Information processing device, learning device, generation method, learning method, generation program, and learning program

The information processing device efficiently generates sound signals to improve user states by vectorizing and dimensionality-reducing higher-order statistics, employing pre-trained models to quickly produce audio enhancements.

WO2026239630A1PCT designated stage Publication Date: 2026-11-19MITSUBISHI ELECTRIC CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
PCT/JP2025/027775
Authority / Receiving Office
WO · WO
Patent Type
Applications
Current Assignee / Owner
Priority Date
2025-05-15
Filing Date
2025-08-05
Publication Date
2026-11-19

AI Technical Summary

Technical Problem

Existing methods for generating sound signals based on auditory high-order statistics are time-consuming due to iterative processing.

Method used

An information processing device that utilizes an acquisition unit to vectorize and dimensionality-reduce higher-order statistics, combined with a trained model to generate sound signals quickly, using pre-trained models and techniques like Variational Autoencoders (VAE) and Generative Adversarial Networks (GANs) to improve user states such as concentration, fatigue, or relaxation.

Benefits of technology

Enables rapid generation of sound signals that enhance user conditions, such as concentration or relaxation, by leveraging pre-trained models to process biometric data and generate audio in a short time frame.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure JP2025027775_19112026_PF_FP_ABST
    Figure JP2025027775_19112026_PF_FP_ABST
Patent Text Reader

Abstract

An information processing device (100) comprises an acquisition unit (120) that acquires a sound signal generation trained model and a high-order statistic obtained by vectorizing and dimensionally compressing a high-order statistic for generating a sound signal for improving the current state of a user, and a generation unit (153) that generates a sound signal by using the dimensionally compressed high-order statistic and the sound signal generation trained model.
Need to check novelty before this filing date? Find Prior Art

Description

Information processing device, learning device, generation method, learning method, generation program, and learning program

[0001] The present disclosure relates to an information processing device, a learning device, a generation method, a learning method, a generation program, and a learning program.

[0002] Various sounds are playing around the user. The various sounds are also referred to as background sounds. Attempts have been made to control the user's concentration level and the like using background sounds. One of the feature quantities of background sounds is a high-order statistic. The high-order statistic is a high-order statistic calculated by applying an auditory filter bank and performing acoustic analysis based on auditory characteristics. This high-order statistic is referred to as an auditory high-order statistic. The user gets the impression that a sound based on the auditory high-order statistic is a sound similar to the background sound. As a method for generating a sound signal based on the auditory high-order statistic, a method using a gradient method has been proposed (see Non-Patent Document 1).

[0003] Josh H. McDermott et al. "Sound Texture Perception via Statistics of the Auditory Periphery: Evidence from Sound Synthesis", Neuron 71, 926-940, 2011 Chris Donahue et al. "ADVERSARIAL AUDIO SYNTHESIS", ICLR, 2019 Ishaan Gulrajani et al. "Improved Training of Wasserstein GANs", NIPS, 2017

[0004] In the gradient method described in Non-Patent Document 1, iterative processing is performed. Therefore, it takes a long time until a sound signal based on the auditory high-order statistic is generated.

[0005] An object of the present disclosure is to generate a sound signal in a short time.

[0006] An information processing device according to one aspect of the present disclosure is provided. The information processing device includes an acquisition unit that acquires higher-order statistics obtained by vectorizing and dimensionality-reducing higher-order statistics for generating sound signals to improve the current state of the user, and a sound signal generation trained model, and a generation unit that generates sound signals using the dimensionality-reduced higher-order statistics and the sound signal generation trained model.

[0007] According to this disclosure, sound signals can be generated in a short amount of time.

[0008] This is a diagram of a control system. This is a diagram of the hardware of an information processing device. This is a block diagram showing the functions of an information processing device. This is a block diagram showing the functions of a sound signal generation unit. This is a flowchart showing an example of processing performed by an information processing device. This is a diagram of the hardware of a learning device. This is a block diagram showing the functions of a learning device. This is a diagram showing a specific example of learning.

