Wake-up word energy calculation method, system, voice wake-up system and storage medium

Abstract

An embodiment of the present invention provides a wake-up word energy calculation method, system, voice wake-up system, and storage medium, wherein the method includes: obtaining a wake-up word audio signal, performing a first conversion on the wake-up word audio signal, obtaining a short-time energy spectrum of the wake-up word audio, taking the logarithm of the short-time energy spectrum, obtaining a logarithmic spectrum of the wake-up word audio, inputting the logarithmic spectrum into a preset neural network model so that the preset neural network model generates a predicted probability matrix based on the logarithmic spectrum, binarizing the predicted probability matrix to obtain a binary matrix, performing a second conversion on the short-time energy spectrum and the binary matrix, and determining the wake-up word speech energy of the wake-up word audio signal. The present invention improves the accuracy of distinguishing between noise time-frequency points and wake-up word time-frequency points in different application scenarios by introducing a preset neural network model to estimate the wake-up word speech component, thereby improving the robustness and accuracy of the final calculation of the wake-up word energy under background noise conditions.

Description

Technical Field

[0001] The present invention relates to the field of smart home technology, and in particular to a wake-up word energy calculation method and system, a voice wake-up system, and a storage medium. Background Art

[0002] With the prevalence of smart homes, more and more home electronic devices are embedded with voice assistant functions, enabling voice control of home devices and enhancing the level of home intelligence. However, when multiple electronic devices use the same or similar wake-up keywords, it is common for a user to issue a single wake-up keyword and multiple electronic devices to respond simultaneously, affecting the user experience. Summary of the Invention

[0003] The purpose of the embodiments of the present invention is to provide a wake-up word energy calculation method, system, voice wake-up system, and storage medium to improve the calculation accuracy and robustness of wake-up word energy under background noise conditions. The specific technical solutions are as follows:

[0004] A method for calculating wake-up word energy, the method comprising:

[0005] Get the wake-up word audio signal.

[0006] Perform a first conversion on the wake-up word audio signal to obtain a short-time energy spectrum of the wake-up word audio.

[0007] The short-time energy spectrum is logarithmically taken to obtain the logarithmic spectrum of the wake-up word audio.

[0008] The logarithmic spectrum is input into a preset neural network model so that the preset neural network model generates a predicted probability matrix according to the logarithmic spectrum.

[0009] The predicted probability matrix is ​​binarized to obtain a binary matrix.

[0010] A second conversion is performed on the short-time energy spectrum and the binary matrix to determine the wake-up word speech energy of the wake-up word audio signal.

[0011] Optionally, the training process of the preset neural network model includes:

[0012] The log spectrum of the noisy speech data is input into the initial neural network for processing to obtain the predicted training probability matrix.

[0013] The error value between the training probability matrix and the label matrix is ​​calculated based on the cross entropy loss function.

[0014] According to the error value, the initial neural network is iteratively updated using a preset optimization algorithm until a training cutoff condition is met, thereby obtaining the preset neural network model.

[0015] Optionally, before training the initial neural network, the process further includes: training data processing and / or training data feature extraction.

[0016] The training data processing includes:

[0017] Performing the first conversion on the training data of the wake-up word speech to obtain a short-time energy spectrum of the training data. Taking the logarithm of the short-time energy spectrum of the training data to obtain a logarithmic spectrum of the training data. Binarizing the logarithmic spectrum of the training data to obtain the label matrix.

[0018] The training data feature extraction includes:

[0019] The noisy training data is inserted into the training data according to the signal-to-noise ratio to obtain the noisy speech data. The noisy speech data is subjected to the first conversion to obtain a short-time energy spectrum of the noisy speech data. The short-time energy spectrum of the noisy speech data is subjected to the logarithm to obtain a logarithmic spectrum of the noisy speech data.

[0020] Optionally, inputting the log spectrum into a preset neural network model so that the preset neural network model generates a predicted probability matrix according to the log spectrum includes:

[0021] The preset neural network maps the received time-frequency points of the logarithmic spectrum into the predicted probability matrix, and each element in the predicted probability matrix represents the probability value that the time-frequency point corresponding to the element belongs to the wake-up data.

