A distributed wake-up method and system

Abstract

The present invention provides a distributed wake-up method and system, which are applied to a first voice device, including: determining a first audio detection signal according to a user's wake-up voice, and receiving an estimated reference signal sent by at least one second voice device; determining noise estimation energy according to at least one estimated reference signal and the first audio detection signal; determining the received wake-up voice energy according to the noise estimation energy, and determining the wake-up state based on the wake-up voice energy. The distributed wake-up method and system provided by the present invention can effectively suppress the influence of the self-noise of the first voice device on the calculation of the distributed wake-up distance estimation value by the artificial intelligence voice device by introducing the estimated reference signal of the second voice device, and greatly improve the accuracy of the distributed wake-up sound source distance calculation in the artificial intelligence voice device self-noise scenario while reducing the amount of calculation, thereby improving the performance of distributed wake-up and the breadth of practical application.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a distributed wake-up method and system. Background Art

[0002] Distributed wake-up is aimed at the problem that multiple artificial intelligence (AI) voice devices are currently deployed in a local space at the same time, which easily causes the same voice command to operate on multiple AI voice devices at the same time. Especially in the environment of distributed wake-up AI voice devices, when the AI ​​voice device is awakened by the user's wake-up voice, if multiple AI voice devices respond at the same time, the same user wake-up voice in one voice command will wake up multiple AI voice devices, making it impossible for users to achieve their actual operation purpose.

[0003] At present, in order to solve the problem of multiple AI voice devices being awakened at the same time, a distributed wake-up solution has been introduced. The common distributed wake-up solution is to calculate the wake-up voice energy received by each AI voice device during wake-up, and each AI voice device automatically compares the energy between each other. The greater the energy, the closer the AI ​​voice device is to the sound source of the user's wake-up voice, and thus it should be awakened first. However, in the scenario of distributed wake-up of AI voice devices, many AI voice devices will generate self-noise during operation. Part of the self-noise is caused by the motor and mechanical friction of the AI ​​voice device. There is no way to directly obtain the reference signal of these noises at the AI ​​voice device itself.

[0004] This simple distributed wake-up calculation based on energy judgment estimates the distance directly according to the voice energy of the sound signal received by each AI voice device. This may cause an AI voice device that receives a lot of self-noise but is actually far away from the sound source to be mistakenly judged as an AI voice device close to the sound source, resulting in a huge distributed wake-up error. Summary of the invention

[0005] In view of the problems existing in the prior art, an embodiment of the present invention provides a distributed wake-up method and system.

[0006] In a first aspect, the present invention provides a distributed wake-up method, which is applied to a first voice device, and the method includes: determining a first audio detection signal according to a user wake-up voice, and receiving an estimated reference signal sent by at least one second voice device, wherein the estimated reference signal is determined according to a second audio detection signal received by each second voice device, and the second audio detection signal is determined by each second voice device according to the user wake-up voice; the first audio detection signal and the second audio detection signal both include noise generated when the first voice device is running; determining noise estimation energy according to at least one estimated reference signal and the first audio detection signal; determining the received wake-up voice energy according to the noise estimation energy, and determining the wake-up state based on the wake-up voice energy.

[0007] In a second aspect, the present invention provides a distributed wake-up method, which is applied to a second voice device, and the method includes: determining a second audio detection signal based on a user wake-up voice; determining an estimated reference signal based on the second audio detection signal; sending the estimated reference signal to a first voice device, so that the first voice device can determine the received wake-up voice energy based on the estimated reference signal and the first audio detection signal; the first audio detection signal is determined by the first voice device based on the user wake-up voice; the first audio detection signal and the second audio detection signal include noise generated when the first voice device is running; the wake-up voice energy is determined by the first voice device based on the noise estimated energy, and is used to determine the wake-up state of the first voice device; the noise estimated energy is determined by the first voice device based on the estimated reference signal and the first audio detection signal.

[0008] In a third aspect, the present invention provides a distributed wake-up method, which is applied to a cloud server, and the method includes: receiving a first audio detection signal sent by a first voice device, and simultaneously receiving a second audio detection signal sent by at least one second voice device; the first audio detection signal is determined by the first voice device according to the user's wake-up voice; the second audio detection signal is determined by each second voice device according to the user's wake-up voice; the first audio detection signal and the second audio detection signal both include noise generated when the first voice device is running; determining an estimated reference signal according to each second audio detection signal; determining noise estimated energy according to at least one estimated reference signal and the first audio detection signal; determining the wake-up voice energy received by the first voice device according to the noise estimated energy; sending the wake-up voice energy to the first voice device; the wake-up voice energy is used to determine the wake-up state of the first voice device.

[0009] In a fourth aspect, the present invention provides a first voice device, which is a first voice device for executing a distributed wake-up method, specifically comprising: a first module, for determining a first audio detection signal based on a user wake-up voice, and receiving an estimated reference signal sent by at least one second voice device, wherein the estimated reference signal is determined based on a second audio detection signal received by each second voice device, and the second audio detection signal is determined by each second voice device based on the user wake-up voice; the first audio detection signal and the second audio detection signal include noise generated when the first voice device is running; a second module, for determining a noise estimation energy based on at least one estimated reference signal and the first audio detection signal; and a third module, for determining the received wake-up voice energy based on the noise estimation energy, and determining a wake-up state based on the wake-up voice energy.

[0010] In a fifth aspect, the present invention provides a distributed wake-up system, comprising: a first voice device and at least one second voice device; the second voice device is used to determine a second audio detection signal according to a user wake-up voice; determine an estimated reference signal according to the second audio detection signal; and send the estimated reference signal to the first voice device; the first voice device is used to determine a first audio detection signal according to the user wake-up voice, and receive the estimated reference signal, the first audio detection signal and the second audio detection signal including noise generated when the first voice device is running; determine noise estimation energy according to at least one estimated reference signal and the first audio detection signal; determine the received wake-up voice energy according to the noise estimation energy, and determine the wake-up state of the first voice device based on the wake-up voice energy.

[0011] The distributed wake-up method and system provided by the present invention can effectively suppress the influence of the self-noise of the first voice device on the calculation of the distributed wake-up distance estimation value by the artificial intelligence voice device by introducing the estimated reference signal of the second voice device. While reducing the amount of calculation, the accuracy of the distributed wake-up sound source distance calculation is greatly improved in the self-noise scenario of the artificial intelligence voice device, thereby improving the performance of distributed wake-up and the breadth of practical application. BRIEF DESCRIPTION OF THE DRAWINGS

[0012] In order to more clearly illustrate the technical solutions in the present invention or 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 some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0013] Figure 1 This is one of the flow charts of the distributed wake-up method provided by an embodiment of the present invention;

[0014] Figure 2 It is a schematic diagram of a distributed artificial intelligence voice device wake-up scenario provided by an embodiment of the present invention;

[0015] Figure 3 This is a second flow chart of the distributed wake-up method provided by an embodiment of the present invention;

[0016] Figure 4 This is a third flow chart of the distributed wake-up method provided by an embodiment of the present invention;

[0017] Figure 5 is a schematic diagram of the structure of a first voice device provided by an embodiment of the present invention;

[0018] Figure 6 is a schematic diagram of the structure of a distributed wake-up system provided by an embodiment of the present invention;

[0019] Figure 7 It is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0020] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the drawings of the present invention. Obviously, the described embodiments are 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 creative work are within the scope of protection of the present invention.

