WO2021074973A1 - モデル生成方法、モデル生成装置、プログラム - Google Patents

モデル生成方法、モデル生成装置、プログラム Download PDF

Info

Publication number
WO2021074973A1
WO2021074973A1 PCT/JP2019/040514 JP2019040514W WO2021074973A1 WO 2021074973 A1 WO2021074973 A1 WO 2021074973A1 JP 2019040514 W JP2019040514 W JP 2019040514W WO 2021074973 A1 WO2021074973 A1 WO 2021074973A1
Authority
WO
WIPO (PCT)
Prior art keywords
data
value
model
replacement
predetermined
Prior art date
Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
Ceased
Application number
PCT/JP2019/040514
Other languages
English (en)
French (fr)
Japanese (ja)
Inventor
裕 清川
Current Assignee (The listed assignees may be inaccurate. Google has not performed a legal analysis and makes no representation or warranty as to the accuracy of the list.)
NEC Corp
Original Assignee
NEC Corp
Priority date (The priority date is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the date listed.)
Filing date
Publication date
Application filed by NEC Corp filed Critical NEC Corp
Priority to JP2021552019A priority Critical patent/JP7420144B2/ja
Priority to PCT/JP2019/040514 priority patent/WO2021074973A1/ja
Priority to US17/763,374 priority patent/US20220335964A1/en
Publication of WO2021074973A1 publication Critical patent/WO2021074973A1/ja
Anticipated expiration legal-status Critical
Ceased legal-status Critical Current

Links

Images

Classifications

    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L21/00—Speech or voice signal processing techniques to produce another audible or non-audible signal, e.g. visual or tactile, in order to modify its quality or its intelligibility
    • G10L21/02—Speech enhancement, e.g. noise reduction or echo cancellation
    • G10L21/0208—Noise filtering
    • G—PHYSICS
    • G10—MUSICAL INSTRUMENTS; ACOUSTICS
    • G10L—SPEECH ANALYSIS TECHNIQUES OR SPEECH SYNTHESIS; SPEECH RECOGNITION; SPEECH OR VOICE PROCESSING TECHNIQUES; SPEECH OR AUDIO CODING OR DECODING
    • G10L15/00—Speech recognition
    • G10L15/06—Creation of reference templates; Training of speech recognition systems, e.g. adaptation to the characteristics of the speaker's voice
    • G10L15/063—Training

