which the signal source data of the second subject's electroencephalogram has been input. The brain information generating unit 62 stores the generated brain information in the brain information storage unit 64. The output unit 56 outputs the brain information recorded in the brain information storage unit 64 to an external device (such as a storage device or a display device).

[0072] According to the estimation system 10 of the embodiment, information on the voice itself that is assumed to be recognized by the subject (second subject) is generated. This makes it possible to check whether the subject correctly recognizes the sound when the subject is a healthy person. Also, when the subject is in a vegetative state, locked-in syndrome, or is an infant, etc., and is unable or has difficulty in expressing his / her will, it is possible to check whether the subject can hear sounds, and whether the subject can not only hear sounds but also recognize them as language in the brain. Also, according to the estimation system 10, by playing back the recognized voice of the subject (second subject) or displaying the waveform of the original voice and the waveform of the recognized voice in a manner that allows easy comparison, it is possible to assist in determining how the presented voice sounds to the subject and to what extent it is recognized.

[0073] Furthermore, the estimation system 10 can present the original voice to a subject (second subject) wearing a hearing aid to check the subject's recognized voice. This can support the development of high-quality hearing aids based on the user's hearing recognition level. Furthermore, the estimation device 18 can also check the recognized voice when the subject is remembering the original voice, which can support the assessment of the subject's cognitive function.

[0074] Moreover, the estimation system 10 can visualize the activity of each area in the subject's brain (for example, brain information 71 in FIG. 8 and FIG. 9). By accumulating the brain information 71 of many individuals, it is possible to visualize the difference between people with impaired hearing function (for example, elderly people or people who use earphones excessively) and people with normal hearing function at the brain area level, and also to support the establishment and evaluation of a training method for maintaining healthy hearing function. After the training method is established, it is possible to diagnose in real time whether the training is effective and whether the brain function is approaching normal.

[0075] Providing brain information 71 also has the following effects: (1) It can help determine whether a person is concentrating on listening to audio in a lecture, etc. (2) It can help determine where in the brain the cause of an inability to hear a non-native language (such as English for Japanese people) lies. (3) It can help determine where in the brain the cause of mishearing audio lies. (4) It can help determine the causes of auditory hallucinations, tinnitus, etc. from the perspective of brain activity, and can support treatment using neurofeedback.

[0076] A supplementary explanation will be given regarding the estimation accuracy of the speech estimation model of the embodiment. The comparison image shown in Fig. 6 shows the experimental results of the estimation system 10, and compares the waveform of the original speech presented to a healthy subject with the waveform of the recognized speech estimated from the electroencephalogram of the healthy subject. As mentioned above, in this comparison image, the waveform of the original speech is shown by a dashed line and the waveform of the recognized speech is shown by a solid line for each of "a", "i", and noise.

[0077] The inventors measure the difference between the waveform of the original voice and the waveform of the recognized voice by the coefficient of determination R 2 When the waveforms match perfectly, R 2 =1. As a result of the experiment, the R 2 is "0.983", and the R for the voice "i" is 2 is "0.957", R in the case of noise 2 The result was 0.997. 2Since the waveforms are similar even when the difference is around 0.7, the estimation accuracy of the voice estimation model, i.e., the accuracy of voice resynthesis using EEG, can be said to be quite high.

[0078] However, R 2 Even if the coefficient of determination R is high, it is meaningless if the voice actually synthesized based on the recognized voice information (i.e., the voice reproduced by the reproduction unit 58) does not sound the same as the original voice. 2 We conducted experiments to confirm the degree to which the speech synthesized based on the recognized speech information sounds identical to the original speech.

[0079] Figures 10(a) and 10(b) are graphs showing the experimental results. Figure 10(a) shows the results when recognized speech was synthesized (played back) from the electroencephalograms of the subject when he was listening to the original speech. Figure 10(b) shows the results when recognized speech was synthesized (played back) from the electroencephalograms of the subject when he was recalling the sound he had heard after listening to the original speech. The horizontal axis shows the coefficient of determination R, which indicates the deviation between the waveform of the original speech and the waveform of the recognized speech. 2 The line graph shows the percentage of speech synthesized based on the recognized speech information that was recognized as being the same as the original speech.