[0022] Optionally, binarizing the predicted probability matrix to obtain a binary matrix includes:

[0023] The predicted probability matrix is ​​binarized according to a preset threshold value in the preset neural network model to obtain the binary matrix, wherein the binarization is to determine whether each element in the predicted probability matrix is ​​greater than the preset threshold value, and if so, set the element to 1. If the element is not greater than the preset threshold value, set the element to 0.

[0024] Optionally, the method is applied to a distributed voice wake-up system, and the method further includes:

[0025] Multiple electronic devices in the distributed voice wake-up system respectively calculate the wake-up word voice energy of the device according to the method, and compare the wake-up word voice energy of the device with the wake-up word voice energy of other devices. The device with the largest wake-up word voice energy performs the wake-up operation, and other devices except the device that performs the wake-up operation do not perform the wake-up operation.

[0026] Optionally, the wake-up word audio signal is a voice signal containing a wake-up keyword and an audio signal of a scene noise signal of a scene in which the distributed voice wake-up system is located.

[0027] A wake-up word energy calculation system is applied to a distributed voice wake-up system. The system includes:

[0028] Signal acquisition module: used to obtain the wake-up word audio signal.

[0029] The first conversion module is used to perform a first conversion on the wake-up word audio signal to obtain a short-time energy spectrum of the wake-up word audio.

[0030] The second conversion module is used to take the logarithm of the short-time energy spectrum to obtain the logarithmic spectrum of the wake-up word audio.

[0031] Matrix generation module: used to input the logarithmic spectrum into a preset neural network model so that the preset neural network model generates a predicted probability matrix based on the logarithmic spectrum.

[0032] The third conversion module is used to binarize the predicted probability matrix to obtain a binary matrix.

[0033] The fourth conversion module is configured to perform a second conversion on the short-time energy spectrum and the binary matrix to determine the wake-up word voice energy of the wake-up word audio signal.

[0034] Optionally, the above system further includes:

[0035] The model training module is configured to input the log spectrum of the noisy speech data into an initial neural network for processing to obtain a predicted training probability matrix. The module then calculates the error between the training probability matrix and the label matrix based on a cross-entropy loss function. Based on this error, the module iteratively updates the initial neural network using a preset optimization algorithm until a training cutoff condition is met, thereby obtaining the preset neural network model.

[0036] Optionally, the above system further includes:

[0037] The training data processing module is used to perform the first conversion on the training data of the wake-up word speech to obtain the short-time energy spectrum of the training data; take the logarithm of the short-time energy spectrum of the training data to obtain the logarithmic spectrum of the training data; and binarize the logarithmic spectrum of the training data to obtain the label matrix.

[0038] and / or a training data feature extraction module configured to insert the noise training data into the training data according to the signal-to-noise ratio to obtain the noisy speech data, perform the first conversion on the noisy speech data to obtain a short-time energy spectrum of the noisy speech data, and perform the logarithmization of the short-time energy spectrum of the noisy speech data to obtain a logarithmic spectrum of the noisy speech data.

[0039] Optionally, the matrix generation module is set to:

[0040] The preset neural network in the matrix generation module maps the received time-frequency points of the logarithmic spectrum into the predicted probability matrix, and each element in the predicted probability matrix represents the probability value that the time-frequency point corresponding to the element belongs to the wake-up data.

[0041] Optionally, the third conversion module is configured to:

[0042] According to the preset threshold value in the preset neural network model, the predicted probability matrix is ​​binarized to obtain the binary matrix, wherein the binarization is to determine whether each element in the predicted probability matrix is ​​greater than the preset threshold value. If so, the element is set to 1; if the element is not greater than the preset threshold value, the element is set to 0.

[0043] Optionally, the above system further includes:

[0044] The device wake-up module is used to control multiple electronic devices in a distributed voice wake-up system, calculate the wake-up word voice energy of the device respectively according to the method, and compare the wake-up word voice energy of the device with the wake-up word voice energy of other devices. The device with the largest wake-up word voice energy performs the wake-up operation, and other devices except the device that performs the wake-up operation do not perform the wake-up operation.