Definitions

  • the present invention relates to a method, an apparatus, and a program for generating a model for removing noise from acoustic data.
  • analysis processing of acoustic data such as detecting a specific event such as an abnormality occurring in the plant may be performed from the acoustic data collected in the plant.
  • noise removal processing such as suppressing or reducing the noise in order to improve the accuracy of the analysis processing for the acoustic data.
  • a method of removing noise from acoustic data the following methods can be considered.
  • a noise removal processing method a method of separating signals based on the difference in the statistical model between the acoustic data to be analyzed and the noise can be considered.
  • filter processing such as smoothing of acoustic data or using a high-pass filter.
  • the noise removal method described above causes the following problems.
  • an object of the present invention is to provide a method, an apparatus, and a program for solving the above-mentioned problem that noise cannot be removed from acoustic data with high accuracy.
  • the model generation method which is one embodiment of the present invention, is From the actual data of the acoustic data, the replacement data in which the predetermined value in the actual data is replaced with the replacement value which is a value different from the predetermined value is generated.
  • a model for removing noise from predetermined acoustic data is generated by learning using the actual data and the replacement data of the acoustic data. It takes the configuration.
  • the model generator which is one embodiment of the present invention, is A data generation unit that generates replacement data from the actual data of acoustic data by replacing a predetermined value in the actual data with a replacement value that is a value different from the predetermined value.
  • a learning unit that learns using the actual data of the acoustic data and the replacement data to generate a model for removing noise from predetermined acoustic data, and a learning unit. With, It takes the configuration.
  • the program which is one form of the present invention is For information processing equipment A data generation unit that generates replacement data from the actual data of acoustic data by replacing a predetermined value in the actual data with a replacement value that is a value different from the predetermined value. A learning unit that learns using the actual data of the acoustic data and the replacement data to generate a model for removing noise from predetermined acoustic data, and a learning unit. To realize, It takes the configuration.
  • the present invention can accurately remove noise from acoustic data by being configured as described above.
  • FIG. 1 It is a block diagram which shows the structure of the noise removing device in Embodiment 1 of this invention. It is a figure which shows the state of the process at the time of generating the model for noise removal by the noise removal apparatus disclosed in FIG. It is a figure which shows the state of the process at the time of generating the model for noise removal by the noise removal apparatus disclosed in FIG. It is a figure which shows the state of the process at the time of generating the model for noise removal by the noise removal apparatus disclosed in FIG. It is a figure which shows the state of the process at the time of generating the model for noise removal by the noise removal apparatus disclosed in FIG. It is a figure which shows the state of the process at the time of generating the model for noise removal by the noise removal apparatus disclosed in FIG. It is a figure which shows the state of the process at the time of generating the model for noise removal by the noise removal apparatus disclosed in FIG.
  • FIG. 1 is a diagram for explaining a configuration of a noise removing device
  • FIGS. 2 to 11 are diagrams for explaining a processing operation of the noise removing device.
  • the noise removing device 10 in this embodiment is connected to a monitoring target P such as a plant. Then, the noise removing device 10 acquires acoustic data such as mechanical sounds in the plant measured by the microphone installed in the monitored target P, and generates a model for removing noise from the acoustic data. Functions as a generator. Further, the noise removing device 10 functions to remove noise from the measured acoustic data by using the generated model.
  • the noise removing device 10 outputs the noise-removed acoustic data to an analysis device (not shown), the acoustic data is analyzed by the analysis device, and the state of the monitoring target P is monitored based on the analysis result.
  • the analysis device can detect that the monitored target P is in a specific state such as an abnormality by analyzing the acoustic data from which noise has been removed.
  • the noise removing device 10 is not necessarily limited to processing the acoustic data measured from the plant, and may process any acoustic data measured at any place.
  • acoustic data such as unreproducible acoustic data, acoustic data that cannot increase the number of trials, and acoustic data that cannot measure only noise at the measurement location is desirable as a processing target.
  • Any acoustic data may be processed.
  • the model generating device may only perform the process of generating a model for removing noise from the acoustic data.
  • the noise removing device 10 is composed of one or a plurality of information processing devices including an arithmetic unit and a storage device. Then, as shown in FIG. 1, the noise removing device 10 includes a measuring unit 11, a clipping unit 12, a defect generation unit 13, a learning unit 14, and a noise removing unit 15, which are constructed by the arithmetic unit executing a program. To be equipped. Further, the noise removing device 10 includes an acoustic data storage unit 16 and a model storage unit 17 formed in the storage device. Hereinafter, each configuration will be described in detail.
  • the measurement unit 11 acquires acoustic data, which is a sound signal measured by a single microphone installed in the monitoring target P, and stores it in the acoustic data storage unit 16.