[0009] The embodiments will be described below with reference to the drawings.

[0010] Embodiment. Figure 1 shows a control system. The control system includes an information processing device 100, a biosensing device 200, and an output device 300. The information processing device 100, the biosensing device 200, and the output device 300 communicate via a network. The network is either a wired network or a wireless network.

[0011] The information processing device 100 is a device that executes a generation method. The biosensing device 200 measures the user's biometric information. For example, biometric information includes information such as heart rate, brain waves, gaze direction, pupil shape, and blink rate per unit time. Heart rate can be obtained from a wristband-type measuring device. Brain waves can be obtained from an EEG sensor attached to the user's head. Gaze direction, pupil shape, and blink rate per unit time can be obtained from a camera. For example, the output device 300 is a speaker, headphones, etc. The output device 300 provides sound to the user.

[0012] Next, the hardware of the information processing device 100 will be described. Figure 2 shows the hardware of the information processing device. The information processing device 100 has a processor 101, a volatile storage device 102, a non-volatile storage device 103, and an interface 104. The processor 101 controls the entire information processing device 100. For example, the processor 101 is a CPU (Central Processing Unit), FPGA (Field Programmable Gate Array), GPU (Graphics Processing Unit), etc. The processor 101 may be a multiprocessor. The information processing device 100 may also have processing circuits.

[0013] The volatile storage device 102 is the main memory of the information processing device 100. For example, the volatile storage device 102 is RAM (Random Access Memory). The non-volatile storage device 103 is the auxiliary storage device of the information processing device 100. For example, the non-volatile storage device 103 is an HDD (Hard Disk Drive) or an SSD (Solid State Drive). The interface 104 communicates with the biosensing device 200 and the output device 300.

[0014] Next, the functions of the information processing device 100 will be explained. Figure 3 is a block diagram showing the functions of the information processing device. The information processing device 100 includes a storage unit 110, an acquisition unit 120, a state identification unit 130, a higher-order statistical quantity identification unit 140, an audio signal generation unit 150, and an output unit 160.

[0015] The storage unit 110 may be implemented as a storage area reserved in a volatile storage device 102 or a non-volatile storage device 103. The storage unit 110 may also be called a memory. Some or all of the acquisition unit 120, state identification unit 130, higher-order statistical quantity identification unit 140, sound signal generation unit 150, and output unit 160 may be implemented by processing circuits. Also, some or all of the acquisition unit 120, state identification unit 130, higher-order statistical quantity identification unit 140, sound signal generation unit 150, and output unit 160 may be implemented as modules of a program executed by the processor 101. For example, the program executed by the processor 101 is also called a generation program or generation program product. For example, the generation program is recorded on a recording medium.

[0016] The memory unit 110 stores various types of information.

[0017] The acquisition unit 120 acquires the user's biological information from the biosensing device 200. The acquisition unit 120 may also acquire biological information via other devices.

[0018] The state identification unit 130 identifies the current user state based on biological information. The user state refers to the level of concentration, fatigue, or relaxation.

[0019] This section describes the process for identifying the user's state. In the following description, the user's state will be referred to as the concentration level. For example, the state identification unit 130 uses a table showing the correspondence between biometric information and concentration levels to identify the concentration level corresponding to the acquired biometric information. Alternatively, the state identification unit 130 can use a trained model to identify the concentration level corresponding to the acquired biometric information. Specifically, when the state identification unit 130 inputs the biometric information into the trained model, the trained model outputs the concentration level. For example, if the trained model is a regression model, it can output the concentration level as a continuous value. Alternatively, if the trained model is a classification model, it can output the concentration level as a discrete value. The table and the trained model are acquired by the acquisition unit 120 from the storage unit 110 or an external device. The external device is a device located outside the information processing device 100. For example, the external device may be a cloud server or external memory. The diagram of the external device is omitted. The state identification unit 130 can similarly identify the degree of fatigue or relaxation using a table or a trained model.