[0045] Optionally, the wake-up word audio signal is a voice signal containing the wake-up keyword and an audio signal of a scene noise signal of the scene in which the distributed voice wake-up system is located.

[0046] A voice wake-up system, comprising:

[0047] Multiple electronic devices are configured to execute instructions to implement the wake-up word energy calculation method as described in any one of the above.

[0048] A computer-readable storage medium, when instructions in the computer-readable storage medium are executed by a processor of an electronic device, enables the device to perform any of the wake-up word energy calculation methods described above.

[0049] The embodiments of the present invention provide a method, system, voice wake-up system and storage medium for calculating the energy of a wake-up word. By introducing a preset neural network model, the wake-up word voice component in the wake-up word audio is estimated. Compared with the prior art, the present invention improves the accuracy of distinguishing between noise time-frequency points and wake-up word time-frequency points in the wake-up word audio under different application scenarios, thereby improving the robustness and accuracy of the final calculation of the wake-up word energy under background noise conditions. At the same time, by introducing a preset neural network model, the present invention dynamically updates the internal parameters involved in calculating the energy of the wake-up word for different application scenarios, thereby improving the applicability of the present invention to different application scenarios. Finally, since the present invention can be deployed based on the existing distributed voice wake-up system, there is no need to modify the hardware equipment, which further enhances the universality of the present invention. It can be seen that the present invention improves the calculation accuracy and robustness of the wake-up word energy under background noise conditions.

[0050] Of course, it is not necessary to achieve all of the advantages described above simultaneously in order to implement any product or method of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0052] Figure 1 A flowchart of a method for calculating wake-up word energy provided by an embodiment of the present invention;

[0053] Figure 2 This is a block diagram of a wake-up word energy calculation system provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0054] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0055] The embodiment of the present invention provides a method for calculating the energy of a wake-up word. Figure 1 As shown, the method includes:

[0056] S101: Obtain a wake-up word audio signal.

[0057] Optionally, in an optional embodiment of the present invention, the device for acquiring the wake-up word audio signal may be a sound collection device deployed on a smart home electronic device.

[0058] Optionally, in another optional embodiment of the present invention, the above-mentioned wake-up word audio signal may be a voice signal containing the wake-up keyword and an audio signal of a scene noise signal of the scene in which the distributed voice wake-up system is located.

[0059] S102: Perform a first conversion on the wake-up word audio signal to obtain a short-time energy spectrum of the wake-up word audio.

[0060] Optionally, in an optional embodiment of the present invention, the above-mentioned first conversion may include short-time Fourier transform (STFT), modulo operation and square operation. The above-mentioned process of performing the first conversion on the wake-up word audio signal to obtain the short-time energy spectrum of the wake-up word audio may be: performing STFT transformation on the wake-up word audio signal to obtain the short-time spectrum of the wake-up word audio signal, and then performing modulo operation and square operation on the short-time spectrum to obtain the short-time energy spectrum of the wake-up word audio signal.

[0061] Wherein, the above-mentioned STFT transform is applicable to the spectrum analysis of slowly time-varying signals. Its method is to first divide the voice signal into frames, and then perform Fourier transform on each frame signal. In this way, each frame voice signal can be considered to be intercepted from each different stationary signal waveform, and the short-time spectrum of each frame voice is an approximation of the spectrum of each stationary signal waveform. Then, the short-time spectrum is subjected to modulo operation and square operation to obtain a short-time energy spectrum that characterizes the frequency distribution of the voice signal. The above-mentioned STFT, modulo operation and square operation are common means for speech preprocessing, and the present invention will not go into details here.

[0062] S103: Taking the logarithm of the short-time energy spectrum to obtain the logarithmic spectrum of the wake-up word audio.

[0063] Optionally, in an optional embodiment of the present invention, the wake-up word audio signal can be converted from time domain data to logarithmic spectrum features by taking the logarithm of the short-time energy spectrum, compressing the dynamic range of the wake-up word feature data in the wake-up word audio signal, thereby ensuring the integrity of the logarithmic spectrum data of the wake-up word audio used for subsequent neural network model calculations, thereby improving the accuracy of the final calculation of the wake-up word energy.