[0080] According to the experimental results, in order to achieve a recognition rate of 80% or more, as shown in the line graph, R 2 It was found that a value of 0.8 to 0.85 was required. Furthermore, the histograms in Figures 10(a) and 10(b) show that the R 2 In Fig. 10(a) and Fig. 10(b), the distribution of R ≥ 0.8 is 2 The data showing this accounted for 77.2% to 79.3% or more of the total, indicating that the estimation accuracy of the speech estimation model is high overall.

[0081] The present disclosure has been described above based on examples. These examples are merely illustrative, and it will be understood by those skilled in the art that various modifications are possible in the combination of each component and each processing step, and that such modifications are also within the scope of the present disclosure. Modifications are shown below.

[0082] A first modified example will be described. In the above embodiment, the speech estimation model is realized by a neural network, but as a modified example, a mathematical model or function as the speech estimation model may be generated by other machine learning techniques. For example, the learning unit 32 of the model generation device 16 may generate a speech estimation model that estimates the recognized speech of the subject (e.g., the category of the recognized speech) based on the input signal source data by a technique of SLR (Sparse Logistic Regression) or SVM (Support Vector Machine).

[0083] A second modification will be described. In the above embodiment, VBMEG is used to estimate the signal source, but the signal source may be estimated by other methods. For example, sLORETA (standardized LORETA) The signal source may be estimated using Low-Resolution Brain Electromagnetic Tomography (sLORETA). sLORETA is a brain function imaging analysis method that superimposes neural activity in the brain obtained by electroencephalograms and magnetoencephalograms onto a brain atlas (in other words, a standard brain).

[0084] A third modified example will be described. In the above embodiment, the fMRI device 14 is used to identify the signal source of the user's electroencephalogram (in other words, brain activity), but a configuration that does not use the fMRI device 14 is also possible. For example, a configuration may be used in which the correspondence between the electroencephalogram state and the signal source is assumed or identified based on anatomical knowledge and / or the shape of the skull estimated from the three-dimensional positions of the electrodes of the electroencephalograph 12, in which case the fMRI device 14 becomes unnecessary. The signal source estimation unit 28 may estimate the signal source based on the above correspondence.

[0085] A fourth modified example will be described. In the above embodiment, the first subject and the second subject are the same person. However, as a modified example, the first subject and the second subject may be different people. For example, the first subject may be a healthy person (a person who can understand voice and express their intention), while the second subject may be a person who has difficulty expressing their intention (which can also be called communication), such as a person with hearing impairment, a person in a vegetative state, or a person with locked-in syndrome. In this case, a voice estimation model created based on the electroencephalogram (and its signal source) of the healthy person as the first subject may be used to synthesize and reproduce the voice that is assumed to be recognized by the person who has difficulty expressing their intention as the second subject.

[0086] A fifth modified example will be described. By applying the technology described in the above embodiment, an information processing device (estimation device 18) can be realized that estimates the voice recalled by the second subject when the second subject recalls (in other words, thinks of or remembers) any voice. The model storage unit 40 of the estimation device 18 of this modified example may store data of a voice estimation model constructed by machine learning using information on a plurality of types of voices (for example, Japanese "a" to "n") and information on the signal source of the electroencephalogram of the first subject to which each of the plurality of types of voices is presented as teacher data. The electroencephalogram acquisition unit 48 of the estimation device 18 may acquire the electroencephalogram of the second subject recalling any voice (for example, any of "a" to "n"). The recognized voice estimation unit 52 of the estimation device 18 may input information on the signal source of the electroencephalogram of the second subject to the above voice estimation model and acquire information on the voice (recalled voice) estimated to have been recalled by the second subject output from the above voice estimation model. According to this embodiment, it is possible to generate information on the actual voice that is assumed to have occurred to the subject (second subject).