[0020] The acquisition unit 120 may acquire previously acquired biometric information from the storage unit 110 or an external device. This biometric information is the user's biometric information. The state identification unit 130 may identify the user's state using the acquired biometric information (i.e., current biometric information) and previously acquired biometric information. Specifically, the state identification unit 130 uses a table or a trained model to identify the user's state based on the acquired biometric information (i.e., current biometric information) and previously acquired biometric information. This allows the information processing device 100 to identify the user's state considering both current and past biometric information.

[0021] The higher-order statistics identification unit 140 identifies higher-order statistics for generating sound signals to improve the user's state, based on the identified user state. For example, if the user's concentration level is low, the higher-order statistics identification unit 140 identifies higher-order statistics for generating sound signals to increase concentration. Also, for example, if the user's fatigue level is high, the higher-order statistics identification unit 140 identifies higher-order statistics for generating sound signals to reduce fatigue. Also, for example, if the user's relaxation level is low, the higher-order statistics identification unit 140 identifies higher-order statistics for generating sound signals to increase relaxation. The higher-order statistics are auditory higher-order statistics. Therefore, hereafter, the higher-order statistics will be referred to as auditory higher-order statistics.

[0022] The process for identifying auditory higher-order statistics is described below. For example, the higher-order statistics identification unit 140 uses a table showing the correspondence between a state (e.g., concentration level) and auditory higher-order statistics to identify auditory higher-order statistics for generating sound signals to improve the user's state. Alternatively, for example, the higher-order statistics identification unit 140 uses a trained model to identify auditory higher-order statistics for generating sound signals to improve the user's state. Specifically, when the state identification unit 130 inputs the user's state to the trained model, the trained model outputs auditory higher-order statistics. The trained model may be a regression model. If the trained model is a classification model, the higher-order statistics identification unit 140 may identify auditory higher-order statistics based on the class output by the trained model. The table and the trained model are acquired by the acquisition unit 120 from the storage unit 110 or an external device.

[0023] The acquisition unit 120 may acquire previously identified auditory higher-order statistics from the storage unit 110 or an external device. The higher-order statistics identification unit 140 may identify auditory higher-order statistics for generating a sound signal to improve the user's state, based on the identified user state and previously identified auditory higher-order statistics. For example, the higher-order statistics identification unit 140 uses a table showing the correspondence between the state (e.g., concentration level), past auditory higher-order statistics, and the auditory higher-order statistics to be output to identify auditory higher-order statistics for generating a sound signal to improve the user's state. Alternatively, for example, if the higher-order statistics identification unit 140 inputs the user's state and previously identified auditory higher-order statistics into a trained model, the trained model outputs auditory higher-order statistics for generating a sound signal to improve the user's state. As will be described later, the information processing device 100 generates a sound signal. By using previously identified auditory higher-order statistics, the information processing device 100 can generate a sound signal that takes into account both the current and past situations.

[0024] Next, the functions of the sound signal generation unit 150 will be described in detail. Figure 4 is a block diagram showing the functions of the sound signal generation unit. The sound signal generation unit 150 includes a vectorization unit 151, an encoder unit 152, a generation unit 153, and a signal processing unit 154.

[0025] The vectorization unit 151 vectorizes the auditory higher-order statistics. In other words, since the auditory higher-order statistics include multiple types of statistics, the vectorization unit 151 converts the auditory higher-order statistics into a vector representation. Alternatively, the above sentence can be expressed as follows: The vectorization unit 151 performs vectorization in such a way that it forms and combines multiple types of statistics into a one-dimensional form.

[0026] The acquisition unit 120 acquires the dimensionality-reduced trained model from the storage unit 110 or an external device. The dimensionality-reduced trained model may be a trained model trained using a Variational Autoencoder (VAE).