  • the measuring unit 11 acquires acoustic data measured at a sampling frequency of 44.1 kHz, and for example, as shown by reference numeral D1 in FIG. 2, digital data in which the number of samples is plotted on the horizontal axis and the amplitude is plotted on the vertical axis. Acoustic data will be acquired.
  • reference numeral D1 in FIG. 2 only the acoustic data corresponding to the period corresponding to the number of samplings of 1000 points is shown, but the period of the acquired acoustic data is not limited to such a period.
  • the measurement unit 11 is not always necessary, and the acoustic data may be stored in the acoustic data storage unit 16 in advance.
  • the clipping unit 12 (data generation unit) divides and cuts out the acoustic data stored in the acoustic data storage unit 16 for each predetermined period, and performs a process of generating a plurality of acoustic data for the predetermined period.
  • the clipping unit 12 generates 5000 divided acoustic data D2 (actual data) obtained by dividing the acoustic data D1 for 1000 sampling points into a period for 64 sampling points.
  • the clipping unit 12 randomly generates the divided acoustic data D2 cut out from the acoustic data D1 in a period of 64 consecutive sampling points.
  • the periods of the plurality of divided acoustic data D2 generated by the clipping unit 12 may overlap each other in the original acoustic data D1.
  • a window for a predetermined period such as 64 sampling points may be prepared, and the acoustic data in the window may be cut out as the divided acoustic data D2 while moving the window.
  • the clipping unit 12 is not necessarily limited to generating the divided sound data D2 for the above-mentioned period (64 points), and may generate the divided sound data D2 for a certain period. Further, the clipping unit 12 does not necessarily have to generate 5000 divided acoustic data, and may generate any number of divided acoustic data. Alternatively, the clipping 12 is not always necessary, and the acoustic data measured by the measuring unit 11 or a plurality of acoustic data stored in advance may be prepared and used as the above-mentioned divided acoustic data D2.
  • the defect generation unit 13 (data generation unit) generates defective data D3 (replacement data) in which some values are missing from each divided acoustic data D2 (actual data) generated as described above.
  • the defect generation unit 13 sets the amplitude value (predetermined value) at a predetermined sampling point (predetermined time point) in the divided acoustic data D2 shown in the upper part of FIG. 3 to a value different from the actual value.
  • the missing data D3 shown in the lower part of FIG. 3 is generated by replacing with the value (replacement value).
  • the missing value may be any value, but for example, as the missing value, the average value of each value in the same divided acoustic data D2 may be calculated and used, or another value may be copied and used.
  • the missing value in the present embodiment may be "0", but it is not necessarily limited to a value that eliminates the amplitude value such as "0", and is included in the divided acoustic data D2 which is the actual data. It may be a value different from the value of the amplitude at a predetermined sampling point of.
  • the defect generation unit 13 generates the missing data D3 by replacing only the amplitude value at one sampling point with the missing value in one divided acoustic data D2.
  • the defect generation unit 13 is not necessarily limited to replacing only the amplitude value at one sampling point with the missing value in one divided acoustic data D2.
  • the defect generation unit 13 may replace the amplitude values at the plurality of sampling points with the defect values in one divided acoustic data D2.
  • the defect generation unit 13 replaces the value of one amplitude with the missing value for each of the divided acoustic data D2 in the same manner as described above, and generates each missing data D3 corresponding to each divided acoustic data D2. At this time, the defect generation unit 13 replaces the amplitude values of different sampling points on the original acoustic data D1 before division with the missing values for each divided acoustic data D2. For example, in the example of FIG. 4, in the divided acoustic data D2 in which the period of the sampling number is 64 points, only the value of the amplitude of the 40th sampling number is replaced with the missing value.
  • the period of the divided acoustic data D2 cut out is different from the example of FIG. 3, but even if the divided acoustic data D2 is the same, the amplitude values of the different sampling points are replaced with missing values.
  • the defect generation unit 13 randomly determines the sampling points to be replaced with the missing values in each of the divided acoustic data D2, and as a result, the replacement with the missing values does not occur at many sampling points. It becomes.
  • the learning unit 14 performs network learning using the divided acoustic data D2 and the missing data D3 generated as described above, and generates a model for removing noise from predetermined acoustic data. Specifically, the learning unit 14 first generates a missing data set D3'that is a collection of a plurality of missing data D3s. At this time, as shown in FIG. 5, the learning unit 14 generates a missing data set D3'consisting of a combination of a plurality of missing data D3 in which the missing values are replaced at different sampling points. As an example, the learning unit 14 generates 100 missing data D3s as one missing data set D3'.
  • the learning unit 14 uses a plurality of missing data D3s included in the missing data set D3'as input values to be input to the model at once, and learns the plurality of missing data D3s collectively. Specifically, in the learning unit 14, for each missing data D3 in the missing data set D3', the amplitude value of the sampling point replaced with the missing value in the missing data D3 is before being replaced with such a missing value. Network learning is performed so as to predict and output a value approaching the amplitude value in the divided acoustic data D2. For example, in the example of FIG. 6, learning is performed so that the value of the missing value F approaches the value T of the actual data before being replaced with the missing value as shown by the arrow. At this time, the learning unit 14 learns to predict the value T of the actual data before being replaced by the missing value F, in particular, from the value of the amplitude of the missing data D3 other than the missing value F.