[0021] It should be noted that, in the description of the embodiments of the present invention, the terms "include", "comprise" or any other variants thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "include one..." do not exclude the existence of other identical elements in the process, method, article or device including the elements. For those of ordinary skill in the art, the specific meanings of the above terms in the present invention can be understood according to the specific circumstances.

[0022] The first voice device and the second voice device are both AI voice devices, which are, for example, electronic devices having a voice processing module and having functions such as voice recognition, voiceprint recognition, or semantic understanding. The AI ​​voice device distributed wake-up scenario may refer to a scenario in which multiple AI voice devices are deployed simultaneously in a local space. The first voice device generates a large amount of noise during operation, while the second voice device does not generate noise during operation.

[0023] In the AI ​​voice device distributed wake-up scenario, there is a first voice device and at least one second voice device. Since the noise generated by the first voice device interferes with itself, it can be called self-noise, and the noise generated is transmitted through the air to interfere with other AI voice devices, such as the second voice device.

[0024] The self-noise generated by some sound-source first voice devices such as smart speakers and phonographs can be divided into two categories.

[0025] One type of self-noise is the sound of audio played by the first voice device itself, such as music played by an audio device. For this type of self-noise, the first voice device can obtain a reference signal and perform self-noise processing.

[0026] The other type is the self-noise generated by the motor and mechanical friction of the first voice device. For example, when a smart washing machine is running, it will produce a vibration sound. For this kind of self-noise, the first voice device has no reference signal, so it cannot perform self-noise processing.

[0027] Since most of the sound received by the running first voice device is its own self-noise, the distance to the sound source of the user's wake-up voice can only be accurately estimated after the self-noise part is removed. Therefore, the embodiment of the present invention needs to estimate the distributed wake-up distance of the AI ​​voice device under the premise of removing the influence of the self-noise of the AI ​​voice device. The distributed wake-up distance is the distance between the AI ​​voice device and the sound source of the user's wake-up voice.

[0028] In the currently known distributed wake-up methods, since the self-noise impact of distributed AI voice devices is not taken into account, it is still difficult to robustly handle the impact of self-noise when estimating the distance from the AI ​​voice device to the sound source of the user's wake-up voice. This greatly limits the practical application value of distributed wake-up in AI voice device scenarios.

[0029] For distributed wake-up under self-noise conditions of AI voice devices, most of the existing technologies perform reverberation estimation, dereverberation, and noise reduction on the signals received by the AI ​​voice devices. These processes have a high amount of computation and are generally limited and cannot be implemented in practical applications.

[0030] The present invention does not have these complex and high-computation calculations, does not affect the characteristics of obtaining the distance information of the signal, and is robust to environmental influences. Therefore, it can greatly improve the application scope of the distributed wake-up method and improve the performance of distributed wake-up under the self-noise conditions of AI voice devices.

[0031] Combine the following Figures 1 to 7 The distributed wake-up method and system provided by the embodiments of the present invention are described.

[0032] Figure 1 is one of the flow charts of the distributed wake-up method provided in an embodiment of the present invention, which is applied to a first voice device, such as Figure 1 As shown, including but not limited to the following steps:

[0033] First, in step S11, a first audio detection signal is determined according to the user wake-up voice, and an estimated reference signal sent by at least one second voice device is received, wherein the estimated reference signal is determined according to the second audio detection signal received by each second voice device, and the second audio detection signal is determined by each second voice device according to the user wake-up voice; the first audio detection signal and the second audio detection signal both include the noise generated when the first voice device is running.

[0034] In the AI ​​voice device distributed wake-up scenario, the distributed interconnection between AI voice devices can be used to transmit information about received noise to each other. Each AI voice device can communicate with each other and obtain the status of all AI voice devices in the current scene. The first voice device and the second voice device are, for example, AI voice devices. The second voice device can distinguish the first voice device as the source of self-noise in the observed sound, thereby effectively helping the running first voice device to remove the influence of self-noise on its own judgment of the distributed wake-up distance.

[0035] Among them, the user's wake-up voice can be used to wake up the AI ​​voice devices in the distributed AI voice device wake-up scenario.

[0036] The estimated reference signal is obtained according to the second audio detection signal when each second audio device observes the self-noise emitted by the first audio device.

[0037] Figure 2 is a schematic diagram of a distributed artificial intelligence voice device wake-up scenario provided by an embodiment of the present invention, such as Figure 2 As shown in the figure, under the self-noise condition of the AI ​​voice device, for the sound signal received by the first voice device (i.e., the running device), in addition to the user wake-up voice issued by the user, there is also the noise generated by the first voice device (i.e., the running device) itself during operation, i.e., self-noise. This self-noise is the noise from the first voice device (i.e., the running device) to other AI voice devices (i.e., device A and device B). Therefore, in order to accurately determine the relative distance between the sound source of the user wake-up voice and the AI ​​voice device, it is necessary to remove the self-noise part during distance determination in order to effectively improve the robustness of the distributed wake-up distance. However, the traditional distributed wake-up algorithm based on energy estimation does not consider the impact of self-noise, especially the problem of self-noise suppression during device operation.

[0038] Specifically, for any first voice device, the collected first time domain signal is converted into a frequency domain signal as a first audio detection signal, and an estimated reference signal sent by at least one second voice device is received.

[0039] The first voice device obtains multiple first time domain signals in multiple preset time periods, and performs fast Fourier transform on each first time domain signal to convert it into multiple frequency domain signals in the user wake-up speech frequency band as the first audio detection signal, each first audio detection signal is in a different frequency band, and each frequency band is a sub-band of the user wake-up speech frequency band. The first time domain signal is collected by the first voice device through self-observation of the self-noise emitted by itself.

[0040] The user wake-up speech frequency band mainly refers to the audio spectrum range of the user wake-up speech. Since the self-noise in the user wake-up speech frequency band will interfere with the AI ​​voice device, it is necessary to remove the self-noise energy from the wake-up speech energy received by the AI ​​voice device.

[0041] The estimated reference signal may be sent directly from the second voice device to the first voice device, or may be sent from the second voice device to a cloud server, which is then forwarded to the first voice device by the cloud server.

[0042] Further, in step S12, noise estimation energy is determined according to at least one estimation reference signal and the first audio detection signal.

[0043] Specifically, the first voice device can update the first filter coefficient of the first voice device according to all the estimated reference signals and the first audio detection signal to achieve a better noise reduction effect.

[0044] The first speech device may also determine noise estimation energy based on the first filter coefficient and the first audio detection signal.

[0045] Further, in step S13, the received wake-up speech energy is determined according to the noise estimation energy, and the wake-up state is determined based on the wake-up speech energy.

[0046] Specifically, the first voice device determines the distributed wake-up distance of the first voice device according to the noise estimation energy.

[0047] Each second voice device obtains multiple second time domain signals in multiple preset time periods, and performs Fast Fourier Transform (FFT) on each second time domain signal to convert it into multiple second audio detection signals, each of which is in a different sub-band in the user wake-up voice band. The larger the number of preset time periods, the larger the number of second audio detection signals, the larger the number of corresponding sub-bands, and the higher the noise reduction accuracy.

[0048] Each second voice device updates its own second filter coefficient according to the second audio detection signal, and then obtains an estimated reference signal according to the second audio detection signal and the second filter coefficient.

[0049] The noise estimation energy is the sound energy in the user wake-up speech frequency band received by the first voice device, and part of the self-noise energy is subtracted by updating the first filter coefficient, so it is closer to the actual energy of the user wake-up speech received by the first voice device.