[0027] The encoder unit 152 generates dimensionally compressed auditory higher-order statistics using vectorized auditory higher-order statistics and a dimensionally compressed trained model. Specifically, when the encoder unit 152 inputs vectorized auditory higher-order statistics to the dimensionally compressed trained model, the dimensionally compressed trained model outputs dimensionally compressed auditory higher-order statistics. These dimensionally compressed auditory higher-order statistics are also called condition vectors.

[0028] The encoder unit 152 may perform preprocessing on the vectorized auditory higher-order statistics. For example, the preprocessing may be standardization. The encoder unit 152 may generate dimensionality-reduced auditory higher-order statistics using the auditory higher-order statistics obtained by preprocessing and a dimensionality-reduced trained model.

[0029] The acquisition unit 120 may acquire dimensionally compressed auditory higher-order statistics from the encoder unit 152. The dimensionally compressed auditory higher-order statistics are obtained by vectorizing and dimensionally compressing auditory higher-order statistics used to generate sound signals to improve the current user state.

[0030] The acquisition unit 120 acquires a trained sound signal generation model from the storage unit 110 or an external device. The generation unit 153 generates a sound signal using dimensionally compressed auditory higher-order statistics and the trained sound signal generation model. Specifically, when the generation unit 153 inputs dimensionally compressed auditory higher-order statistics to the trained sound signal generation model, the trained sound signal generation model outputs a sound signal. This sound signal is a sound signal intended to improve the user's state.

[0031] The acquisition unit 120 may acquire latent vectors from the storage unit 110 or an external device. The latent vectors may be represented as vectors based on random numbers. For example, the latent vectors may be represented as a set of multiple randomly derived values. The generation unit 153 may generate sound signals using the latent vectors, dimensionality-reduced auditory higher-order statistics, and a sound signal generation trained model. Specifically, when the generation unit 153 inputs the latent vectors and dimensionality-reduced auditory higher-order statistics to the sound signal generation trained model, the sound signal generation trained model outputs a sound signal. By inputting latent vectors to the sound signal generation trained model, the sound signal generation trained model can introduce diversity to the sound signals generated based on the dimensionality-reduced auditory higher-order statistics. Note that if the dimensionality-reduced trained model is learning using VAE, the generation unit 153 does not need to use latent vectors.

[0032] An example of how to train a pre-trained model for generating sound signals will be explained later.

[0033] If the generated sound signal has multiple channels, the signal processing unit 154 performs processing to convert the generated sound signal into a single-channel sound signal. Specifically, the signal processing unit 154 adjusts the sound signal by performing averaging and variance on the generated sound signal. For example, the signal processing unit 154 shapes the sound signal of each channel so that the sound signal of each channel matches the mean and variance of the subband signal, which is a type of auditory higher-order statistic. The signal processing unit 154 applies a filter bank to the shaped sound signal of each channel and generates a single-channel sound signal based on the sound signals of all channels.

[0034] The information processing device 100 does not need to have a signal processing unit 154 if the generated sound signal is a single-channel sound signal.

[0035] The output unit 160 outputs an audio signal to the output device 300. This allows the user to hear sounds that improve their condition via the output device 300. As a result, the user's condition improves.

[0036] Next, an example of processing performed by the information processing device 100 will be explained using a flowchart. Figure 5 is a flowchart of an example of processing performed by the information processing device. (Step S11) The acquisition unit 120 acquires the user's biometric information from the biosensing device 200. (Step S12) The state identification unit 130 identifies the user's state based on the biometric information. (Step S13) The higher-order statistics identification unit 140 identifies auditory higher-order statistics for generating sound signals to improve the user's state based on the user's state. (Step S14) The vectorization unit 151 vectorizes the auditory higher-order statistics. (Step S15) The acquisition unit 120 acquires the dimensionality-reduced trained model from the storage unit 110. The encoder unit 152 generates dimensionality-reduced auditory higher-order statistics using the vectorized auditory higher-order statistics and the dimensionality-reduced trained model.