  • the learning unit 14 Prior to the above-mentioned learning, the learning unit 14 has a loss, which is the difference between the missing value in the missing data D2 and the value T of the actual data before being replaced by the missing value F in the corresponding divided acoustic data D2. Calculate the value. Then, the learning unit 14 learns a model that predicts a value that minimizes the loss value with respect to the value T of the actual data as the value of the sampling point replaced by the missing value in the missing data D2.
  • the learning unit 14 learns about a large number of missing data D3s by inputting a plurality of missing data sets D3', and predicts the value of the sampling point replaced by the missing value. To generate. Then, the learning unit 14 stores the generated model in the model storage unit 17.
  • the model generated in this way has a function of removing missing values, and can be applied to noise removal.
  • the noise removing unit 15 removes noise in predetermined acoustic data by using the model stored in the model storage unit 17. Specifically, the noise removing unit 15 first acquires the acoustic data in the monitored target P measured by the measuring unit 11 as described above. Then, the noise removing unit 15 reads out the model stored in the model storage unit 17, inputs the acquired acoustic data to the model, and acquires the output thereof. Then, the noise removing unit 15 can acquire the acoustic data from which the noise has been removed as an output. The noise removing unit 15 outputs the output acoustic data to a predetermined analysis processing device or stores it for analysis processing.
  • the noise removing device 10 acquires acoustic data D1 which is a sound signal measured by a single microphone installed in the monitoring target P (step S1). Then, as shown in FIG. 2, the noise removing device 10 randomly divides the acoustic data D1 into a period of a fixed number of samplings to generate a plurality of divided acoustic data D2 (step S2).
  • the noise removing device 10 deletes a part of the amplitude value in each divided acoustic data D2 and generates the missing data D3 corresponding to each divided acoustic data D2 (step S3).
  • the noise removing device 10 generates the missing data D3 by replacing only the amplitude value at one sampling point with the missing value for one divided acoustic data D2.
  • the noise removing device 10 generates the missing data D3 by replacing the amplitude values of different sampling points on the original acoustic data D1 before the division with the missing values for each divided acoustic data D2.
  • the missing data D3 is generated as shown in the lower figures of FIGS. 3 and 4.
  • the noise removing device 10 generates a missing data set D3'in which a plurality of missing data D3s are put together (step S4). At this time, as shown in FIG. 5, the noise removing device 10 generates a missing data set D3'of a plurality of missing data D3 in which the missing values are replaced at different sampling points.
  • the noise removing device 10 determines that each missing data D3 in each missing data set D3'is replaced with a missing value in the missing data D3 and a missing value F in the corresponding divided acoustic data D2.
  • the loss value which is the difference between the data value T and the data value T, is calculated (step S5).
  • the noise removing device 10 performs network learning using the missing data D3 and the loss value (step S6). Specifically, the noise removing device 10 uses a plurality of missing data D3s included in the missing data set D3'as input values to be input to the model at one time, and the values at the sampling points replaced with the missing values in each missing data D3. As a result, network training of the model is performed so as to predict a value that minimizes the loss value with respect to the value of the actual data before being replaced with the missing value. That is, the noise removing device 10 learns the input missing value in the missing data D3 so that the value of the sampling point replaced with the missing value in the missing data D3 is used as the teacher signal. As a result, the generated model is learned to predict the value of the actual data before being replaced by the missing value as the value at the sampling point replaced by the missing value in the missing data D3.
  • the noise removing device 10 learns about a large number of missing data D3s by inputting a plurality of missing data sets D3', and generates a model that predicts the value of the sampling point replaced by the missing value (step S7). After that, the noise removing device 10 stores the generated model in the model storage unit 17.
  • the model generated as described above has a function of removing missing values from acoustic data, and can also be applied to noise removal.
  • the noise removing device 10 acquires the acoustic data in the monitored target P measured by the measuring unit 11 (step S11). Then, the noise removing device 10 inputs the acquired acoustic data to the model stored in the model storage unit 17 (step S12), and acquires the output (step S13). Then, the noise removing device 10 outputs the output acoustic data to a predetermined analysis processing device or stores it for analysis processing.
  • the acoustic data is deleted, and a model learned to predict the value of the actual data before the deletion is generated as the value of the defective portion.
  • noise can be accurately removed from the acoustic data. Therefore, even if the acoustic data has no reproducibility, the number of trials cannot be increased, or the acoustic data cannot measure only the noise at the measurement location, the noise can be removed with high accuracy.
  • the analysis system can be improved by performing various analysis processes using the noise-removed acoustic data. For example, it can also be used for detecting the occurrence of a specific event such as an abnormality from acoustic data measured in a plant or the like.
  • the value of the amplitude at one sampling point is deleted in one divided acoustic data D2, and the value of the defective portion is predicted from the value of the actual data of the other portion. Is being generated. Therefore, since a model that calculates one predicted value from a plurality of values can be generated, a model that predicts the value of the defective portion more effectively can be generated, and noise removal can be effectively performed.