[0050] In addition, the distributed wake-up distances of each AI voice device can be compared, and the minimum value can be determined among the distributed wake-up distances of all AI voice devices, so that the AI ​​voice device corresponding to the minimum value is awakened by the user's wake-up voice, and other AI voice devices in the distributed AI voice device wake-up scenario are not awakened.

[0051] The distributed wake-up method provided in the embodiment of the present invention can effectively suppress the influence of the self-noise of the first voice device on the calculation of the distributed wake-up distance estimation value by the artificial intelligence voice device by introducing the estimated reference signal of the second voice device. While reducing the amount of calculation, it greatly improves the accuracy of the distributed wake-up sound source distance calculation in the artificial intelligence voice device self-noise scenario, thereby improving the performance of distributed wake-up and the breadth of practical application.

[0052] Optionally, determining the first audio detection signal according to the user wake-up voice specifically includes:

[0053] Acquire a first time domain signal according to the user wake-up voice;

[0054] The first time domain signal is converted into a frequency domain signal to obtain the first audio detection signal.

[0055] The first time domain signal may be collected in multiple preset time periods within the length (time) of the user's wake-up voice, and the preset time periods may be continuous, and the preset time periods may be equal or unequal.

[0056] In each preset time period, in the user wake-up voice frequency band, the first voice device uses each sound collection channel to collect the first time domain signal. Among them, the first voice device uses a microphone array, so the number of sound collection channels is determined by the microphone array of the device. For example, if there are 4 microphones, there will be 4 sound collection channels.

[0057] Specifically, the first voice device obtains multiple first time domain signals in multiple preset time periods, and performs fast Fourier transform on each first time domain signal to convert it into multiple frequency domain signals, each frequency domain signal is used as a first audio detection signal, and each first audio detection signal is in a different frequency band. The larger the number of preset time periods, the larger the number of frequency bands corresponding to the first audio detection signal, and thus the higher the noise reduction accuracy.

[0058] For a first voice device numbered s, a first audio detection signal x in a target frequency band f is determined from among a plurality of first audio detection signals. s，f (t), the first audio detection signal x s，f The calculation formula of (t) is as follows:

[0059] x s,f (t) = [X(1) s,f (t),X(2) s,f (t),...,X(M) s,f (t)] k ;

[0060] Among them, M is the number of sound collection channels of the first voice device numbered s; f is the target frequency band, and after the first time domain signal is subjected to fast Fourier transform, f represents the frequency point; t is the order of the preset time periods; k is the number of the preset time periods, t=0, 1,…, k-1; s is the number of the first voice device.

[0061] Optionally, determining the noise estimation energy according to at least one estimation reference signal and the first audio detection signal includes:

[0062] updating a first filter coefficient of the first voice device according to the estimated reference signal of each second voice device and the first audio detection signal;

[0063] The noise estimation energy is determined according to the updated first filter coefficient and the first audio detection signal.

[0064] like Figure 2 As shown in the figure, the self-noise generated by the running device is pure noise to device A and device B. Therefore, in order to remove the self-noise of the running device, the observed values ​​of the noise emitted by other devices such as device A and device B are needed as reference.

[0065] Specifically, based on the distributed wake-up network, all second voice devices send the calculated estimated reference signals to the first voice device, and the first voice device obtains the reference signal statistic sn according to all the estimated reference signals. f (t), the calculation formula is as follows:

[0066] s f (t) = [sn 0,f (t) T ,sn 1,f (t) T ,…,sn L-1,f (t) T ] T ;

[0067] Among them, sn 0，f (t),…,sn L-1，f (t) is the estimated reference signal sent by each second voice device; L is the number of second voice devices used to observe the first voice device; (X) T It means taking the transpose of any matrix or vector X.

[0068] Optionally, the first filter coefficient of the first voice device is updated according to the following formula, and the formula is specifically:

[0069] W s,f (t) = (I + β s,f (t)sn f (t)sn f (t) H )W s,f (t-1)-β s,f (t)sn f (t)x s,f (t);

[0070] Among them, W s,f (t) is the updated first filter coefficient, I is the unit matrix; β s,f (t) is a first filter update function of the first speech device; sn f (t) is the statistical value of the reference signal; (X) H represents the conjugate transpose of any matrix or vector X; W s,f (t-1) is the first filter coefficient before updating; x s,f (t) is the first audio detection signal on the target frequency band f; s is the serial number of the first voice device;

[0071] The reference signal statistic is determined based on the estimated reference signals of the second voice devices.

[0072] According to the distributed wake-up method provided by the embodiment of the present invention, the first voice device uses the first frequency domain signal and the estimated reference signals of each second voice device to update its own first filter coefficient, thereby achieving the purpose of eliminating noise interference.

[0073] Optionally, determining the noise estimation energy according to the updated first filter coefficient and the first audio detection signal includes:

[0074] determining a self-noise estimation signal according to the updated first filter coefficient and the first audio detection signal;

[0075] The self-noise estimation signal is input into an energy estimation model to obtain the noise estimation energy, wherein the energy estimation model is obtained by training a deep neural network model using a self-noise estimation sample signal, and the self-noise estimation sample signal is labeled with a self-noise energy.

[0076] The self-noise estimation signal may be the product of the updated first filter coefficient and the first audio detection signal.

[0077] A large number of self-noise estimation sample signals are collected in the distributed wake-up scenario of AI voice devices, and a self-noise energy label is attached to each self-noise estimation sample signal; then the energy estimation model is trained using the self-noise estimation sample signals with self-noise energy labels.

[0078] The self-noise energy annotation includes the self-noise energy annotation with error and the real self-noise energy annotation.

[0079] The energy estimation model is a deep neural network model (neural network, nnet), which can be built based on a long short-term memory network (Long Short-Term Memory, LSTM) or a convolutional layer, and can correct the erroneous self-noise estimation signal and output the noise estimation energy.

[0080] Noise Estimation Energy P s The calculation formula of (f) is:

[0081] P s (f) = nnet(W s,f (t)x s,f (t));

[0082] Among them, nnet is the energy estimation model; W s,f (t) is the first filter coefficient; x s,f (t) is the first audio detection signal on the target frequency band f; s is the serial number of the first voice device.

[0083] According to the distributed wake-up method provided by the embodiment of the present invention, by introducing the estimated reference signal of the second voice device, it is possible to effectively suppress the influence of the self-noise of the first voice device on the calculation of the distributed wake-up distance estimate by the artificial intelligence voice device. While reducing the amount of calculation, the accuracy of the distributed wake-up sound source distance calculation is greatly improved in the self-noise scenario of the artificial intelligence voice device, thereby improving the performance of distributed wake-up and the breadth of practical application.

[0084] Optionally, determining the received wake-up speech energy according to the noise estimation energy specifically includes:

[0085] Using the noise estimation energy, obtaining an energy estimation correction term;

[0086] According to the energy estimation correction item, the collected speech energy is corrected to determine the wake-up speech energy; wherein the collected speech energy is obtained by collecting the sound signal on the frequency band of the user's wake-up speech.

[0087] The first voice device estimates the energy of the noise on all sub-bands in the user wake-up voice band, and obtains the energy estimation correction item P self , the specific calculation formula is as follows:

[0088] P self =10log 10 (∑ f P s (f));

[0089] Among them, ∑ f P s (f) is the estimated noise energy P on all sub-bands s (f) performing summation; s is the serial number of the first voice device.