[0064] S104: Input the logarithmic spectrum into a preset neural network model, so that the preset neural network model generates a predicted probability matrix according to the logarithmic spectrum.

[0065] Optionally, in an optional embodiment of the present invention, the preset neural network model may be a convolutional neural network (CNN). The present invention calculates the probability value of the time-frequency point in the logarithmic spectrum of the input wake-up word audio belonging to the wake-up data by modeling the scene noise and the wake-up word audio classification network based on CNN, and maps it into a probability matrix.

[0066] Since the prior art obtains the threshold value for distinguishing scene noise and wake-up audio by calculating the energy of multiple frames of data, it actually assumes that the scene noise is stationary noise and is much smaller than the wake-up word energy. However, in actual application scenarios, the above assumptions are difficult to meet, resulting in a serious inaccuracy in the final calculated wake-up word energy. At the same time, since the coefficients and parameters used to calculate the threshold value in the prior art are usually obtained through a pre-set static scene, and then no longer updated according to the actual application scenario in subsequent actual deployment, its universality deteriorates, further resulting in an inaccurate final calculated wake-up word energy. Therefore, the present invention introduces CNN to estimate the wake-up word speech component in the wake-up word audio, so that compared with the prior art, the present invention achieves adaptation to different application scenarios, and at the same time achieves dynamic adjustment of internal parameters for different application scenarios, thereby improving the accuracy of the final calculated wake-up word energy.

[0067] S105. Binarize the predicted probability matrix to obtain a binary matrix.

[0068] Optionally, in an optional embodiment of the present invention, since the dimension of the predicted probability matrix generated in the above step S104 is different from the dimension of the short-time energy spectrum of the wake-up word audio generated in the above step S102, it is impossible to obtain a scalar for calculating the wake-up word energy. Therefore, through a matrix binarization operation, the above-mentioned predicted probability matrix is ​​converted into a binary matrix, and the binary matrix has the same dimension as the short-time energy spectrum of the above-mentioned wake-up word audio. At the same time, according to the preset threshold value in the above-mentioned preset neural network model, when performing the above-mentioned binarization operation, the elements in the above-mentioned predicted probability matrix that are greater than the preset threshold value can be set to 1, and the elements that are not greater than the preset threshold value can be set to 0. The interference data in the data used to calculate the wake-up word speech energy is reduced, and the accuracy of the final calculation of the wake-up word speech energy is improved.

[0069] S106: Perform a second conversion on the short-time energy spectrum and the binary matrix to determine the wake-up word voice energy of the wake-up word audio signal.

[0070] Optionally, in an optional embodiment of the present invention, the above-mentioned second transformation includes but is not limited to: matrix Hadamard product and matrix dimension summation. The matrix Hadamard product is a commonly used matrix multiplication operation. After performing a matrix Hadamard product operation on the above-mentioned short-time energy spectrum and the binary matrix, a two-dimensional matrix is ​​obtained. The two dimensions of the two-dimensional matrix are then summed to obtain the wake-up word speech energy. The present invention can realize the selection of time-frequency points belonging to wake-up data in the short-time energy spectrum of the above-mentioned wake-up word audio signal through the above-mentioned matrix Hadamard product operation, thereby improving the accuracy of the final calculation of the wake-up word speech energy.

[0071] The present invention estimates the wake-up word speech component in the wake-up word audio by introducing a preset neural network model, so that compared with the prior art, the present invention improves the accuracy of distinguishing between noise time-frequency points and wake-up word time-frequency points in the wake-up word audio under different application scenarios, thereby improving the robustness and accuracy of the final calculation of the wake-up word energy under background noise conditions. At the same time, the present invention dynamically updates the internal parameters involved in the calculation of the wake-up word energy for different application scenarios by introducing a preset neural network model, thereby improving the applicability of the present invention to different application scenarios. Finally, since the present invention can be deployed based on the existing distributed voice wake-up system, there is no need to modify the hardware equipment, which further improves the universality of the present invention. It can be seen that the present invention improves the calculation accuracy and robustness of the wake-up word energy under background noise conditions.