  • a plurality of divided acoustic data D2 having different defective parts are collectively learned. Therefore, it is possible to generate a model that can appropriately handle all kinds of acoustic data, and it is possible to perform noise removal more effectively.
  • the graph of FIG. 9 is an output when a model is generated by learning using acoustic data (gray line: noise addition signal) with Gaussian noise added, and acoustic data with Gaussian noise added to the model is input.
  • solid black line: signal before noise addition are shown. Looking at this graph, it can be said that the model output reproduces the signal before noise addition to some extent, and it can be seen that the noise is appropriately removed.
  • the graph of FIG. 10 learns using acoustic data (gray line: noise addition signal) in which a random impulse signal is added as noise to generate a model, and acoustic data in which random impulse noise is added to the model.
  • the output when is input (dotted line: model output) and the acoustic data before adding noise (solid black line: signal before noise addition) are shown. Looking at this graph, it can be said that the model output reproduces the signal before noise addition to some extent, and it can be seen that the noise is appropriately removed.
  • the graph of FIG. 11 learns using acoustic data (gray line: noise addition signal) obtained by adding a periodic impulse signal as noise to generate a model, and the noise of the periodic impulse signal is added to the model.
  • the output (dotted line: model output) when the acoustic data with the addition of noise is input and the acoustic data before the noise is added (solid black line: signal before noise addition) are shown. Looking at this graph, it cannot be said that the model output reproduces the signal before noise addition. That is, even if the model generated by the method in the present embodiment is used, the periodic impulse signal is not removed as noise.
  • FIGS. 12 to 14 are block diagrams showing the configuration of the model generation device according to the second embodiment
  • FIG. 14 is a flowchart showing the operation of the model generation device.
  • the outline of the configuration of the model generation device and the model generation method described in the above-described embodiment is shown.
  • the model generation device 100 is composed of a general information processing device, and is equipped with the following hardware configuration as an example.
  • -CPU Central Processing Unit
  • -ROM Read Only Memory
  • RAM Random Access Memory
  • 103 storage device
  • -Program group 104 loaded into RAM 303
  • a storage device 105 that stores the program group 304.
  • a drive device 106 that reads and writes the storage medium 110 external to the information processing device.
  • -Communication interface 107 that connects to the communication network 111 outside the information processing device -I / O interface 108 for inputting / outputting data -Bus 109 connecting each component
  • the model generation device 100 can construct and equip the data generation unit 121 and the learning unit 122 shown in FIG. 19 by acquiring the program group 104 by the CPU 101 and executing the program group 104.
  • the program group 104 is stored in, for example, a storage device 105 or a ROM 102 in advance, and the CPU 101 loads the program group 104 into the RAM 103 and executes the program group 104 as needed. Further, the program group 104 may be supplied to the CPU 101 via the communication network 111, or may be stored in the storage medium 110 in advance, and the drive device 106 may read the program and supply the program to the CPU 101.
  • the extraction unit 121 and the calculation unit 122 described above may be constructed by an electronic circuit.
  • FIG. 12 shows an example of the hardware configuration of the information processing device which is the model generation device 100, and the hardware configuration of the information processing device is not limited to the above case.
  • the information processing device may be composed of a part of the above-described configuration, such as not having the drive device 106.
  • the model generation device 100 executes the model generation method shown in the flowchart of FIG. 14 by the functions of the data generation unit 121 and the learning unit 122 constructed by the program as described above.
  • the model generator 100 is From the actual data of the acoustic data, replacement data in which the predetermined value in the actual data is replaced with a replacement value which is a value different from the predetermined value is generated (step S101). Learning is performed using the actual data of the acoustic data and the replacement data to generate a model for removing noise from the predetermined acoustic data (step S102).
  • the model generation device 100 and the model generation method in the present embodiment are configured as described above, thereby replacing a predetermined value of the acoustic data with a replacement value, and using the replacement data and the actual data from the acoustic data.
  • Appendix 2 The model generation method described in Appendix 1. Using the actual data of the acoustic data and the replacement data, the model that predicts the actual data from the replacement data is generated. Model generation method.
  • Appendix 3 The model generation method according to Appendix 1 or 2. From the substitution data, the model that predicts the predetermined value in the actual data replaced with the substitution value is generated. Model generation method.
  • Appendix 4 The model generation method according to any one of Appendix 1 to 3.
  • the difference between the replacement value and the predetermined value in the actual data replaced by the replacement value is calculated as a loss value, and the replacement value is replaced with the replacement value based on the replacement data and the loss value.
  • Appendix 5 The model generation method according to any one of Appendix 1 to 4. With respect to the actual data for one predetermined period, only the predetermined value at one time point in the actual data is replaced with the replacement value to generate the replacement data. Model generation method.