[0090] Energy estimation correction term P self Used to correct the collected voice energy, and then obtain the distributed wake-up distance without the influence of self-noise. The distributed wake-up distance is the distance between the first voice device and the sound source of the user's wake-up voice. The collected voice energy is obtained by the first voice device collecting the sound signal on the user's wake-up voice frequency band.

[0091] According to the distributed wake-up method provided by an embodiment of the present invention, the first language device jointly estimates the self-noise of the sound-emitting device with the help of other devices in the distributed system, which greatly improves the application scope and accuracy of the distributed wake-up algorithm in artificial intelligence voice devices, and especially optimizes the accuracy and robustness of distributed wake-up in self-noise scenarios.

[0092] Figure 3 FIG. 2 is a flow chart of a distributed wake-up method provided in an embodiment of the present invention, which is applied to a second voice device, such as Figure 3 As shown, the method includes but is not limited to the following steps:

[0093] First, in step S21, a second audio detection signal is determined according to the user wake-up voice.

[0094] Specifically, each second voice device can collect a second time domain signal based on the observation of the self-noise emitted by the first voice device, and convert the second time domain signal into a frequency domain signal as a second audio detection signal. Each second voice device obtains multiple second time domain signals in multiple preset time periods, and performs a fast Fourier transform on each second time domain signal to convert it into multiple second audio detection signals in the user wake-up voice frequency band, and each second audio detection signal is in a different frequency band. Among them, the larger the number of preset time periods, the larger the number of frequency bands corresponding to the second audio detection signal, and thus the higher the accuracy of noise reduction.

[0095] Further, in step S22, an estimated reference signal is determined according to the second audio detection signal.

[0096] In the AI ​​voice device distributed wake-up scenario, the distributed interconnection between AI voice devices can be used to transmit information about received noise to each other. Each AI voice device can communicate with each other and obtain the status of all AI voice devices in the current scenario. The AI ​​voice device includes a first voice device and a second voice device. The second voice device can distinguish the first voice device as the source of self-noise in the observed sound, thereby effectively helping the running first voice device to remove the influence of self-noise on its own judgment of the distributed wake-up distance.

[0097] Among them, the user wake-up voice can be used to wake up the AI ​​voice device in the distributed AI voice device wake-up scenario. The estimated reference signal is obtained by each second voice device based on the second time domain signal collected when observing the self-noise emitted by the first voice device.

[0098] like Figure 2 As shown in the figure, under the self-noise condition of the AI ​​voice device, for the sound signal received by the first voice device (i.e., the running device), in addition to the user wake-up voice issued by the user, there is also the noise generated by the first voice device (i.e., the running device) itself during operation, i.e., self-noise. This self-noise is the noise from the first voice device (i.e., the running device) to other AI voice devices (i.e., device A and device B). Therefore, in order to accurately determine the relative distance between the sound source of the user wake-up voice and the AI ​​voice device, it is necessary to remove the self-noise part during distance determination in order to effectively improve the robustness of the distributed wake-up distance. However, the traditional distributed wake-up algorithm based on energy estimation does not consider the impact of self-noise, especially the problem of self-noise suppression during device operation.

[0099] Specifically, for any second voice device, the second filter coefficient is obtained according to the second audio detection signal corresponding to the first voice device, and then the estimated reference signal is obtained according to the second audio detection signal and the second filter coefficient.

[0100] The user wake-up speech frequency band mainly refers to the audio spectrum range of the user wake-up speech. Since the self-noise in the user wake-up speech frequency band will interfere with the AI ​​voice device, it is necessary to remove the self-noise energy from the wake-up speech energy received by the AI ​​voice device.

[0101] Further, in step S23, the estimated reference signal is sent to the first voice device, so that the first voice device determines the received wake-up voice energy based on the estimated reference signal and the first audio detection signal; the first audio detection signal is determined by the first voice device based on the user wake-up voice; the first audio detection signal and the second audio detection signal include the noise generated when the first voice device is running; the wake-up voice energy is determined by the first voice device based on the noise estimation energy, and is used to determine the wake-up state of the first voice device; the noise estimation energy is determined by the first voice device based on the estimated reference signal and the first audio detection signal.

[0102] Specifically, at least one second voice device sends the estimated reference signal to the first voice device.

[0103] The first speech device calculates the noise estimation energy P according to these estimated reference signals s (f), and then estimate the energy P according to the noise s (f) Determine a distributed wake-up distance of the first voice device.

[0104] In addition, the distributed wake-up distances of each AI voice device can be compared, and the minimum value can be determined among the distributed wake-up distances of all AI voice devices, so that the AI ​​voice device corresponding to the minimum value is awakened by the user's wake-up voice, and other AI voice devices in the distributed AI voice device wake-up scenario are not awakened.

[0105] For a first voice device numbered s, a first audio detection signal x in a target frequency band f is determined from among a plurality of first audio detection signals. s，f (t), the first audio detection signal x s，f (t), the calculation formula is as follows:

[0106] x s,f (t) = [X(1) s,f (t),X(2) s,f (t),...,X(M) s,f (t)]k ;

[0107] Among them, M is the number of sound collection channels of the first voice device numbered s; f is the target frequency band, and after the first time domain signal is subjected to fast Fourier transform, f represents the frequency point; t is the order of the preset time periods; k is the number of the preset time periods, t=0, 1,…, k-1; s is the number of the first voice device.

[0108] like Figure 2 As shown in the figure, the self-noise generated by the running device is pure noise to device A and device B. Therefore, in order to remove the self-noise of the running device, the observed values ​​of the noise emitted by other devices such as device A and device B are needed as reference.

[0109] Based on the distributed wake-up network, all second voice devices send the calculated estimated reference signals to the first voice device, and the first voice device obtains the reference signal statistic sn according to all the estimated reference signals. f (t), the calculation formula is as follows:

[0110] s f (t) = [sn 0,f (t) T ,sn 1,f (t) T ,…,sn L-1,f (t) T ] T ;

[0111] Among them, sn 0，f (t),…,sn L-1，f (t) is the estimated reference signal sent by each second voice device; L is the number of second voice devices used to observe the first voice device; (X) T It means taking the transpose of any matrix or vector X.

[0112] The first voice device uses the first time domain signal to update its first filter coefficient, thereby eliminating noise interference. The updated calculation formula is as follows:

[0113] W s,f (t) = (I + β s,f (t)sn f (t)sn f (t) H )W s,f (t-1)-β s,f (t)sn f (t)x s,f (t);

[0114] Among them, W s,f(t) is the updated first filter coefficient, I is the unit matrix; β s,f (t) is a first filter update function of a first speech device; sn f (t) is the statistical value of the reference signal; (X) H represents the conjugate transpose of any matrix or vector X; W s,f (t-1) is the first filter coefficient before updating; x s,f (t) is the first audio detection signal on the target frequency band f; s is the serial number of the first voice device.

[0115] A large number of self-noise estimation sample signals are collected in the distributed wake-up scenario of AI voice devices, and a self-noise energy label is attached to each self-noise estimation sample signal; then the energy estimation model is trained using the self-noise estimation sample signals with self-noise energy labels.

[0116] The self-noise energy annotation includes the self-noise energy annotation with error and the real self-noise energy annotation.

[0117] The energy estimation model is a deep neural network model, which can be built based on LSTM or convolutional layers. It can correct the erroneous self-noise estimation signal and output the noise estimation energy.

[0118] Noise Estimation Energy P s The calculation formula of (f) is:

[0119] P s (f) = nnet(W s,f (t)x s,f (t));

[0120] Among them, nnet is the energy estimation model; W s,f (t) is the first filter coefficient; x s,f (t) is the first audio detection signal on the target frequency band f; s is the serial number of the first voice device.

