A method, device, and equipment for voice-based proximity wake-up
By using a deep learning model to predict audio feature vectors in a multi-device environment, the latency problem of existing voice proximity wake-up solutions is solved, achieving more accurate proximity wake-up and a better user experience.
Patent Information
- Application Number
- CN202410082310.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-01-19
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2044-01-19
AI Technical Summary
Existing voice wake-up solutions suffer from time delays in multi-device environments, impacting user experience.
By generating a wake-up event within a preset wake-up delay time, the smart voice device calculates the audio feature vector and sends it to the host device for prediction. The deep learning model then uses the probability of waking up the nearest device returned by the audio feature vector to perform the wake-up.
It achieves more accurate wake-up from the nearest location, reduces processing latency, improves user experience, and has stronger generalization ability in noisy environments.
Smart Images

Figure CN117953869B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of smart home technology, and in particular to a method, apparatus, and device for voice-activated wake-up. Background Technology
[0002] With the development of artificial intelligence technology, human-computer interaction has become a mainstream trend, and products with voice functionality are becoming increasingly common. These products typically require voice activation before proceeding. However, when multiple devices are present in a space and use the same wake-up word, multiple devices will respond when the user performs a voice activation. Usually, users only want the nearest device to respond.
[0003] Since wake-up based on proximity is an auxiliary function, it should not take up too much runtime. Existing solutions, such as wake-up based on sound source sorting, ultimately require calculating distance scores through models and comparing them pairwise. If there are many devices, the delay may affect the user experience. Summary of the Invention
[0004] In view of this, the purpose of the present invention is to provide a voice-based wake-up method, apparatus and device, which aims to solve the problems that existing wake-up solutions cause more time delays and result in poor user experience.
[0005] To achieve the above objectives, the present invention provides a voice proximity wake-up method, the method comprising:
[0006] In response to detecting that a user has issued voice data containing a wake-up word, at least one smart voice device is determined to generate a wake-up event based on a preset wake-up delay time.
[0007] The audio feature vector of the voice data is calculated by the determined intelligent voice device and sent to the host device. The wake-up model in the host device is used to predict the audio feature vector. The host device and multiple intelligent voice devices are networked on the same WIFI local area network.
[0008] The corresponding smart voice device is woken up according to the prediction result, wherein the prediction result is the probability of waking up the device returned by the wake-up model in the host device based on the audio feature vector.
[0009] Preferably, the at least one intelligent voice device that determines the generation of a wake-up event based on a preset wake-up delay time includes:
[0010] When the voice data is detected by the VAD module in the intelligent voice device, the voice data is input into the voice wake-up algorithm to obtain the corresponding wake-up value;
[0011] The wake-up value is compared with a set threshold. When the wake-up value is greater than the set threshold, the corresponding smart voice device is determined to generate the wake-up event.
[0012] Preferably, the step of calculating the audio feature vector of the voice data using the determined intelligent voice device includes:
[0013] The first audio segment of a preset length is extracted from the voice data, and the first audio segment contains a wake word;
[0014] FBANK feature extraction is performed on the first audio to obtain the audio feature vector.
[0015] Preferably, the network structure of the wake-up model includes multiple convolutional layers, a SE-BLOCK module, a channel attention module, a pooling layer, a pyramid fusion structure, and a fully connected layer.
[0016] Preferably, the training process of the wake-up model includes:
[0017] The training dataset is constructed by placing different devices at different locations in the same space to collect audio data, adding noise and random masking to the raw audio data, and setting the nearest device as the true label.
[0018] The training dataset is input into the model for training, and the model is optimized using a cross-entropy loss function and the Adam optimizer to obtain the trained wake-up model.
[0019] Preferably, waking up the corresponding smart voice device based on the prediction result includes:
[0020] The first voice device is woken up based on the highest probability of waking up the device returned by the host device.
[0021] When it is determined that the first voice device cannot be woken up, the second voice device that can be woken up is woken up according to the probability of waking up the device from high to low.
[0022] To achieve the above objectives, the present invention also provides a voice proximity wake-up device, the device comprising:
[0023] The detection unit is used to determine at least one smart voice device that generates a wake-up event based on a preset wake-up delay time when the user emits voice data containing a wake-up word.
[0024] A computing unit is configured to calculate the audio feature vector of the voice data through the determined intelligent voice device, and send the audio feature vector to the host device so as to predict the audio feature vector through the wake-up model in the host device, wherein the host device and multiple intelligent voice devices are networked on the same WIFI local area network;
[0025] A wake-up unit is used to wake up a first voice device so that the first voice device can perform a corresponding operation, wherein the first voice device is the intelligent voice device with the highest probability of being woken up by the host device based on the audio feature vector.