[0037] (Step S16) The acquisition unit 120 acquires the sound signal generation trained model from the storage unit 110. The generation unit 153 generates a sound signal using the dimensionality-reduced auditory higher-order statistics and the sound signal generation trained model. If the dimensionality-reduced trained model has not been trained using VAE, the generation unit 153 generates a sound signal using the latent vector, the dimensionality-reduced auditory higher-order statistics and the sound signal generation trained model. (Step S17) If the generated sound signal is a sound signal with multiple channels, the signal processing unit 154 performs processing to make the generated sound signal a single-channel sound signal. (Step S18) The output unit 160 outputs the sound signal to the output device 300.

[0038] In the gradient method described in Non-Patent Document 1, iterative processing is performed. Therefore, it takes a long time until an audio signal based on auditory higher-order statistics is generated.

[0039] According to this embodiment, the information processing device 100 generates sound signals using a pre-trained model for sound signal generation. The pre-trained model for sound signal generation does not perform iterative processing. Therefore, the pre-trained model for sound signal generation can generate sound signals in a short amount of time. Thus, the information processing device 100 can generate sound signals in a short amount of time by using the pre-trained model for sound signal generation.

[0040] The above describes a case where the embodiment is realized by a single device. The embodiment may be realized by multiple devices. For example, multiple information processing devices have an acquisition unit 120, a state identification unit 130, a higher-order statistical identification unit 140, a vectorization unit 151, an encoder unit 152, a generation unit 153, a signal processing unit 154, and an output unit 160. A specific example is given. The first information processing device has an acquisition unit 120, a state identification unit 130, a higher-order statistical identification unit 140, a vectorization unit 151, and an encoder unit 152. The second information processing device has an acquisition unit 120, a generation unit 153, a signal processing unit 154, and an output unit 160. The acquisition unit 120 of the second information processing device acquires dimensionality-reduced auditory higher-order statistics from the first information processing device. The generation unit 153 of the second information processing device generates an audio signal using the dimensionality-reduced auditory higher-order statistics and a trained model for sound signal generation. The signal processing unit 154 of the second information processing device performs processing to convert the generated sound signal into a single-channel sound signal if the generated sound signal has multiple channels. The output unit 160 of the second information processing device outputs the sound signal to the output device 300. Thus, the embodiment may be implemented using multiple devices.

[0041] Next, we will explain an example of a training method for a pre-trained model for generating sound signals. The pre-trained model for generating sound signals is generated by a training device. We will then describe the hardware of the training device.

[0042] Figure 6 shows the hardware of the learning device. The learning device 400 has a processor 401, a volatile memory device 402, and a non-volatile memory device 403. The processor 401 controls the entire learning device 400. For example, the processor 401 is a CPU, FPGA, GPU, etc. The processor 401 may also be a multiprocessor. The learning device 400 may also have processing circuits. The volatile memory device 402 is the main memory of the learning device 400. For example, the volatile memory device 402 is RAM. The non-volatile memory device 403 is the auxiliary memory of the learning device 400. For example, the non-volatile memory device 403 is an HDD or SSD.

[0043] Next, the functions of the learning device 400 will be described. FIG. 7 is a block diagram illustrating the functions of the learning device. The learning device 400 includes a storage unit 410, an acquisition unit 420, a vectorization unit 430, an encoder unit 440, a generation unit 450, and an identification unit 460. The storage unit 410 may be implemented as a storage area secured in the volatile storage device 402 or the non-volatile storage device 403. Part or all of the acquisition unit 420, the vectorization unit 430, the encoder unit 440, the generation unit 450, and the identification unit 460 may be implemented by a processing circuit. Further, part or all of the acquisition unit 420, the vectorization unit 430, the encoder unit 440, the generation unit 450, and the identification unit 460 may be implemented as modules of a program executed by the processor 401. For example, the program executed by the processor 401 is also referred to as a learning program or a learning program product. For example, the learning program is recorded on a recording medium.