[0121] The first speech device s estimates the energy of the noise on all sub-bands and obtains the energy estimation correction term P self , the specific calculation formula is as follows:

[0122] P self =10log 10 (∑ f P s (f));

[0123] Among them, ∑ f P s (f) is the estimated noise energy P in all frequency bands s (f) performing summation; s is the serial number of the first voice device.

[0124] The distributed wake-up method provided in the embodiment of the present invention can effectively suppress the influence of the self-noise of the first voice device on the calculation of the distributed wake-up distance estimation value by the artificial intelligence voice device by introducing the estimated reference signal of the second voice device. While reducing the amount of calculation, it greatly improves the accuracy of the distributed wake-up sound source distance calculation in the artificial intelligence voice device self-noise scenario, thereby improving the performance of distributed wake-up and the breadth of practical application.

[0125] Optionally, determining an estimated reference signal according to the second audio detection signal specifically includes:

[0126] updating a second filter coefficient according to the second audio detection signal;

[0127] Calculating a noise signal according to the updated second filter coefficient and the second audio detection signal;

[0128] The noise signal is mapped to the estimated reference signal according to a conversion parameter.

[0129] In each preset period, in the user wake-up voice band, any second voice device uses each sound collection channel to collect the second time domain signal. Among them, all second voice devices use microphone arrays, so the number of sound collection channels is determined by the microphone array of the device. For example, if there are 4 microphones, there will be 4 channels of data.

[0130] The second time domain signal may be collected in multiple preset time periods within the length (time) of the user's wake-up voice, and the preset time periods may be continuous, and the preset time periods may be equal or unequal.

[0131] For the second voice device numbered l, a second audio detection signal x in the target frequency band f is determined from among the plurality of second audio detection signals. l，f (t), the second audio detection signal x l，f The calculation formula of (t) is as follows:

[0132] x l,f (t) = [X(1) l,f (t),X(2) l,f (t),...,X(M) l,f (t)] k ;

[0133] Among them, M is the number of sound collection channels of the second voice device numbered l; f is the target frequency band, and after the second time domain signal is subjected to fast Fourier transform, f represents the frequency point; t is the order of the preset time periods; k is the number of the preset time periods, t=0, 1,…, k-1; l is the number of the second voice device.

[0134] Further, at the frequency point f, the second voice device l uses the second audio detection signal x l,f (t) to update its second filter coefficient, thereby achieving the purpose of eliminating noise interference. The updated calculation formula is as follows:

[0135] w l,f (t) = (I + α l,f (t)x l,f (t) H ) l,f (t-1)-α l,f (t)x l,f (t)d l,f (t) H ;

[0136] Among them, w l,f (t) is the updated second filter coefficient; f is the target frequency band; I is the unit matrix; α l,f (t) is the second filter update function of the second voice device 1; l,f (t) is the second audio detection signal of the second voice device 1 in the target frequency band f; (X) H represents the conjugate transpose of any matrix or vector X; w l,f (t-1) is the second filter coefficient before updating; d l,f (t) is the noise reference vector of the second speech device l, which is estimated by the second speech device in a sound source-free speech environment.

[0137] After the second voice device 1 updates the second filter coefficient, the estimated reference signal sn in the target frequency band f can be obtained. l,f (t), the calculation formula is as follows:

[0138] s l,f (t) = δ l,f (t)w l,f (t)x l,f (t)+γ l,f (t);

[0139] Among them, δ l,f (t) is the conversion coefficient; w l,f (t) is the second filter coefficient; x l,f (t) is the second audio detection signal of the second voice device 1; l,f (t) is the conversion deviation; f is the target frequency band.

[0140] The conversion parameters include the conversion factor δ l,f (t) and conversion deviation γ l,f (t), the noise signal w generated by the first voice device calculated by each second voice devicel,f (t)x l,f (t), mapped into an estimated reference signal.

[0141] According to the distributed wake-up method provided by the present invention, through the observation of noise by the second voice device, the second voice device obtains the second filter coefficient to update its own parameters, thereby achieving the purpose of reducing self-noise, and can effectively improve the accuracy of the second voice device at the sound source of the user's wake-up voice.

[0142] Figure 4 This is a flow chart of the third distributed wake-up method provided by an embodiment of the present invention, which is applied to a cloud server, such as Figure 4 As shown, including but not limited to the following steps:

[0143] First, in step S31, the cloud server receives a first audio detection signal sent by a first voice device, and simultaneously receives a second audio detection signal sent by at least one second voice device; the first audio detection signal is determined by the first voice device based on the user wake-up voice; the second audio detection signal is determined by each second voice device based on the user wake-up voice; the first audio detection signal and the second audio detection signal both include the noise generated when the first voice device is running.

[0144] The estimated reference signal is obtained by each second voice device based on the second time domain signal collected when observing the self-noise emitted by the first voice device.

[0145] Since the processing power of some AI voice devices is limited, the collected signals can be simply processed and uploaded to the cloud server. The cloud server will calculate and process the data and send the processing results to the corresponding AI voice device.

[0146] Specifically, the cloud server receives a second audio detection signal sent by at least one second voice device on the one hand, and receives a first audio detection signal sent by the first voice device on the other hand.

[0147] The user wake-up speech frequency band mainly refers to the audio spectrum range of the user wake-up speech. Since the self-noise in the user wake-up speech frequency band will interfere with the AI ​​voice device, it is necessary to remove the self-noise energy from the collected voice energy received by the AI ​​voice device.

[0148] For any AI voice device, including the first voice device and the second voice device, in the preset period, when receiving the self-noise emitted by the first voice device, each sound collection channel is used to obtain the time domain signal in the same frequency band as the user's wake-up voice band. Among them, all AI voice devices use microphone arrays, so the number of sound collection channels is determined by the microphone array of the device. For example, if there are 4 microphones, there will be 4 channels of data.

[0149] For the first voice device numbered s, a first audio detection signal x in the target frequency band f is determined from among the multiple first audio detection signals. s，f (t), obtain the first audio detection signal x s，f The calculation formula of (t) is as follows:

[0150] x s,f (t) = [X(1) s,f (t),X(2) s,f (t),...,X(M) s,f (t)] k ;

[0151] Among them, M is the number of sound collection channels of the first voice device numbered s; f is the target frequency band, and after the first time domain signal is subjected to fast Fourier transform, f represents the frequency point; t is the order of the preset time periods; k is the number of the preset time periods, t=0, 1,…, k-1; s is the number of the first voice device.

[0152] For the second voice device numbered l, a second audio detection signal x in the target frequency band f is determined from among the plurality of second audio detection signals. l，f (t), obtain the second audio detection signal x l，f The calculation formula of (t) is as follows:

[0153] x l,f (t) = [X(1) l,f (t),X(2) l,f (t),...,X(M) l,f (t)] k ;

[0154] Among them, M is the number of sound collection channels of the second voice device numbered l; f is the target frequency band, and after the second time domain signal is subjected to fast Fourier transform, f represents the frequency point; t is the order of the preset time periods; k is the number of the preset time periods, t=0, 1,…, k-1; l is the number of the second voice device.

[0155] Further, in step S32, an estimation reference signal is determined according to each second audio detection signal.