[0026] To achieve the above objectives, the present invention also proposes a voice proximity wake-up device, including a processor, a memory, and a computer program stored in the memory, wherein the computer program is executed by the processor to implement the steps of a voice proximity wake-up method as described in the above embodiments.
[0027] Preferably, the voice-activated wake-up device includes smart home devices.
[0028] To achieve the above objectives, the present invention also proposes a computer-readable storage medium storing a computer program that is executed by a processor to implement the steps of a voice proximity wake-up method as described in the above embodiments.
[0029] Beneficial effects:
[0030] The above solution, in response to detecting voice data containing a wake-up word emitted by a user, determines at least one smart voice device that will generate a wake-up event based on a preset wake-up delay time; calculates the audio feature vector of the voice data through the determined smart voice device, and sends the audio feature vector to a host device; the wake-up model in the host device predicts the audio feature vector; and wakes up the corresponding voice device based on the prediction result. The prediction result is the probability of waking up the device returned by the wake-up model in the host device based on the audio feature vector, which can more accurately wake up the nearest device, reduce processing latency, and improve the user experience.
[0031] The above solution abstracts the end-to-end wake-up process into multi-class classification and uses a wake-up model built based on deep learning to predict and wake up the nearest device. This results in stronger generalization under various noise conditions, which is difficult for traditional algorithms to overcome. Attached Figure Description
[0032] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0033] Figure 1 This is a flowchart illustrating a voice proximity wake-up method according to an embodiment of the present invention.
[0034] Figure 2 This is a schematic diagram of the overall process of a proximity wake-up scheme provided in an embodiment of the present invention.
[0035] Figure 3 This is a schematic diagram of the processing flow of an intelligent voice device provided in an embodiment of the present invention.
[0036] Figure 4 This is a schematic diagram of the processing flow of a host device provided in an embodiment of the present invention.
[0037] Figure 5 This is a schematic diagram of the network structure of a wake-up model provided in an embodiment of the present invention.
[0038] Figure 6 This is a schematic diagram of a voice-activated wake-up device according to an embodiment of the present invention.
[0039] The realization of the invention's objective, its functional characteristics, and advantages will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation
[0040] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention. Therefore, the following detailed description of the embodiments of the present invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to represent selected embodiments of the invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.
[0041] In the description of this invention, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined with "first" and "second" may explicitly or implicitly include one or more of that feature.
[0042] The present invention will be described in detail below with reference to the embodiments.
[0043] Existing wake-up-from-nearest-location schemes typically include the following:
[0044] 1) Sensor-based wake-up solutions: such as infrared, ultrasonic, and radar sensors. When the device is woken up, the sensors are used to measure the distance to nearby speakers. However, these sensors mainly utilize the propagation characteristics of waves to achieve distance measurement. In real-world applications, various obstructions are often encountered, and the human body is irregular, leading to errors in calculating propagation time. Furthermore, using sensors increases hardware costs and does not utilize voice information.
[0045] 2) SNR-based wake-up scheme: This scheme requires extracting a certain length of audio upon receiving a wake-up signal. The audio must include both a silence zone and a wake-up zone. The corresponding signal power is then calculated to obtain the signal-to-noise ratio (SNR) of the speech. Finally, the device with the highest SNR is selected for wake-up. The drawback of this approach is that the silence and wake-up zones require empirical segmentation, and the silence zone is susceptible to various types of noise, such as inherent background noise from the device itself.
[0046] 3) Wake-up scheme based on sound source ranking: This method acquires N voice signals from N devices within the same time period; performs feature extraction on the N voice signals to obtain N voice features; and uses a trained sound source distance model to calculate the distance scores of all devices to determine the final ranking relationship for wake-up. While this method is less affected by environmental characteristics and random noise and has strong generalization ability, the latency may affect the user experience if there are many devices, as it requires calculating distance scores separately using the model and performing pairwise comparisons.
[0047] Based on this, the present invention proposes a voice proximity wake-up method, which can be used for end-to-end proximity wake-up judgment, thereby reducing latency and improving user experience.
[0048] Reference Figure 1 The diagram shown is a flowchart of a voice-based wake-up method according to an embodiment of the present invention.
[0049] In this embodiment, the method includes:
[0050] S11, in response to detecting that the user has issued voice data containing a wake-up word, at least one smart voice device that generates a wake-up event is determined based on a preset wake-up delay time.