[0087] The estimation device 18 of the fifth modified example may execute processing and output in relation to the recalled voice in the same manner as the recognized voice in the embodiment. For example, (1) the playback unit 58 of the estimation device 18 may play back the voice indicated by the information on the recalled voice. Also, (2) information regarding the degree of influence of each of the multiple brain regions of the second subject on the recalled voice may be recorded in the voice estimation model. The brain information generation unit 62 of the estimation device 18 may generate brain information indicating the degree of influence of each of the multiple brain regions of the second subject on the recalled voice by referring to the information recorded in the voice estimation model.

[0088] The evocative voice is a voice (including words) that comes to the mind of the second subject, and includes voices that the second subject does not show to the outside. The evocative voice also includes a voice (including words) that comes to the mind of the second subject unconsciously. That is, according to the estimation system 10 of the fifth modified example, information on the voice that comes to the mind of the second subject can be obtained without the second subject consciously thinking about it. For example, if the second subject mainly thinks about the facade that he / she will show to the outside world, and thinks about his / her true feelings in the back of his / her mind, information including both the voice related to the facade and the voice related to his / her true feelings can be obtained.

[0089] A sixth modified example will be described. In the above embodiment and modified examples, electroencephalograms are used as signals indicating the brain activities of the first and second subjects. In this modified example, magnetoencephalograms may be used as signals indicating the brain activities of the first and second subjects. In this case, the estimation system 10 may include a magnetoencephalograph that measures a magnetic field generated by electrical activity of the brain, instead of the electroencephalograph 12 shown in FIG. 2. The brain activity acquisition unit of the model generation device 16 and the estimation device 18 may acquire data of magnetoencephalograms measured by the magnetoencephalograph. The signal source estimation unit of the model generation device 16 and the estimation device 18 may estimate the signal source of magnetoencephalograms based on the form of magnetoencephalograms.

[0090] As another aspect of the sixth modified example, the measurement results by a NIRS brain measuring device (which may also be called optical topography (registered trademark)) may be used as signals indicating the brain activities of the first and second subjects. The NIRS brain measuring device may measure signals that are indicators of blood flow in the cerebral cortex, increases and decreases in hemoglobin, oxygen exchange, and the like. In this case, the estimation system 10 may include a NIRS brain measuring device instead of the electroencephalograph 12 shown in FIG. 2. The brain activity acquisition unit of the model generating device 16 and the estimation device 18 may acquire data of a signal measured by the NIRS brain measuring device. The signal source estimation unit of the model generating device 16 and the estimation device 18 may estimate the signal source of the signal based on the form of the signal measured by the NIRS brain measuring device.

[0091] Any combination of the above-mentioned embodiments and modifications is also useful as an embodiment of the present disclosure. A new embodiment resulting from the combination has the effects of each of the combined embodiments and modifications. In addition, it will be understood by those skilled in the art that the functions to be performed by each component described in the claims can be realized by each component shown in the embodiments and modifications alone or in cooperation with each other. [Industrial Applicability]

[0092] The technology of the present disclosure can be applied to a device or system that estimates speech that a person recognizes or recalls. [Explanation of symbols]

[0093] 10 estimation system, 18 estimation device, 26 brain wave acquisition unit, 28 signal source estimation unit, 40 model memory unit, 48 brain wave acquisition unit, 50 signal source estimation unit, 52 recognized voice estimation unit, 54 recognized voice memory unit, 58 playback unit, 60 image generation unit, 62 brain information generation unit.

Claims

1. A model constructed by machine learning using teacher data including information on a predetermined sound and information on a signal source of a signal indicating brain activity of a first subject to which the predetermined sound is presented, the model being accessible to a model storage unit that stores a model that outputs information on a sound that is estimated to be recognized by the subject based on input information on the signal source of the signal indicating the brain activity of the subject, a brain activity acquisition unit that acquires a signal indicating brain activity of the second subject to which the predetermined sound is presented; a signal source estimation unit that estimates a signal source of the signal indicating brain activity from among a plurality of regions of the brain of the second subject based on a state of the signal indicating brain activity acquired by the brain activity acquisition unit; a recognized speech acquisition unit that inputs information about the signal source estimated by the signal source estimation unit into the model and acquires information about a recognized speech output from the model, the recognized speech being a speech estimated to be recognized by the second test subject; An information processing device comprising:

2. 2 . The information processing apparatus according to claim 1 , further comprising a playback unit that plays back a voice indicated by the information on the recognized voice acquired by the recognized voice acquisition unit.