[0044] The functions of the acquisition unit 420, the vectorization unit 430, the encoder unit 440, the generation unit 450, and the identification unit 460 will be described with reference to the drawings.

[0045] FIG. 8 is a diagram illustrating a specific example of learning. FIG. 8 shows a dimension compression trained model 11, a sound signal generation model 12, a discriminator 13, and a dimension compression trained model 14. The dimension compression trained model 11 may be a trained model learned using VAE.

[0046] The encoder unit 440 uses the dimension compression trained model 11. The generation unit 450 uses the sound signal generation model 12. The identification unit 460 uses the discriminator 13. The encoder unit 440 uses the dimension compression trained model 14. The dimension compression trained model 11 is also referred to as the first dimension compression trained model. The dimension compression trained model 14 is also referred to as the second dimension compression trained model.

[0047] The learning device 400 performs learning using a GAN (Generative Adversarial Networks). The generation unit 450 using the sound signal generation model 12 may be regarded as a Generator. The discrimination unit 460 using the discriminator 13 may be regarded as a Discriminator. Note that the sound signal generation model 12 may be the generator of WaveGAN described in Non-Patent Document 2. The discriminator 13 may be the discriminator of WaveGAN described in Non-Patent Document 2.

[0048] The acquisition unit 420 acquires auditory higher-order statistics of a correct sound signal that is a ground-truth sound signal. For example, the acquisition unit 420 acquires the auditory higher-order statistics from the storage unit 410. The acquisition unit 420 may acquire the auditory higher-order statistics from an external device. The external device is a device that exists outside the learning device 400. For example, the external device is a cloud server, an external memory, or the like. The diagram of the external device is omitted. The acquisition unit 420 acquires the dimension compression learned model 11 and the dimension compression learned model 14 from the storage unit 410 or the external device. The acquisition unit 420 acquires a latent vector from the storage unit 410 or the external device.

[0049] The vectorization unit 430 vectorizes the auditory higher-order statistics of the correct sound signal. The encoder unit 440 generates dimension-compressed auditory higher-order statistics using the vectorized auditory higher-order statistics and the dimension compression learned model 11. The acquisition unit 420 may acquire the dimension-compressed auditory higher-order statistics from the encoder unit 440. Note that the dimension-compressed auditory higher-order statistics are higher-order statistics obtained by vectorizing and dimension-compressing the auditory higher-order statistics of the correct sound signal.

[0050] The generation unit 450 generates a sound signal using the latent vector, the dimension-compressed auditory higher-order statistics, and the sound signal generation model 12. The discrimination unit 460 learns the discriminator 13 using the correct sound signal. Further, the discrimination unit 460 uses the discriminator 13 to determine whether or not the sound signal generated by the generation unit 450 is fake. When it is determined that the sound signal is fake, the generation unit 450 learns the sound signal generation model 12 so as to deceive the discriminator 13. By repeating the learning, the sound signal generation model 12 becomes a sound signal generation learned model.

[0051] In this way, the learning device 400 can generate a trained model for sound signal generation by repeating the learning process. That is, the generation unit 450 can generate a trained model for sound signal generation capable of generating sound signals by learning the sound signal generation model 12 in a way that deceives the discriminator 13. After the trained model for sound signal generation is generated, for example, the learning device 400 transmits the trained model for sound signal generation to the information processing device 100 via the network.

[0052] Furthermore, if the dimensionality-reduced trained model 11 is trained using VAE, the generation unit 450 does not need to train the sound signal generation model 12 using latent vectors.

[0053] The acquisition unit 420 may acquire auditory higher-order statistics of the correct sound signal from the storage unit 410 or an external device. The vectorization unit 430 may vectorize the auditory higher-order statistics of the correct sound signal. The encoder unit 440 may generate dimensionally compressed auditory higher-order statistics using the vectorized auditory higher-order statistics and the dimensionality-reduced trained model 14. The dimensionally compressed auditory higher-order statistics may be input to the identification unit 460. By inputting the dimensionally compressed auditory higher-order statistics to the identification unit 460, the identification unit 460 can easily determine whether a sound signal that conforms to the conditions has been generated.