[0051] Furthermore, in step S11, the determination of at least one intelligent voice device that generates a wake-up event based on a preset wake-up delay time includes:
[0052] S11-1, When the voice data is detected by the VAD module in the intelligent voice device, the voice data is input to the voice wake-up algorithm to obtain the corresponding wake-up value;
[0053] S11-2, compare the wake-up value with a set threshold, and when the wake-up value is greater than the set threshold, determine that the corresponding smart voice device generates the wake-up event.
[0054] Reference Figure 2 As shown. In this embodiment, multiple devices are aggregated and networked, with the host device set as the master node and the others as child nodes. The host device refers to a powerful edge computing device, which is a third-party device. All the smart voice devices that need to be woken up are networked with the host device in a local area network. Due to device processing latency, a wake-up delay is set to wait for other devices to process. The device that generates the wake-up event is determined by a voice wake-up algorithm. Smart voice devices typically detect voice signals through VAD (Voice Activation Detection). When VAD detects voice, it inputs the audio segment into the voice wake-up algorithm and outputs a wake-up value P. When the wake-up value P exceeds a set threshold T, a device wake-up event is considered to have occurred. In the same space, if multiple devices use the same wake-up keyword, multiple device wake-up events will occur simultaneously. Our goal is to obtain a response from only the nearest device.
[0055] S12, the audio feature vector of the voice data is calculated by the determined intelligent voice device, and the audio feature vector is sent to the host device. The wake-up model in the host device is used to predict the audio feature vector. The host device and multiple intelligent voice devices are networked on the same WIFI local area network.
[0056] Furthermore, in step S12, calculating the audio feature vector of the voice data using the determined intelligent voice device includes:
[0057] S12-1, reverse the extraction of a first audio segment of a preset length from the voice data, wherein the first audio segment contains a wake-up word;
[0058] S12-2, Perform FBANK feature extraction on the first audio to obtain the audio feature vector.
[0059] In this embodiment, the system collects the audio feature vectors of the smart voice devices that have received wake-up events within a preset wake-up delay time. The audio feature vectors are obtained by independently processing the voice data through the smart voice devices. Specifically, a preset length of audio is extracted, such as 800ms (set manually based on experience, usually containing the wake-up word). The audio is then subjected to FBANK feature extraction (more feature vectors can also be used) to obtain the audio feature vectors. The frame segmentation is 25ms, the frame shift is 10ms, and 80 Mel filters are used, ultimately resulting in an [80, 78]-dimensional audio feature vector.
[0060] Reference Figure 3 As shown, the smart voice device initially sets a wake-up parameter `wakeup = 0`. When the wake-up value P obtained through the voice wake-up algorithm is greater than T, it checks whether `wakeup` equals 1. If it does, no action is taken; otherwise, `wakeup = 1`. The audio is then extracted in reverse and the FBANK feature is calculated. The `wakeup` parameter and the FBANK feature (audio feature vector) are sent to the host device via the local area network, and the device waits for the corresponding result. The `wakeup` parameter is checked to avoid sending signals to the host device too frequently, as this could result in two consecutive keyword wake-ups in some situations.
[0061] S13, wake up the corresponding smart voice device according to the prediction result, wherein the prediction result is the probability of waking up the device returned by the wake-up model in the host device based on the audio feature vector.
[0062] Reference Figure 4 As shown. In this embodiment, the audio feature vector calculated by the smart voice device is uploaded to the host device. The host device directly obtains the response device through the trained wake-up model, and finally returns the algorithm result to the corresponding smart voice device to perform the corresponding operation. Specifically, the host device function initially sets one parameter, judge = 0, delay = 0. When the host device receives the audio feature vector, it first determines whether the preset judge is equal to wakeup. If not, it sets judge to 1 and delay to 1, saves the first triggered audio vector feature to an array to be sent to the wake-up model, and sets the delay time to 100ms. During this period, it receives audio feature vectors from other devices. If the timeout occurs, it directly sends a command not to wake up other devices; otherwise, it sends all smart voice device features to the wake-up model to obtain the nearest wake-up device with the highest probability, sends wakeup = 1 to that device, sends the corresponding command to that device, sends wakeup = 0 to other unwakeable devices, and finally sets judge = 0.
[0063] Furthermore, the network structure of the wake-up model includes multiple convolutional layers, a SE-BLOCK module, a channel attention module, a pooling layer, a pyramid fusion structure, and a fully connected layer.