[0072] Optionally, preset the neural network model training process, including:

[0073] The log spectrum of the noisy speech data is input into the initial neural network for processing to obtain the predicted training probability matrix.

[0074] The error between the training probability matrix and the label matrix is ​​calculated based on the cross entropy loss function.

[0075] According to the error value, the initial neural network is iteratively updated using a preset optimization algorithm until the training cutoff condition is met to obtain a preset neural network model.

[0076] Optionally, in an optional embodiment of the present invention, during the training of the preset neural network model, the noisy speech data may be synthesized. The noisy training data is inserted into the wake-up word speech training data according to a preset signal-to-noise ratio to obtain the noisy speech data. The training data may be recorded wake-up word speech without background noise.

[0077] Optionally, in another optional embodiment of the present invention, the acquisition process of the above-mentioned label matrix can be obtained by performing a first conversion, taking the logarithm and binarization on the training data used to train the above-mentioned initial neural network. After taking the logarithm to obtain the logarithmic spectrum of the training data, the above-mentioned preset threshold value is selected according to the background noise level to which the noise training data belongs, and the time-frequency points in the above-mentioned logarithmic spectrum that are greater than the preset threshold value are set to 1, and the time-frequency points that are less than the preset threshold value are set to zero. Since the label matrix is ​​generated by the above-mentioned training data through conversion, the time-frequency points in the label matrix all belong to the time-frequency points of the wake-up data. The present invention uses the label matrix and combines the cross-entropy loss function and the optimization algorithm based on adaptive moment estimation (Adaptive Moment Estimation, Adam) to iterate the above-mentioned initial neural network more finely, thereby improving the accuracy of the preset neural network in screening and judging the wake-up data in the wake-up word audio.

[0078] Optionally, in another optional embodiment of the present invention, the above-mentioned training cutoff condition may be that the loss of the above-mentioned initial neural network on the validation set does not decrease within a preset period.

[0079] Optionally, before training the initial neural network, the method further includes: training data processing and / or training data feature extraction.

[0080] Among them, training data processing includes:

[0081] The training data of the wake-up word speech is first converted to obtain the short-time energy spectrum of the training data; the short-time energy spectrum of the training data is logarithmized to obtain the logarithmic spectrum of the training data; the logarithmic spectrum of the training data is binarized to obtain a label matrix.

[0082] Among them, training data feature extraction includes:

[0083] The noise training data is inserted into the training data according to the signal-to-noise ratio to obtain noisy speech data; the noisy speech data is subjected to a first conversion to obtain a short-time energy spectrum of the noisy speech data; and the short-time energy spectrum of the noisy speech data is logarithmized to obtain a logarithmic spectrum of the noisy speech data.

[0084] Optionally, the log spectrum is input into a preset neural network model so that the preset neural network model generates a predicted probability matrix based on the log spectrum, including:

[0085] The preset neural network maps the time-frequency points of the received logarithmic spectrum into a predicted probability matrix, and each element in the predicted probability matrix represents the probability value of the time-frequency point corresponding to the element belonging to the wake-up data.

[0086] Optionally, the predicted probability matrix is ​​binarized to obtain a binary matrix, including:

[0087] According to the preset threshold value in the preset neural network model, the predicted probability matrix is ​​binarized to obtain a binary matrix, wherein the binarization is to determine whether each element in the predicted probability matrix is ​​greater than the preset threshold value. If so, the element is set to 1; if the element is not greater than the preset threshold value, the element is set to 0.

[0088] Optionally, the above method is applied to a distributed voice wake-up system, and the method further includes:

[0089] Multiple electronic devices in a distributed voice wake-up system calculate the wake-up word voice energy of each device according to the method, and compare the wake-up word voice energy of each device with the wake-up word voice energy of other devices. The device with the largest wake-up word voice energy performs the wake-up operation, and other devices except the device that performs the wake-up operation do not perform the wake-up operation.

[0090] Optionally, in an optional embodiment of the present invention, each electronic device in the distributed voice wake-up system is configured to execute the above-mentioned wake-up word energy calculation method. The above-mentioned electronic devices include but are not limited to smart home electronic devices and smart communication devices.