[0054] For example, the learning loss function may be the WGAN-GP described in Non-Patent Document 3. The loss function of the sound signal generation model 12 may include the difference between the auditory higher-order statistics of the sound signal generated by the sound signal generation model 12 and the auditory higher-order statistics of the correct sound signal. The generation unit 450 may train the sound signal generation model 12 so that this difference becomes small.

[0055] During the learning phase, the learning device 400 can generate a trained model for generating sound signals that can generate sound signals in a short amount of time.

[0056] The above describes a case where the learning phase is implemented by a single device. However, the learning phase may be implemented by multiple devices. For example, multiple learning devices may have an acquisition unit 420, a vectorization unit 430, an encoder unit 440, a generation unit 450, and an identification unit 460. A specific example is given. The first learning device has an acquisition unit 420, a vectorization unit 430, and an encoder unit 440. The second learning device has an acquisition unit 420, a generation unit 450, and an identification unit 460. The acquisition unit 420 of the second learning device acquires dimensionally compressed auditory higher-order statistics from the first learning device. The generation unit 450 of the second learning device generates an audio signal using the dimensionally compressed auditory higher-order statistics and the audio signal generation model 12. The identification unit 460 of the second learning device uses a classifier 13 to determine the authenticity of the audio signal generated by the generation unit 450. If the sound signal is determined to be false, the generation unit 450 learns the sound signal generation model 12 in a way that deceives the discriminator 13. Through repeated learning, the sound signal generation model 12 becomes a learned sound signal generation model. Thus, the learning phase may be implemented by multiple devices.

[0057] The embodiments are merely examples, and various modifications are possible within the scope of this disclosure. Furthermore, the features of the embodiments can be combined with each other as appropriate.

[0058] 11-dimensional compressed pre-trained model, 12-sound signal generation model, 13-classifier, 14-dimensional compressed pre-trained model, 100-information processing device, 101-processor, 102-volatile memory device, 103-non-volatile memory device, 104-interface, 110-storage unit, 120-acquisition unit, 130-state identification unit, 140-higher-order statistical identification unit, 150-sound signal generation unit, 151-vectorization unit, 152-encoder unit, 153-generation unit, 154-signal processing unit, 160-output unit, 200-biosensing device, 300-output device, 400-learning device, 401-processor, 402-volatile memory device, 403-non-volatile memory device, 410-storage unit, 420-acquisition unit, 430-vectorization unit, 440-encoder unit, 450 Generation unit, 460 Identification unit.

Claims

1. An information processing device comprising: an acquisition unit that acquires higher-order statistics obtained by vectorizing and dimensionality-reducing higher-order statistics for generating sound signals to improve the current state of the user, and a pre-trained model for generating sound signals; and a generation unit that generates sound signals using the dimensionality-reduced higher-order statistics and the pre-trained model for generating sound signals.

2. The information processing device according to claim 1, further comprising: a state identification unit; a higher-order statistical quantity identification unit; a vectorization unit; an encoder unit; and an output unit, wherein the information processing device communicates with an output device that provides sound to the user; the acquisition unit acquires the user's biometric information and a dimensionality-reduced trained model; the state identification unit identifies the user's state based on the biometric information; the higher-order statistical quantity identification unit identifies the higher-order statistical quantity for generating a sound signal to improve the user's state based on the user's state; the vectorization unit vectorizes the higher-order statistical quantity; the encoder unit generates the dimensionality-reduced higher-order statistical quantity using the vectorized higher-order statistical quantity and the dimensionality-reduced trained model; and the output unit outputs the sound signal to the output device.