[0156] Specifically, at the frequency point f, the cloud server uses the second audio detection signal x l,f (t) to update the second filter coefficient of the second voice device 1, thereby eliminating the interference of noise on the second voice device 1. The updated calculation formula is as follows:

[0157] w l,f (t) = (I + α l,f (t)x l,f (t) H ) l,f (t-1)-α l,f (t)x l,f (t)d l,f (t) H ;

[0158] Among them, w l,f (t) is the new second filter coefficient; f is the target frequency band; I is the unit matrix; α l,f (t) is the second filter update function of the second voice device 1; l,f (t) is the second audio detection signal of the second voice device 1 in the target frequency band f; (X) H represents the conjugate transpose of any matrix or vector X; w l,f (t-1) is the second filter coefficient before updating; d l,f (t) is the noise reference vector of the second speech device l, which is estimated by the second speech device in a sound source-free speech environment.

[0159] After the cloud server obtains the new second filter coefficient of the second voice device 1, the new second filter coefficient is sent to the second voice device 1 to update the second filter coefficient of the second voice device 1.

[0160] Before step S32, the cloud server also needs to receive the second filter update function, the second filter coefficient before updating, and the noise reference vector of each second voice device.

[0161] Furthermore, the cloud server uses the second audio detection signal and the second filter system to obtain an estimated reference signal sn of the second voice device l in the target frequency band f. l,f (t), the calculation formula is as follows:

[0162] s l,f (t) = δ l,f (t)w l,f (t)x l,f (t)+γ l,f (t);

[0163] Among them, δ l,f (t) is the conversion coefficient; wl,f (t) is the second filter coefficient; x l,f (t) is the second audio detection signal of the second voice device 1; l,f (t) is the conversion deviation; f is the target frequency band.

[0164] The conversion parameters include the conversion factor δ l,f (t) and conversion deviation γ l,f (t), the noise signal w generated by the first voice device calculated by each second voice device l,f (t)x l,f (t), mapped into an estimated reference signal.

[0165] Further, in step S33, the cloud server determines the noise estimation energy according to at least one estimation reference signal and the first audio detection signal.

[0166] The cloud server obtains the reference signal statistic sn according to the estimated reference signal on the target frequency band f observed by all second voice devices for the first voice device. f (t), denoted as:

[0167] s f (t) = [sn 0,f (t) T ,sn 1,f (t) T ,…,sn L-1,f (t) T ] T ;

[0168] Among them, sn 0，f (t),…,sn L-1，f (t) is the estimated reference signal corresponding to each second speech device in the target frequency band f; L is the number of second speech devices observing the first speech device; (X) T It means taking the transpose of any matrix or vector X.

[0169] In addition, on the cloud server, the first filter coefficient of the first voice device is updated according to the first time domain signal of the first voice device s, thereby achieving the purpose of eliminating the interference of noise on the first voice device.

[0170] The first filter coefficient of the first voice device s is updated by the following calculation formula:

[0171] W s,f (t) = (I + β s,f (t)sn f (t)sn f (t) H )W s,f (t-1)-βs,f (t)sn f (t)x s,f (t);

[0172] Among them, W s,f (t) is the new first filter coefficient; f is the target frequency band; I is the unit matrix; β s,f (t) is a first filter update function of a first speech device; sn f (t) is the statistical value of the reference signal; (X) H represents the conjugate transpose of any matrix or vector X; W s,f (t-1) is the first filter coefficient before updating; x s,f (t) is the first audio detection signal on the target frequency band f; s is the serial number of the first voice device.

[0173] A large number of self-noise estimation sample signals are collected in the distributed wake-up scenario of AI voice devices, and a self-noise energy label is attached to each self-noise estimation sample signal; then the energy estimation model is trained using the self-noise estimation sample signals with self-noise energy labels.

[0174] The self-noise energy annotation includes the self-noise energy annotation with error and the real self-noise energy annotation.

[0175] The energy estimation model is a deep neural network model (neural network, nnet), which can be built based on a long short-term memory network (Long Short-Term Memory, LSTM) or a convolutional layer, and can correct the erroneous self-noise estimation signal and output the noise estimation energy.

[0176] On the cloud server, the noise estimation energy P s The calculation formula of (f) is:

[0177] P s (f) = nnet(W s,f (t)x s,f (t));

[0178] Among them, nnet is the energy estimation model; W s,f (t) is the new first filter coefficient; x s,f (t) is the first audio detection signal of the first voice device in the target frequency band f; s is the serial number of the first voice device.

[0179] Before step S33, the cloud server also needs to receive the first filter update function of the first voice device and the first filter coefficient before updating.

[0180] After step S33, the cloud server sends the second filter coefficient corresponding to the second audio detection signal to each second voice device, and sends the first filter coefficient to the first voice device.

[0181] Further, in step S34, the wake-up speech energy received by the first speech device is determined according to the noise estimation energy.

[0182] Specifically, the cloud server estimates the energy of the noise on all sub-bands and obtains the energy estimation correction term P self , the specific calculation formula is as follows:

[0183] P self =10log 10 (∑ f P s (f));

[0184] Among them, ∑ f P s (f) is the estimated noise energy P in all frequency bands s (f) performing summation; s is the serial number of the first voice device.

[0185] Energy estimation correction term P self The first voice device numbered s is used to correct the collected voice energy, and then obtains the wake-up voice energy without the influence of self-noise. The collected voice energy is obtained by the first voice device collecting the sound signal on the user's wake-up voice frequency band.

[0186] Before step S34, the cloud server also needs to receive the collected voice energy sent by the first voice device.

[0187] Further, in step S35, the wake-up speech energy is sent to the first speech device; the wake-up speech energy is used to determine the wake-up state of the first speech device.

[0188] Specifically, the cloud server sends the wake-up voice energy to the first voice device, and the first voice device calculates the distributed wake-up distance of the first voice device based on the wake-up voice energy, and each second voice device also obtains its own distributed wake-up distance based on the observed wake-up voice energy. The distributed wake-up distances of each AI voice device can be compared to determine the AI ​​voice device corresponding to the minimum distance, and then determine that the AI ​​voice device corresponding to the minimum value is awakened by the user's wake-up voice, and other AI voice devices in the distributed AI voice device wake-up scenario are not awakened.

[0189] The distributed wake-up method provided in the embodiment of the present invention can effectively suppress the influence of the self-noise of the first voice device on the calculation of the distributed wake-up distance estimation value by the artificial intelligence voice device by introducing the estimated reference signal of the second voice device. While reducing the amount of calculation, it greatly improves the accuracy of the distributed wake-up sound source distance calculation in the artificial intelligence voice device self-noise scenario, thereby improving the performance of distributed wake-up and the breadth of practical application.

[0190] Figure 5 is a structural diagram of a first voice device provided in an embodiment of the present invention, specifically a first voice device, which is used to execute a distributed wake-up method, such as Figure 5 As shown, the first voice device includes but is not limited to the following modules:

[0191] The first module 501 is used to determine a first audio detection signal according to a user wake-up voice, and receive an estimated reference signal sent by at least one second voice device, wherein the estimated reference signal is determined according to a second audio detection signal received by each second voice device, and the second audio detection signal is determined by each second voice device according to the user wake-up voice; the first audio detection signal and the second audio detection signal include noise generated when the first voice device is running;

[0192] A second module 502 is configured to determine noise estimation energy based on at least one estimation reference signal and the first audio detection signal;

[0193] The third module 503 is configured to determine the received wake-up speech energy according to the noise estimation energy, and determine the wake-up state based on the wake-up speech energy.