[0064] Furthermore, the training process of the wake-up model includes:
[0065] The training dataset is constructed by placing different devices at different locations in the same space to collect audio data, adding noise and random masking to the raw audio data, and setting the nearest device as the true label.
[0066] The training dataset is input into the model for training, and the model is optimized using a cross-entropy loss function and the Adam optimizer to obtain the trained wake-up model.
[0067] In this embodiment, the wake-up model is primarily a convolutional classification model. The construction process of this model is further explained below, including: taking 5 devices as an example, the audio feature vectors obtained from each device are concatenated by channel to obtain a feature map of [80, 78, 5]. Assuming a maximum of 10 devices in the same space (this can be manually configured according to actual needs), a tensor of size [80, 78, 10] with all zeros is generated. The first 5 channels of the feature map obtained above are assigned to this tensor as input. The model input is [80, 78, 10]. After three convolutional layers, the channels are expanded to 20, 40, and 80, respectively. Then, feature pyramid fusion is performed, fusing the coarse audio features at the bottom layer with the fine audio features at the top layer to utilize more audio details. (Refer to...) Figure 5 As shown, the convolutional layers all use 3x3 convolutional kernels, with RULE activation and batch normalization (BN). Then, the SE-BLOCK module uses channel attention to weight the obtained channels, emphasizing some features and weakening others. A 2x2 pooling layer is then used to obtain the first feature tensor [40, 39, 140]. Twelve residual blocks are then used consecutively to increase model complexity and improve feature extraction. These twelve residual blocks are then combined with a 3x3 convolutional layer with a stride of 2 and padding of 1 to downsample the feature tensor to [20, 20, 200]. Finally, global pooling is performed by channel, followed by a fully connected layer, resulting in a 1x10 output probability. The model output is a 10-dimensional probability vector, which, after passing through a softmax function, is equivalent to summing the 10 probability vectors to 1. For example, [0.9, 0.1, 0, 0....] indicates that the device with the highest probability is selected for response.
[0068] During data acquisition, the process needs to be conducted in the same space. For example, if there are five smart voice devices (A, B, C, D, and E), the devices need to be placed in different locations, or the speaker needs to move to different locations, for data acquisition. During acquisition, only the FBANK feature vectors extracted from the devices by channel splicing need to be saved, along with the corresponding encoding of the nearest device as the ground truth label. One-hot encoding will be used here; for example, if A is the nearest device, the corresponding ground truth label would be [1, 0, 0, 0, 0], and so on. In other words, the model trained using the nearest device as the modeling basis is considered to have a generalization ability based on the probability of identifying the nearest device. For example, if the model predicts the probabilities of devices 1, 2, and 3 as [0.8, 0.2, 0], then there is a 0.8 probability that device 1 is the nearest, so device 1 is activated.
[0069] In the specific training process, 10,000 sets of audio data collected from different device locations in the same space were used as the training set, 250 sets were used as the validation set, and 250 sets were used as the test set. Data augmentation was performed on the original audio data, including: playing music in the current environment, adding noise such as different speech, and performing random masking, randomly masking 0-5 frames in the temporal domain.
[0070] The loss function used is the cross-entropy function, and the optimizer is the Adam optimizer. The initial learning rate is set to 1e-3-1e-8 for cyclic training, with 1K iterations per cyclic. To prevent overfitting, weight decay is applied to the model. A training batch size of 24 is used. The model is activated when the loss function approaches stability. Testing shows that it achieves a certain accuracy on the test set.
[0071] In addition, the network architecture of the above model can be modified by using more audio features for classification to achieve proximity wake-up.
[0072] Furthermore, in step S13, waking up the corresponding smart voice device based on the prediction result includes:
[0073] The first voice device is woken up based on the highest probability of waking up the device returned by the host device.
[0074] When it is determined that the first voice device cannot be woken up, the second voice device that can be woken up is woken up according to the probability of waking up the device from high to low.
[0075] In this embodiment, the audio feature vector uploaded by the smart voice device is concatenated into the input model to obtain the device wake-up probability. The device with the highest wake-up probability is selected for wake-up. If the device finds that it cannot be woken up for some reason, the device with the second highest wake-up probability is selected for wake-up, and so on.
[0076] This solution abstracts the end-to-end implementation of proximity wake-up into a classification model. It uses the audio feature vector to directly predict wake-up through the wake-up model, saving the process of pairwise comparison and avoiding potential time delays that could lead to a poor user experience. Furthermore, this wake-up solution uses a deep learning model, which has stronger generalization ability under various noise conditions, a challenge that traditional algorithms struggle to overcome.