  • Appendix 6 The model generation method according to any one of Appendix 1 to 5.
  • the plurality of the replacement data are generated by replacing the predetermined value at a predetermined time in the actual data with the replacement value.
  • the model is generated by learning based on the plurality of the actual data and the plurality of the replacement data. Model generation method.
  • Appendix 7 The model generation method described in Appendix 6 A plurality of the replacement data are generated by replacing the predetermined values at different time points in the actual data with the replacement values for each of the actual data in the plurality of predetermined periods. Model generation method.
  • Appendix 8 The model generation method according to Appendix 6 or 7.
  • the model is generated by simultaneously learning the plurality of the actual data and the plurality of replacement data corresponding to each of the plurality of the actual data. Model generation method.
  • Appendix 9 The model generation method described in Appendix 8. A plurality of the actual data and a plurality of the replacement data having different time points when the predetermined value in the actual data is replaced with the replacement value are simultaneously learned to generate the model. Model generation method.
  • the replacement data in which the predetermined value in the actual data is replaced with the replacement value which is a value different from the predetermined value is generated.
  • a model for removing noise from predetermined acoustic data is generated by learning using the actual data and the replacement data of the acoustic data. Input predetermined acoustic data to the generated model and acquire the output from the model. Noise removal method.
  • a data generation unit that generates replacement data from the actual data of acoustic data by replacing a predetermined value in the actual data with a replacement value that is a value different from the predetermined value.
  • a learning unit that learns using the actual data of the acoustic data and the replacement data to generate a model for removing noise from predetermined acoustic data, and a learning unit.
  • a model generator equipped with.
  • the model generator according to Appendix 11 uses the actual data of the acoustic data and the replacement data to generate the model that predicts the actual data from the replacement data. Model generator.
  • the model generator according to any one of Appendix 11 to 11.2.
  • the learning unit calculates the difference between the substitution value and the predetermined value in the actual data replaced by the substitution value as a loss value, and based on the substitution data and the loss value, the learning unit said. Generate the model that predicts the predetermined value in the actual data that has been replaced by the replacement value. Model generator.
  • the model generator according to any one of Appendix 11 to 11.3.
  • the data generation unit generates the replacement data by substituting only the predetermined value at one time point in the actual data with the replacement value for the actual data for one predetermined period.
  • Model generator
  • the model generator according to any one of Appendix 11 to 11.4.
  • the data generation unit generates a plurality of the replacement data by replacing the predetermined value at a predetermined time in the actual data with the replacement value for each of the actual data in the plurality of predetermined periods.
  • the learning unit generates the model by learning based on the plurality of the actual data and the plurality of the replacement data. Model generator.
  • the model generator according to Appendix 11.5. The data generation unit generates a plurality of the replacement data by replacing the predetermined values at different time points in the actual data with the replacement values for each of the actual data in the plurality of predetermined periods. Model generator.
  • a data generation unit that generates replacement data from the actual data of acoustic data by replacing a predetermined value in the actual data with a replacement value that is a value different from the predetermined value.
  • a learning unit that learns using the actual data of the acoustic data and the replacement data to generate a model for removing noise from predetermined acoustic data, and a learning unit.
  • a noise removal unit that inputs predetermined acoustic data to the generated model and acquires the output from the model. Noise removal device equipped with.
  • a data generation unit that generates replacement data from the actual data of acoustic data by replacing a predetermined value in the actual data with a replacement value that is a value different from the predetermined value.
  • a learning unit that learns using the actual data of the acoustic data and the replacement data to generate a model for removing noise from predetermined acoustic data, and a learning unit.
  • a data generation unit that generates replacement data from the actual data of acoustic data by replacing a predetermined value in the actual data with a replacement value that is a value different from the predetermined value.
  • a learning unit that learns using the actual data of the acoustic data and the replacement data to generate a model for removing noise from predetermined acoustic data, and a learning unit.
  • a noise removal unit that inputs predetermined acoustic data to the generated model and acquires the output from the model.
  • Non-temporary computer-readable media include various types of tangible storage media.
  • Examples of non-temporary computer-readable media include magnetic recording media (eg, flexible disks, magnetic tapes, hard disk drives), magneto-optical recording media (eg, magneto-optical disks), CD-ROMs (Read Only Memory), CD-Rs, Includes CD-R / W, semiconductor memory (for example, mask ROM, PROM (Programmable ROM), EPROM (Erasable PROM), flash ROM, RAM (RandomAccessMemory)).
  • the program may also be supplied to the computer by various types of temporary computer readable media. Examples of temporary computer-readable media include electrical, optical, and electromagnetic waves.
  • the temporary computer-readable medium can supply the program to the computer via a wired communication path such as an electric wire and an optical fiber, or a wireless communication path.