[0194] During the operation of the device, the first module 501 determines a first audio detection signal based on the user's wake-up voice, and receives an estimated reference signal sent by at least one second voice device, wherein the estimated reference signal is determined based on the second audio detection signal received by each second voice device, and the second audio detection signal is determined by each second voice device based on the user's wake-up voice; the first audio detection signal and the second audio detection signal include noise generated when the first voice device is running; the second module 502 determines noise estimation energy based on at least one estimated reference signal and the first audio detection signal; the third module 503 determines the received wake-up voice energy based on the noise estimation energy, and determines the wake-up state based on the wake-up voice energy.

[0195] First, the first module 501 determines a first audio detection signal based on the user wake-up voice, and receives an estimated reference signal sent by at least one second voice device, wherein the estimated reference signal is determined based on the second audio detection signal received by each second voice device, and the second audio detection signal is determined by each second voice device based on the user wake-up voice; the first audio detection signal and the second audio detection signal both include the noise generated when the first voice device is running.

[0196] Among them, the user's wake-up voice can be used to wake up the AI ​​voice devices in the distributed AI voice device wake-up scenario.

[0197] The estimated reference signal is obtained according to the second audio detection signal when each second audio device observes the self-noise emitted by the first audio device.

[0198] The estimated reference signal may be sent directly from the second voice device to the first voice device, or may be sent from the second voice device to a cloud server, and then forwarded to the first voice device by the cloud server.

[0199] Further, the second module 502 determines noise estimation energy according to at least one estimation reference signal and the first audio detection signal.

[0200] Specifically, the first voice device can update the first filter coefficient of the first voice device according to all the estimated reference signals and the first audio detection signal to achieve a better noise reduction effect.

[0201] The first speech device may also determine noise estimation energy based on the first filter coefficient and the first audio detection signal.

[0202] Further, the third module 503 determines the received wake-up speech energy according to the noise estimation energy, and determines the wake-up state based on the wake-up speech energy.

[0203] Specifically, the first voice device determines the distributed wake-up distance of the first voice device according to the noise estimation energy.

[0204] For the specific steps performed by the first voice device in this embodiment, please refer to the description of the above embodiment, which will not be repeated here.

[0205] The first voice device provided in an embodiment of the present invention can effectively suppress the influence of the self-noise of the first voice device on the calculation of the distributed wake-up distance estimation value by the artificial intelligence voice device by introducing the estimated reference signal of the second voice device. While reducing the amount of calculation, the accuracy of the distributed wake-up sound source distance calculation is greatly improved in the self-noise scenario of the artificial intelligence voice device, thereby improving the performance of distributed wake-up and the breadth of practical application.

[0206] Figure 6 is a schematic diagram of the structure of a distributed wake-up system provided by an embodiment of the present invention, such as Figure 6 As shown, including:

[0207] A first voice device 601 and at least one second voice device 602;

[0208] The second voice device 602 is used to determine a second audio detection signal according to the user wake-up voice; determine an estimated reference signal according to the second audio detection signal; and send the estimated reference signal to the first voice device;

[0209] The first voice device 601 is used to determine a first audio detection signal according to the user's wake-up voice, and receive the estimated reference signal, the first audio detection signal and the second audio detection signal including the noise generated when the first voice device is running; determine the noise estimation energy according to at least one estimated reference signal and the first audio detection signal; determine the received wake-up voice energy according to the noise estimation energy, and determine the wake-up state of the first voice device based on the wake-up voice energy.

[0210] During the operation of the system, the second voice device 602 determines a second audio detection signal according to the user's wake-up voice; determines an estimated reference signal according to the second audio detection signal; and sends the estimated reference signal to the first voice device; the first voice device 601 determines a first audio detection signal according to the user's wake-up voice, and receives the estimated reference signal, the first audio detection signal and the second audio detection signal including the noise generated when the first voice device is running; determines the noise estimation energy according to at least one estimated reference signal and the first audio detection signal; determines the received wake-up voice energy according to the noise estimation energy, and determines the wake-up state of the first voice device based on the wake-up voice energy.

[0211] In the AI ​​voice device distributed wake-up scenario, the distributed interconnection between AI voice devices can be used to transmit information about received noise to each other. Each AI voice device can communicate with each other and obtain the status of all AI voice devices in the current scenario. The AI ​​voice device includes a first voice device and a second voice device. The second voice device can distinguish the first voice device as the source of self-noise in the observed sound, thereby effectively helping the running first voice device to remove the influence of self-noise on its own judgment of the distributed wake-up distance.

[0212] In addition, the distributed wake-up distances of each AI voice device can be compared, and the minimum value can be determined among the distributed wake-up distances of all AI voice devices, so that the AI ​​voice device corresponding to the minimum value is awakened by the user's wake-up voice, and other AI voice devices in the distributed AI voice device wake-up scenario are not awakened.

[0213] Among them, the user wake-up voice can be used to wake up the AI ​​voice device in the distributed AI voice device wake-up scenario. The estimated reference signal is obtained by each second voice device based on the second time domain signal collected when observing the self-noise emitted by the first voice device.

[0214] The specific steps performed by the first voice device 601 and the second voice device 602 in the distributed wake-up system of the embodiment of the present invention are described in the previous embodiment and will not be repeated here.

[0215] like Figure 2 As shown in the figure, the self-noise generated by the running device is pure noise to device A and device B. Therefore, in order to remove the self-noise of the running device, the observed values ​​of the noise emitted by other devices such as device A and device B are needed as reference.

[0216] The distributed wake-up system provided by the embodiment of the present invention can effectively suppress the influence of the self-noise of the first voice device on the calculation of the distributed wake-up distance estimate by the artificial intelligence voice device by introducing the estimated reference signal of the second voice device. While reducing the amount of calculation, it greatly improves the accuracy of the distributed wake-up sound source distance calculation in the self-noise scenario of the artificial intelligence voice device, thereby improving the performance of distributed wake-up and the breadth of practical application.

[0217] It should be noted that the distributed wake-up system provided in the embodiment of the present invention can be implemented based on the distributed wake-up method described in any of the above embodiments during specific execution, which will not be described in detail in this embodiment.

[0218] Figure 7 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention, such as Figure 7 As shown, the electronic device may include: a processor 710, a communication interface 720, a memory 730 and a communication bus 740, wherein the processor 710, the communication interface 720 and the memory 730 communicate with each other through the communication bus 740. The processor 710 may call the logic instructions in the memory 730 to execute the distributed wake-up method provided in the above-mentioned various embodiments.

[0219] In addition, the logic instructions in the above-mentioned memory 730 can be implemented in the form of a software functional unit and can be stored in a computer-readable storage medium when it is sold or used as an independent product. Based on such an understanding, the technical solution of the embodiment of the present invention is essentially or the part that contributes to the prior art or the part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium, including a number of instructions to enable a computer device (which can be a personal computer, a server, or a network device, etc.) to perform all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (Read-Only Memory, ROM), random access memory (Random Access Memory, RAM), disk or optical disk and other media that can store program codes.

[0220] On the other hand, the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer can execute the distributed wake-up method provided in the above-mentioned embodiments.

[0221] On the other hand, the present invention further provides a non-transitory computer-readable storage medium having a computer program stored thereon, which is implemented when the computer program is executed by a processor to execute the distributed wake-up method provided in the above embodiments.

[0222] Through the description of the above implementation methods, those skilled in the art can clearly understand that each implementation method can be implemented by means of software plus a necessary general hardware platform, and of course, it can also be implemented by hardware. Based on this understanding, the above technical solution is essentially or the part that contributes to the prior art can be embodied in the form of a software product, and the computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, a disk, an optical disk, etc., including a number of instructions for a computer device (which can be a personal computer, a server, or a network device, etc.) to execute the methods described in each embodiment or some parts of the embodiments.