[0091] Optionally, in another optional embodiment of the present invention, whether the device performs the wake-up operation is determined based on the calculated wake-up word speech energy of the device and the wake-up word speech energy calculated by other devices in the distributed voice wake-up system:

[0092]

[0093] Among them, E i is the wake-up word speech energy of the i device, maxE j The maximum wake-up word speech energy among the wake-up word speech energies calculated for j devices except device i. i >maxE j In the case of a lag, if the wake-up word speech energy of the device is determined to be the highest, the wake-up operation will be performed by the device itself, and the wake-up operation will not be performed on other devices. This method is deployed in each electronic device of the distributed voice wake-up system, thereby improving the robustness of the distributed voice wake-up system to background noise.

[0094] Optionally, the wake-up word audio signal is an audio signal including a voice signal containing the wake-up keyword and a scene noise signal of the scene in which the distributed voice wake-up system is located.

[0095] The present invention estimates the wake-up word speech component in the wake-up word audio by introducing a preset neural network model, so that compared with the prior art, the present invention improves the accuracy of distinguishing between noise time-frequency points and wake-up word time-frequency points in the wake-up word audio under different application scenarios, thereby improving the robustness and accuracy of the final calculation of the wake-up word energy under background noise conditions. At the same time, the present invention dynamically updates the internal parameters involved in the calculation of the wake-up word energy for different application scenarios by introducing a preset neural network model, thereby improving the applicability of the present invention to different application scenarios. Finally, since the present invention can be deployed based on the existing distributed voice wake-up system, there is no need to modify the hardware equipment, which further improves the universality of the present invention. It can be seen that the present invention improves the calculation accuracy and robustness of the wake-up word energy under background noise conditions.

[0096] Corresponding to the above-mentioned wake-up word energy calculation method embodiment, the present invention also provides a wake-up word energy calculation system, which is applied to a distributed voice wake-up system, such as Figure 2 As shown, the system includes:

[0097] Signal acquisition module 201: used to acquire the wake-up word audio signal.

[0098] The first conversion module 202 is configured to perform a first conversion on the wake-up word audio signal to obtain a short-time energy spectrum of the wake-up word audio.

[0099] The second conversion module 203 is configured to take the logarithm of the short-time energy spectrum to obtain the logarithmic spectrum of the wake-up word audio.

[0100] Matrix generation module 204: used to input the log spectrum into the preset neural network model, so that the preset neural network model generates a predicted probability matrix according to the log spectrum.

[0101] The third conversion module 205 is used to binarize the predicted probability matrix to obtain a binary matrix.

[0102] The fourth conversion module 206 is configured to perform a second conversion on the short-time energy spectrum and the binary matrix to determine the wake-up word speech energy of the wake-up word audio signal.

[0103] Optionally, the above system further includes:

[0104] The model training module processes the log spectrum of noisy speech data into the initial neural network to obtain a predicted training probability matrix. The error between the training probability matrix and the label matrix is ​​calculated using a cross-entropy loss function. Based on this error, the initial neural network is iteratively updated using a preset optimization algorithm until the training cutoff condition is met, resulting in the preset neural network model.

[0105] Optionally, the above system further includes:

[0106] The training data processing module is used to perform a first conversion on the training data of the wake-up word speech to obtain the short-time energy spectrum of the training data; take the logarithm of the short-time energy spectrum of the training data to obtain the logarithmic spectrum of the training data; and binarize the logarithmic spectrum of the training data to obtain a label matrix.

[0107] And / or, a training data feature extraction module is configured to insert the noise training data into the training data according to the signal-to-noise ratio to obtain noisy speech data, perform a first conversion on the noisy speech data to obtain a short-time energy spectrum of the noisy speech data, and take the logarithm of the short-time energy spectrum of the noisy speech data to obtain a logarithmic spectrum of the noisy speech data.

[0108] Optionally, the matrix generation module 204 is configured to:

[0109] The preset neural network in the matrix generation module 204 maps the received time-frequency points of the log spectrum into a predicted probability matrix, where each element in the predicted probability matrix represents the probability value that the time-frequency point corresponding to the element belongs to the wake-up data.