Landscapes

  • Engineering & Computer Science (AREA)
  • Computational Linguistics (AREA)
  • Health & Medical Sciences (AREA)
  • Audiology, Speech & Language Pathology (AREA)
  • Human Computer Interaction (AREA)
  • Physics & Mathematics (AREA)
  • Acoustics & Sound (AREA)
  • Multimedia (AREA)
  • Quality & Reliability (AREA)
  • Signal Processing (AREA)
  • Artificial Intelligence (AREA)
  • Measurement Of Mechanical Vibrations Or Ultrasonic Waves (AREA)
PCT/JP2019/040514 2019-10-15 2019-10-15 モデル生成方法、モデル生成装置、プログラム Ceased WO2021074973A1 (ja)

Priority Applications (3)

Application Number Priority Date Filing Date Title
JP2021552019A JP7420144B2 (ja) 2019-10-15 2019-10-15 モデル生成方法、モデル生成装置、プログラム
PCT/JP2019/040514 WO2021074973A1 (ja) 2019-10-15 2019-10-15 モデル生成方法、モデル生成装置、プログラム
US17/763,374 US20220335964A1 (en) 2019-10-15 2019-10-15 Model generation method, model generation apparatus, and program

Applications Claiming Priority (1)

Application Number Priority Date Filing Date Title
PCT/JP2019/040514 WO2021074973A1 (ja) 2019-10-15 2019-10-15 モデル生成方法、モデル生成装置、プログラム

Publications (1)

Publication Number Publication Date
WO2021074973A1 true WO2021074973A1 (ja) 2021-04-22

Family

ID=75538672

Family Applications (1)

Application Number Title Priority Date Filing Date
PCT/JP2019/040514 Ceased WO2021074973A1 (ja) 2019-10-15 2019-10-15 モデル生成方法、モデル生成装置、プログラム

Country Status (3)

Country Link
US (1) US20220335964A1 (https=)
JP (1) JP7420144B2 (https=)
WO (1) WO2021074973A1 (https=)

Families Citing this family (1)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
CN114531696B (zh) * 2020-11-23 2024-11-19 维沃移动通信有限公司 Ai网络部分输入缺失的处理方法和设备

Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH1097278A (ja) * 1996-09-20 1998-04-14 Nippon Telegr & Teleph Corp <Ntt> 音声認識方法および装置
US20020002414A1 (en) * 2000-03-10 2002-01-03 Chang-Meng Hsiung Method for providing control to an industrail process using one or more multidimensional variables
JP2002014692A (ja) * 2000-06-28 2002-01-18 Matsushita Electric Ind Co Ltd 音響モデル作成装置及びその方法
JP2009128906A (ja) * 2007-11-19 2009-06-11 Mitsubishi Electric Research Laboratories Inc 音響信号と雑音信号とを含む混成信号の雑音を除去するための方法およびシステム
JP2013541023A (ja) * 2010-12-07 2013-11-07 ミツビシ・エレクトリック・リサーチ・ラボラトリーズ・インコーポレイテッド 試験音声信号の雑音を除去する結果として試験雑音除去音声信号内で減衰したスペクトル成分を復元するための方法
JP2015097355A (ja) * 2013-11-15 2015-05-21 キヤノン株式会社 収音装置及びその制御方法、プログラム

Family Cites Families (9)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
US6952674B2 (en) * 2002-01-07 2005-10-04 Intel Corporation Selecting an acoustic model in a speech recognition system
US7617103B2 (en) * 2006-08-25 2009-11-10 Microsoft Corporation Incrementally regulated discriminative margins in MCE training for speech recognition
JP5229478B2 (ja) * 2008-12-25 2013-07-03 日本電気株式会社 統計モデル学習装置、統計モデル学習方法、およびプログラム
JP6004792B2 (ja) * 2011-07-06 2016-10-12 本田技研工業株式会社 音響処理装置、音響処理方法、及び音響処理プログラム
JP5961530B2 (ja) * 2012-11-06 2016-08-02 日本電信電話株式会社 音響モデル生成装置とその方法とプログラム
JP6237774B2 (ja) * 2013-09-09 2017-11-29 日本電気株式会社 情報処理システム、情報処理方法及びプログラム
KR102209689B1 (ko) * 2015-09-10 2021-01-28 삼성전자주식회사 음향 모델 생성 장치 및 방법, 음성 인식 장치 및 방법
WO2019144066A1 (en) * 2018-01-22 2019-07-25 Jack Copper Systems and methods for preparing data for use by machine learning algorithms
US11270717B2 (en) * 2019-05-08 2022-03-08 Microsoft Technology Licensing, Llc Noise reduction in robot human communication