[0223] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A distributed wake-up method, It is characterized in that Applied to a first voice device, the method includes: Determine a first audio detection signal according to the user wake-up voice, and receive an estimated reference signal sent by at least one second voice device, wherein the estimated reference signal is determined according to the second audio detection signal received by each second voice device, and the second audio detection signal is determined by each second voice device according to the user wake-up voice; the first audio detection signal and the second audio detection signal both include noise generated when the first voice device is running; determining a noise estimation energy based on at least one estimation reference signal and the first audio detection signal; Determining a received wake-up speech energy according to the noise estimation energy, and determining a wake-up state based on the wake-up speech energy; The step of determining the noise estimation energy according to at least one estimation reference signal and the first audio detection signal comprises: updating a first filter coefficient of the first voice device according to an estimated reference signal of each second voice device and the first audio detection signal, wherein the estimated reference signal is obtained according to the second audio detection signal when each second voice device observes the self-noise emitted by the first voice device; The noise estimation energy is determined according to the updated first filter coefficient and the first audio detection signal.

2. The distributed wake-up method according to claim 1, It is characterized in that The first filter coefficient of the first voice device is updated according to the following formula, which is specifically: W s,f (t)=(I+β s,f (t)sn f (t)sn f (t) H )W s,f (t-1)-β s,f (t)sn f (t)x s,f (t); Among them, W s,f (t) is the updated first filter coefficient, I is the unit matrix; β s,f (t) is a first filter update function of the first speech device; sn f (t) is the statistical value of the reference signal; (X) H represents the conjugate transpose of any matrix or vector X; W s,f (t-1) is the first filter coefficient before updating; x s,f (t) is the first audio detection signal on the target frequency band f; s is the serial number of the first voice device; The reference signal statistic is determined based on the estimated reference signals of the second voice devices.

3. The distributed wake-up method according to claim 1, It is characterized in that Determining the noise estimation energy according to the updated first filter coefficient and the first audio detection signal includes: determining a self-noise estimation signal according to the updated first filter coefficient and the first audio detection signal; The self-noise estimation signal is input into an energy estimation model to obtain the noise estimation energy, wherein the energy estimation model is obtained by training a deep neural network model using a self-noise estimation sample signal, and the self-noise estimation sample signal is labeled with a self-noise energy.

4. The distributed wake-up method according to claim 1, It is characterized in that Determining the received wake-up speech energy according to the noise estimation energy specifically includes: Using the noise estimation energy, obtaining an energy estimation correction term; According to the energy estimation correction item, the collected speech energy is corrected to determine the wake-up speech energy; wherein the collected speech energy is obtained by collecting the sound signal on the frequency band of the user's wake-up speech.

5. The distributed wake-up method according to any one of claims 1 to 4, It is characterized in that Determining a first audio detection signal according to a user wake-up voice specifically includes: Acquire a first time domain signal according to the user wake-up voice; The first time domain signal is converted into a frequency domain signal to obtain the first audio detection signal.

6. A distributed wake-up method, It is characterized in that Applied to a second voice device, the method comprises: Determine a second audio detection signal according to the user wake-up voice; Determine an estimated reference signal according to the second audio detection signal, wherein the estimated reference signal is obtained according to the second audio detection signal when each second audio device observes the self-noise emitted by the first audio device; The estimated reference signal is sent to a first voice device, so that the first voice device determines the received wake-up voice energy according to the estimated reference signal and a first audio detection signal; the first audio detection signal is determined by the first voice device according to the user wake-up voice; the first audio detection signal and the second audio detection signal include noise generated when the first voice device is running; The wake-up speech energy is determined by the first speech device according to the noise estimation energy and is used to determine the wake-up state of the first speech device; The noise estimation energy is determined by the first speech device according to the estimation reference signal and the first audio detection signal; The step of determining an estimated reference signal according to the second audio detection signal specifically includes: updating a second filter coefficient according to the second audio detection signal; Calculating a noise signal according to the updated second filter coefficient and the second audio detection signal; The noise signal is mapped to the estimated reference signal according to a conversion parameter.

7. A distributed wake-up method, It is characterized in that Applied to a cloud server, the method includes: Receive a first audio detection signal sent by a first voice device, and simultaneously receive a second audio detection signal sent by at least one second voice device; the first audio detection signal is determined by the first voice device according to the user wake-up voice; the second audio detection signal is determined by each second voice device according to the user wake-up voice; the first audio detection signal and the second audio detection signal both include noise generated when the first voice device is running; determining an estimation reference signal according to each second audio detection signal; determining a noise estimation energy based on at least one estimation reference signal and the first audio detection signal; Determining, according to the noise estimation energy, the wake-up speech energy received by the first speech device; The wake-up speech energy is sent to the first speech device; the wake-up speech energy is used to determine the wake-up state of the first speech device; The step of determining the noise estimation energy according to at least one estimation reference signal and the first audio detection signal comprises: updating a first filter coefficient of the first voice device according to an estimated reference signal of each second voice device and the first audio detection signal, wherein the estimated reference signal is obtained according to the second audio detection signal when each second voice device observes the self-noise emitted by the first voice device; The noise estimation energy is determined according to the updated first filter coefficient and the first audio detection signal.

8. A first voice device, It is characterized in that Used to execute distributed wake-up methods, including: A first module is used to determine a first audio detection signal according to a user wake-up voice, and receive an estimated reference signal sent by at least one second voice device, wherein the estimated reference signal is determined according to a second audio detection signal received by each second voice device, and the second audio detection signal is determined by each second voice device according to the user wake-up voice; the first audio detection signal and the second audio detection signal include noise generated when the first voice device is running; A second module, configured to determine noise estimation energy based on at least one estimation reference signal and the first audio detection signal; A third module is used to determine the received wake-up speech energy according to the noise estimation energy, and determine the wake-up state based on the wake-up speech energy; The step of determining the noise estimation energy according to at least one estimation reference signal and the first audio detection signal comprises: updating a first filter coefficient of the first voice device according to an estimated reference signal of each second voice device and the first audio detection signal, wherein the estimated reference signal is obtained according to the second audio detection signal when each second voice device observes the self-noise emitted by the first voice device; The noise estimation energy is determined according to the updated first filter coefficient and the first audio detection signal.

9. A distributed wake-up system, It is characterized in that include: a first voice device and at least one second voice device; The second voice device is used to determine a second audio detection signal according to the user wake-up voice; determining an estimated reference signal according to the second audio detection signal; sending the estimated reference signal to the first speech device; The first voice device is used to determine a first audio detection signal according to a user wake-up voice and receive the estimated reference signal, wherein the first audio detection signal and the second audio detection signal include noise generated when the first voice device is running; determining a noise estimation energy based on at least one estimation reference signal and the first audio detection signal; Determining the received wake-up speech energy according to the noise estimation energy, and determining the wake-up state of the first speech device based on the wake-up speech energy; The step of determining the noise estimation energy according to at least one estimation reference signal and the first audio detection signal comprises: updating a first filter coefficient of the first voice device according to an estimated reference signal of each second voice device and the first audio detection signal, wherein the estimated reference signal is obtained according to the second audio detection signal when each second voice device observes the self-noise emitted by the first voice device; The noise estimation energy is determined according to the updated first filter coefficient and the first audio detection signal.

Images