[0110] Optionally, the third conversion module 205 is configured to:

[0111] According to the preset threshold value in the preset neural network model, the predicted probability matrix is ​​binarized to obtain a binary matrix, wherein the binarization is to determine whether each element in the predicted probability matrix is ​​greater than the preset threshold value. If so, the element is set to 1; if the element is not greater than the preset threshold value, the element is set to 0.

[0112] Optionally, the above system further includes:

[0113] The device wake-up module is used to control multiple electronic devices in the distributed voice wake-up system. The wake-up word voice energy of each device is calculated according to the method, and the wake-up word voice energy of the device is compared with the wake-up word voice energy of other devices. The device with the largest wake-up word voice energy performs the wake-up operation, and other devices except the device that performs the wake-up operation do not perform the wake-up operation.

[0114] Optionally, the wake-up word audio signal is an audio signal containing a voice signal of the wake-up keyword and a scene noise signal of the scene in which the distributed voice wake-up system is located.

[0115] An embodiment of the present invention provides a voice wake-up system, the system comprising:

[0116] Multiple electronic devices are configured to execute instructions to implement any of the above wake-up word energy calculation methods.

[0117] An embodiment of the present invention provides a computer-readable storage medium. When instructions in the computer-readable storage medium are executed by a processor of an electronic device, the device can execute any of the above-mentioned wake-up word energy calculation methods.

[0118] Memory may include non-permanent memory in a computer-readable medium, random access memory (RAM), and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory includes at least one memory chip. Memory is an example of a computer-readable medium.

[0119] Computer-readable media include permanent and non-permanent, removable and non-removable media that can be used to store information by any method or technology. Information can be computer-readable instructions, data structures, program modules or other data. Examples of computer storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices or any other non-transmission media that can be used to store information that can be accessed by a computing device. As defined herein, computer-readable media does not include temporary computer-readable media (transitory media), such as modulated data signals and carrier waves.

[0120] Those skilled in the art will appreciate that the embodiments of the present application may be provided as methods, systems, or computer program products. Therefore, the present application may take the form of a complete hardware embodiment, a complete software embodiment, or an embodiment combining software and hardware. Furthermore, the present application may take the form of a computer program product implemented on one or more computer-usable storage media (including but not limited to magnetic disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0121] It should be noted that, in this article, relational terms such as first and second, etc. are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. It should also be noted that the terms "comprise", "include" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, commodity or device comprising a series of elements includes not only those elements, but also other elements not explicitly listed, or also includes elements inherent to such process, method, commodity or device. In the absence of further restrictions, an element defined by the sentence "comprises a ..." does not exclude the presence of other identical elements in the process, method, commodity or device comprising the element.

[0122] Each embodiment in this specification is described in a related manner. Similar parts between the various embodiments can be referred to in conjunction with each other. Each embodiment focuses on the differences from the other embodiments. In particular, the system embodiment is generally similar to the method embodiment, so the description is relatively simple. For related parts, refer to the description of the method embodiment.

[0123] The above are merely embodiments of the present application and are not intended to limit the present application. For those skilled in the art, the present application may have various changes and variations. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present application should all be included within the scope of the claims of the present application.

Claims

1. A method for calculating wake-up word energy, characterized in that: The method is applied to a distributed voice wake-up system, and the method includes: Get the wake-up word audio signal; Performing a first conversion on the wake-up word audio signal to obtain a short-time energy spectrum of the wake-up word audio; Taking the logarithm of the short-time energy spectrum to obtain the logarithmic spectrum of the wake-up word audio; Inputting the logarithmic spectrum into a preset neural network model so that the preset neural network model generates a predicted probability matrix according to the logarithmic spectrum; Binarizing the predicted probability matrix according to a preset threshold value in a preset neural network model to obtain a binary matrix to reduce interference data in data used to calculate the wake-up word speech energy; Performing a second conversion on the short-time energy spectrum and the binary matrix to determine the wake-up word voice energy of the wake-up word audio signal; Among them, multiple electronic devices in the distributed voice wake-up system respectively calculate the wake-up word voice energy of the device according to the method, and compare the wake-up word voice energy of the device with the wake-up word voice energy of other devices. The device with the largest wake-up word voice energy performs the wake-up operation, and other devices except the device that performs the wake-up operation do not perform the wake-up operation.