[0077] Reference Figure 6 The diagram shown is a structural schematic of a voice-based wake-up device according to an embodiment of the present invention.
[0078] In this embodiment, the device 20 includes:
[0079] The detection unit 21 is used to determine at least one smart voice device that generates a wake-up event based on a preset wake-up delay time when the user emits voice data containing a wake-up word.
[0080] The computing unit 22 is used to calculate the audio feature vector of the voice data through the determined intelligent voice device, and send the audio feature vector to the host device so as to predict the audio feature vector through the wake-up model in the host device, wherein the host device and multiple intelligent voice devices are networked based on the same WIFI local area network;
[0081] The wake-up unit 23 is used to wake up the first voice device so that the first voice device can perform a corresponding operation, wherein the first voice device is the smart voice device with the highest probability of being woken up by the host device based on the audio feature vector.
[0082] In another embodiment, the device further includes:
[0083] The second wake-up unit is used to wake up the second voice device when the first voice device cannot be woken up, so that the second voice device can perform corresponding operations, wherein the second voice device is the intelligent voice device with the second highest probability of being woken up by the host device.
[0084] Each unit module of the device 20 can execute the corresponding steps in the above method embodiment, so the details of each unit module will not be elaborated here. Please refer to the description of the corresponding steps above for details.
[0085] This invention also provides a voice proximity wake-up device, which includes the voice proximity wake-up apparatus as described above, wherein the voice proximity wake-up apparatus can employ... Figure 6 The structure of the embodiment, correspondingly, can be executed Figure 1The technical solutions of the method embodiments shown are similar in implementation principle and technical effect. For details, please refer to the relevant records in the above embodiments, which will not be repeated here.
[0086] Furthermore, the smart home devices include smart switch panels, smart screens, smart door locks, switch buttons, smart voice devices, alarm devices, smart robot vacuums, smart curtains, smart TVs, smart air conditioners, smart refrigerators, smart toilets, etc., which will not be listed here.
[0087] The device includes: a mobile phone, digital camera, or tablet computer, or other device with a camera function; or a device with an image processing function; or a device with an image display function. The device may include components such as a memory, processor, input unit, display unit, and power supply.
[0088] The memory can be used to store software programs and modules. The processor executes various functional applications and data processing by running the software programs and modules stored in the memory. The memory can mainly include a program storage area and a data storage area. The program storage area can store the operating system, application programs required for at least one function (such as image playback function), etc.; the data storage area can store data created according to the use of the device. In addition, the memory can include high-speed random access memory, and can also include non-volatile memory, such as at least one disk storage device, flash memory device, or other volatile solid-state storage device. Accordingly, the memory can also include a memory controller to provide access to the memory for the processor and input units.
[0089] The input unit can be used to receive input numerical, character, or image information, and to generate keyboard, mouse, joystick, optical, or trackball signal inputs related to user settings and function control. Specifically, in addition to a camera, the input unit of this embodiment may also include a touch-sensitive surface (e.g., a touch screen) and other input devices.
[0090] The display unit can be used to display information input by the user or information provided to the user, as well as various graphical user interfaces of the device. These graphical user interfaces can be composed of graphics, text, icons, video, and any combination thereof. The display unit may include a display panel, optionally configured as an LCD (Liquid Crystal Display), OLED (Organic Light-Emitting Diode), or other similar display panel. Furthermore, a touch-sensitive surface may cover the display panel. When the touch-sensitive surface detects a touch operation on or near it, it transmits the information to the processor to determine the type of touch event. Subsequently, the processor provides corresponding visual output on the display panel based on the type of touch event.
[0091] This invention also provides a computer-readable storage medium, which may be a computer-readable storage medium included in the memory described in the above embodiments; or it may be a standalone computer-readable storage medium not assembled into a device. The computer-readable storage medium stores at least one instruction, which is loaded and executed by a processor to implement... Figure 1 The illustrated method is a voice-based wake-up method. The computer-readable storage medium may be a read-only memory, a disk, or an optical disk, etc.
[0092] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Similar or identical parts between embodiments can be referred to interchangeably. For the device embodiments, equipment embodiments, and storage medium embodiments, since they are basically similar to the method embodiments, the descriptions are relatively simple, and relevant parts can be referred to the descriptions in the method embodiments.
[0093] Furthermore, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.