Patent Citations (6)

* Cited by examiner, † Cited by third party
Publication number Priority date Publication date Assignee Title
JPH1097278A (ja) * 1996-09-20 1998-04-14 Nippon Telegr & Teleph Corp <Ntt> 音声認識方法および装置
US20020002414A1 (en) * 2000-03-10 2002-01-03 Chang-Meng Hsiung Method for providing control to an industrail process using one or more multidimensional variables
JP2002014692A (ja) * 2000-06-28 2002-01-18 Matsushita Electric Ind Co Ltd 音響モデル作成装置及びその方法
JP2009128906A (ja) * 2007-11-19 2009-06-11 Mitsubishi Electric Research Laboratories Inc 音響信号と雑音信号とを含む混成信号の雑音を除去するための方法およびシステム
JP2013541023A (ja) * 2010-12-07 2013-11-07 ミツビシ・エレクトリック・リサーチ・ラボラトリーズ・インコーポレイテッド 試験音声信号の雑音を除去する結果として試験雑音除去音声信号内で減衰したスペクトル成分を復元するための方法
JP2015097355A (ja) * 2013-11-15 2015-05-21 キヤノン株式会社 収音装置及びその制御方法、プログラム

Also Published As

Publication number Publication date
JPWO2021074973A1 (https=) 2021-04-22
JP7420144B2 (ja) 2024-01-23
US20220335964A1 (en) 2022-10-20

Similar Documents

Publication Publication Date Title
US12154590B2 (en) Data processing method, data processing device, and program
JP5197853B2 (ja) モニタリング装置
CN108363365A (zh) 用于监测结构的状况的方法和装置
JP7162740B2 (ja) 振動検出装置、異常判定方法および異常判定システム
JP6960766B2 (ja) 雑音抑圧装置、雑音抑圧方法及びプログラム
CN116564332B (zh) 频响分析方法、装置、设备及存储介质
KR20230152936A (ko) 인공지능 및 신호처리 기술을 이용한 비접촉식 전력설비 진단 방법 및 이를 이용한 장치
JP7546850B2 (ja) 打音検査についての情報処理方法及び装置
JP7420144B2 (ja) モデル生成方法、モデル生成装置、プログラム
JP7288218B2 (ja) 異常音判定システム、異常音判定装置、及びプログラム
JP2001091414A (ja) 異常判定方法および装置
JP2015114294A (ja) 音響装置の検査装置及び音響装置の検査方法並びに音響装置の検査プログラム
JP7482016B2 (ja) 故障検知装置、方法およびプログラム
US9978393B1 (en) System and method for automatically removing noise defects from sound recordings
CN114184270A (zh) 一种设备振动数据处理方法、装置、设备及存储介质
CN114026403A (zh) 机器状态的声学分析
CN114018498A (zh) 气阀装置的漏气状态评估方法、装置、终端及存储介质
US20210033450A1 (en) Vibration Waveform DC Disturbance Removal
KR20230062189A (ko) 인공지능 기반의 전력 설비 진단 방법 및 이를 이용한 장치
JP4747323B2 (ja) 音響監視支援装置及び音響監視支援方法
JP7631741B2 (ja) 雑音評価装置、音響特性測定装置、雑音評価方法、音響特性測定方法、および雑音評価プログラム
JP7494407B1 (ja) 情報処理システム、情報処理方法、およびプログラム
TWM674483U (zh) 訊號前處理系統
CN120832565A (zh) 基于孪生神经网络的变压器异常检测方法及装置
JP2007107902A5 (https=)

Legal Events

Date Code Title Description
121 Ep: the epo has been informed by wipo that ep was designated in this application

Ref document number: 19949086

Country of ref document: EP

Kind code of ref document: A1

ENP Entry into the national phase

Ref document number: 2021552019

Country of ref document: JP

Kind code of ref document: A

NENP Non-entry into the national phase

Ref country code: DE

122 Ep: pct application non-entry in european phase

Ref document number: 19949086

Country of ref document: EP

Kind code of ref document: A1