3. Further comprising a brain information generating unit, In the model into which the information about the signal source is input, information about the degree of influence of each of a plurality of brain regions of the second subject on the recognized voice is recorded, The information processing device according to claim 1 or 2, characterized in that the brain information generation unit generates brain information indicating the degree of influence of each of a plurality of areas of the brain of the second subject on the recognized voice by referring to the information recorded in the model.

4. 4. The information processing apparatus according to claim 1, further comprising an image generating unit that generates an image showing both the waveform of the predetermined voice and the waveform of the recognized voice.

5. A model constructed by machine learning using teacher data including information on a predetermined sound and information on a signal source of a signal indicating brain activity of a first subject to which the predetermined sound is presented, the model being accessible to a model storage unit that stores a model that outputs information on a sound that is estimated to be recognized by the subject based on input information on the signal source of the signal indicating the brain activity of the subject, a brain activity acquisition unit that acquires a signal indicating brain activity of the second subject who recalls an arbitrary sound; a signal source estimation unit that estimates a signal source of the signal indicating brain activity from among a plurality of regions of the brain of the second subject based on a state of the signal indicating brain activity acquired by the brain activity acquisition unit; a voice acquisition unit that inputs information about the signal source estimated by the signal source estimation unit into the model and acquires information about the voice estimated to have been recalled by the second subject, the information being output from the model; A brain information generating unit; Equipped with The model is a neural network having multiple successive convolutional layers without a pooling layer in between, The plurality of convolution layers extracts a signal source having a large influence on the speech estimated to be recalled by the second subject by performing a plurality of filtering processes, The information processing device is characterized in that the brain information generation unit generates brain information indicating the degree of influence of each of multiple areas of the second subject's brain on the sound estimated to have been recalled by the second subject, by referring to information recorded in a convolutional layer located last among the multiple convolutional layers.

6. a model constructed by machine learning using teacher data including information on a predetermined voice and information on a signal source of a signal indicating brain activity of a first subject to which the predetermined voice is presented, the model storing a model that outputs information on a voice that is estimated to be recognized by the subject based on information on the signal source of the signal indicating the brain activity of the subject that is input; acquiring a signal indicating brain activity of a second subject to which the predetermined sound is presented; estimating a signal source of the signal indicating brain activity from among a plurality of regions of the brain of the second subject based on a state of the acquired signal indicating brain activity; inputting information about the estimated signal source into the model to obtain information about a recognized speech output from the model, the recognized speech being a speech that is estimated to be recognized by the second test subject; 2. An information processing method comprising:

7. a model constructed by machine learning using teacher data including information on a predetermined voice and information on a signal source of a signal indicating brain activity of a first subject to which the predetermined voice is presented, the model storing a model that outputs information on a voice that is estimated to be recognized by the subject based on information on the signal source of the signal indicating the brain activity of the subject that is input; acquiring a signal indicating brain activity of a second subject recalling an arbitrary sound; estimating a signal source of the signal indicating brain activity from among a plurality of regions of the brain of the second subject based on a state of the acquired signal indicating brain activity; inputting information about the estimated signal source into the model and acquiring information about the speech estimated to have been recalled by the second subject, the information being output from the model; Run The model is a neural network having multiple successive convolutional layers without a pooling layer in between, The plurality of convolution layers extracts a signal source having a large influence on the speech estimated to be recalled by the second subject by performing a plurality of filtering processes, The information processing method is characterized in that the computer further executes a step of generating brain information indicating the degree of influence of each of a plurality of areas of the brain of the second subject on the sound estimated to have been recalled by the second subject, by referring to information recorded in a convolutional layer located last among the plurality of convolutional layers.

Images