[0094] The foregoing description illustrates and describes preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be construed as excluding other embodiments. It can be used in various other combinations, modifications, and environments, and can be altered within the scope of the inventive concept by means of the foregoing teachings or techniques or knowledge in related fields. Any modifications and variations made by those skilled in the art that do not depart from the spirit and scope of the present invention should be within the protection scope of the appended claims.
Claims
1. A voice wake-up method, characterized by, The method includes: In response to detecting that a user has issued voice data containing a wake-up word, at least one smart voice device that generates a wake-up event is determined based on a preset wake-up delay time. The audio feature vector of the voice data is calculated by the determined intelligent voice device and sent to the host device. The wake-up model in the host device is used to predict the audio feature vector. The host device and multiple intelligent voice devices are networked on the same WIFI local area network. The wake-up model's network structure includes multiple convolutional layers, a SE-BLOCK module, a channel attention module, pooling layers, a pyramid fusion structure, and a fully connected layer. The prediction of the audio feature vector by the wake-up model in the host device includes: concatenating the audio feature vector along the channel dimension to obtain a multi-channel feature map, and inputting the multi-channel feature map into the wake-up model. The processing of the wake-up model includes: The multi-channel feature map is used as input, and the channels are expanded through multiple convolutional layers. The low-level audio features and high-level audio features are fused through a pyramid fusion structure. The obtained channels are weighted using the SE-BLOCK module and the channel attention module, and then pooled through a pooling layer to obtain the first feature tensor. Multiple residual blocks are used to extract features from the first feature tensor, and downsampling is performed in conjunction with convolutional layers. Finally, global pooling is performed by channel, connected to a fully connected layer, and the output is processed by the softmax function to obtain the wake-up device probability vector for each smart voice device. The corresponding smart voice device is woken up according to the prediction result, wherein the prediction result is the probability of waking up the device returned by the wake-up model in the host device based on the audio feature vector.
2. The method of claim 1, wherein, The at least one intelligent voice device that determines the generation of a wake-up event based on a preset wake-up delay time includes: When the voice data is detected by the VAD module in the intelligent voice device, the voice data is input into the voice wake-up algorithm to obtain the corresponding wake-up value; The wake-up value is compared with a set threshold. When the wake-up value is greater than the set threshold, the corresponding smart voice device is determined to generate the wake-up event.
3. The method of claim 1, wherein, The step of calculating the audio feature vector of the voice data using the determined intelligent voice device includes: The first audio segment of a preset length is extracted from the voice data, and the first audio segment contains a wake word; FBANK feature extraction is performed on the first audio to obtain the audio feature vector.
4. The method of claim 1, wherein, The training process of the wake-up model includes: The training dataset is constructed by placing different devices at different locations in the same space to collect audio data, adding noise and random masking to the raw audio data, and setting the nearest device as the true label. The training dataset is input into the model for training, and the model is optimized using a cross-entropy loss function and the Adam optimizer to obtain the trained wake-up model.
5. The method of claim 1, wherein, The step of waking up the corresponding smart voice device based on the prediction result includes: The first voice device is woken up based on the highest probability of waking up the device returned by the host device. When it is determined that the first voice device cannot be woken up, the second voice device that can be woken up is woken up according to the probability of waking up the device from high to low.
6. A voice proximity wake-up apparatus, comprising: Using a voice proximity wake-up method as described in any one of claims 1-5, the apparatus comprises: The detection unit is used to determine at least one smart voice device that generates a wake-up event based on a preset wake-up delay time when the user emits voice data containing a wake-up word. A computing unit is configured to calculate the audio feature vector of the voice data through the determined intelligent voice device, and send the audio feature vector to the host device so as to predict the audio feature vector through the wake-up model in the host device, wherein the host device and multiple intelligent voice devices are networked on the same WIFI local area network; A wake-up unit is used to wake up a first voice device so that the first voice device can perform a corresponding operation, wherein the first voice device is the intelligent voice device with the highest probability of being woken up by the host device based on the audio feature vector.
7. A voice proximity wake-up device, comprising: The device includes a processor, a memory, and a computer program stored in the memory, the computer program being executed by the processor to implement the steps of a voice proximity wake-up method as described in any one of claims 1 to 5.
8. The voice wake-up device of claim 7, wherein, The voice-activated wake-up device includes smart home devices.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that is executed by a processor to implement the steps of a voice proximity wake-up method as described in any one of claims 1 to 5.
Citation Information
Patent Citations
Voice assistant devices and voice waking-up method thereof
CN110335601A
Sound source distance sorting method and related product
CN115083390A
Voice wake-up method and device, mobile terminal, vehicle and storage medium
CN116682428A