2. The method according to claim 1, characterized in that The training process of the preset neural network model includes: The log spectrum of the noisy speech data is input into the initial neural network for processing to obtain the predicted training probability matrix; Calculate the error between the training probability matrix and the label matrix based on the cross entropy loss function; According to the error value, the initial neural network is iteratively updated using a preset optimization algorithm until a training cutoff condition is met, thereby obtaining the preset neural network model.

3. The method according to claim 2, characterized in that Before training the initial neural network, the method further includes: processing training data and / or extracting training data features. The training data processing includes: Performing the first conversion on the training data of the wake-up word speech to obtain a short-time energy spectrum of the training data; performing the logarithmization on the short-time energy spectrum of the training data to obtain a logarithmic spectrum of the training data; performing the binarization on the logarithmic spectrum of the training data to obtain the label matrix; The training data feature extraction includes: The noise training data is inserted into the training data according to the signal-to-noise ratio to obtain the noisy speech data; the first conversion is performed on the noisy speech data to obtain the short-time energy spectrum of the noisy speech data; and the logarithm of the short-time energy spectrum of the noisy speech data is taken to obtain the logarithmic spectrum of the noisy speech data.

4. The method according to claim 1, wherein Inputting the logarithmic spectrum into a preset neural network model so that the preset neural network model generates a predicted probability matrix according to the logarithmic spectrum includes: The preset neural network maps the received time-frequency points of the logarithmic spectrum into the predicted probability matrix, and each element in the predicted probability matrix represents the probability value that the time-frequency point corresponding to the element belongs to the wake-up data.

5. The method according to claim 4, characterized in that Binarizing the predicted probability matrix according to a preset threshold value in a preset neural network model to obtain a binary matrix includes: According to the preset threshold value in the preset neural network model, the predicted probability matrix is ​​binarized to obtain the binary matrix, wherein the binarization is to determine whether each element in the predicted probability matrix is ​​greater than the preset threshold value. If so, the element is set to 1; if the element is not greater than the preset threshold value, the element is set to 0.

6. The method according to claim 1, characterized in that The wake-up word audio signal is a voice signal containing the wake-up keyword and an audio signal of a scene noise signal of the scene in which the distributed voice wake-up system is located.

7. A wake-up word energy calculation system, characterized in that: The system is applied to a distributed voice wake-up system, and the system includes: Signal acquisition module: used to obtain the wake-up word audio signal; A first conversion module is configured to perform a first conversion on the wake-up word audio signal to obtain a short-time energy spectrum of the wake-up word audio; A second conversion module is used to take the logarithm of the short-time energy spectrum to obtain a logarithmic spectrum of the wake-up word audio; A matrix generation module is used to input the logarithmic spectrum into a preset neural network model so that the preset neural network model generates a predicted probability matrix based on the logarithmic spectrum; A third conversion module is configured to binarize the predicted probability matrix according to a preset threshold value in a preset neural network model to obtain a binary matrix, so as to reduce interference data in the data used to calculate the wake-up word speech energy; A fourth conversion module: configured to perform a second conversion on the short-time energy spectrum and the binary matrix to determine the wake-up word voice energy of the wake-up word audio signal; Among them, multiple electronic devices in the distributed voice wake-up system respectively calculate the wake-up word voice energy of the device according to the system, and compare the wake-up word voice energy of the device with the wake-up word voice energy of other devices. The device with the largest wake-up word voice energy performs the wake-up operation, and other devices except the device that performs the wake-up operation do not perform the wake-up operation.

8. A voice wake-up system, characterized in that: The system comprises: Multiple electronic devices are configured to execute instructions to implement the wake-up word energy calculation method as described in any one of claims 1 to 6.

9. A computer-readable storage medium, characterized in that When the instructions in the computer-readable storage medium are executed by a processor of an electronic device, the device is enabled to execute the wake-up word energy calculation method according to any one of claims 1 to 6.

Images