3. The information processing apparatus according to claim 2, wherein the acquisition unit acquires previously acquired biological information, and the state identification unit uses the acquired biological information and the previously acquired biological information to identify the state of the user.

4. The information processing apparatus according to claim 2 or 3, wherein the acquisition unit acquires previously identified higher-order statistics, and the higher-order statistics identification unit identifies the higher-order statistics for generating an audio signal to improve the user's state, based on the user's state and the previously identified higher-order statistics.

5. The information processing apparatus according to any one of claims 1 to 4, wherein the acquisition unit acquires a latent vector, and the generation unit generates the sound signal using the latent vector, the dimensionality-reduced higher-order statistics, and the sound signal generation trained model.

6. The information processing apparatus according to any one of claims 1 to 5, further comprising a signal processing unit that performs a process to convert the generated sound signal into a single-channel sound signal when the generated sound signal is a sound signal having multiple channels.

7. A learning device comprising: an acquisition unit that acquires higher-order statistics obtained by vectorizing and dimensionality-reducing the higher-order statistics of a correct sound signal, which is a correct sound signal; a generation unit that generates a sound signal using the dimensionality-reduced higher-order statistics and a sound signal generation model; and a discrimination unit that determines whether the sound signal is false or not using a discriminator, wherein the generation unit learns the sound signal generation model to deceive the discriminator if the sound signal is false.

8. The learning device according to claim 7, further comprising: a vectorization unit; and an encoder unit, wherein the acquisition unit acquires higher-order statistics of the correct sound signal and a first dimensionality-reduced trained model; the vectorization unit vectorizes the higher-order statistics of the correct sound signal; and the encoder unit generates the dimensionality-reduced higher-order statistics using the vectorized higher-order statistics and the first dimensionality-reduced trained model.

9. The learning device according to claim 8, wherein the acquisition unit acquires a second dimensionality-reduced trained model, the vectorization unit vectorizes the higher-order statistics of the correct sound signal, the encoder unit generates the dimensionality-reduced higher-order statistics using the vectorized higher-order statistics and the second dimensionality-reduced trained model, and the dimensionality-reduced higher-order statistics are input to the identification unit.

10. The learning device according to any one of claims 7 to 9, wherein the loss function of the sound signal generation model includes the difference between the higher-order statistics of the sound signal generated by the sound signal generation model and the higher-order statistics of the correct sound signal.

11. The learning device according to any one of claims 7 to 10, wherein the acquisition unit acquires latent vectors, and the generation unit learns the sound signal generation model using the latent vectors.

12. A method for generating an audio signal, wherein an information processing device obtains higher-order statistics obtained by vectorizing and dimensionality-reducing higher-order statistics for generating an audio signal to improve the current state of the user, and an audio signal generation trained model, and generates an audio signal using the dimensionality-reduced higher-order statistics and the audio signal generation trained model.

13. A learning method comprising: a learning device obtaining higher-order statistics obtained by vectorizing and dimensionality-reducing the higher-order statistics of a correct sound signal, which is a correct sound signal; generating a sound signal using the dimensionality-reduced higher-order statistics and a sound signal generation model; determining whether the sound signal is false or not using a classifier; and, if the sound signal is false, training the sound signal generation model to deceive the classifier.

14. A generation program that causes an information processing device to obtain higher-order statistics obtained by vectorizing and dimensionality-reducing higher-order statistics for generating sound signals to improve the current state of the user, and a trained model for generating sound signals, and to generate sound signals using the dimensionality-reduced higher-order statistics and the trained model for generating sound signals.

15. A learning program that causes a learning device to perform the following processes: obtain higher-order statistics obtained by vectorizing and dimensionality-reducing the higher-order statistics of the correct sound signal, which is the correct sound signal; generate a sound signal using the dimensionality-reduced higher-order statistics and a sound signal generation model; determine whether the sound signal is false or not using a classifier; and if the sound signal is false, train the sound signal generation model to deceive the classifier.