A voice wake-up method and system based on time-domain binary neural network

Abstract

The present invention discloses a voice wake-up method and system based on a time-domain binary neural network. The method comprises the following steps: obtaining an audio file to be recognized, thereby obtaining a voice signal to be processed; extracting acoustic features from the voice signal to be processed, and performing dimensionality conversion processing on the voice signal to obtain acoustic features after dimensionality conversion processing; inputting the acoustic features after dimensionality conversion processing into a pre-trained time-domain binary neural network (TBNN) model to obtain probability outputs of keywords and non-keywords; and determining whether to wake up the system based on whether the highest probability among the probability outputs of keywords and non-keywords is a keyword. The present invention greatly reduces the number of parameters and the amount of computation of the neural network classifier, while significantly improving the wake-up speed and reducing the power consumption of the voice wake-up system.

Description

Technical Field

[0001] The present invention relates to a voice wake-up method and system based on a time-domain binary neural network, belonging to the technical field of voice wake-up. Background Art

[0002] The voice wake-up system usually runs on mobile devices, which have small memory and limited computing power. Therefore, the voice wake-up system should meet the requirements of high accuracy, small memory usage and low computational complexity. However, high-performance deep convolutional neural network models are highly complex and computationally intensive, and often require a large amount of memory, making it difficult to deploy them on mobile devices with smaller memory. Based on the above problems, the present invention proposes a new time-domain binary neural network (TBNN), which greatly reduces the memory usage and computational complexity of the voice wake-up system. However, the problem of the non-differentiable Sign function in the binary network will make the network difficult to train or insufficiently trained, which ultimately makes it difficult for the language wake-up accuracy to meet the usage standards. Summary of the Invention

[0003] The purpose of the present invention is to overcome the deficiencies in the prior art and provide a voice wake-up method and system based on a time-domain binary neural network, which greatly reduces the number of parameters and computational complexity of the neural network classifier, while significantly improving the wake-up speed and reducing the power consumption of the voice wake-up system.

[0004] To achieve the above object, the present invention is implemented by adopting the following technical solutions:

[0005] In a first aspect, the present invention provides a voice wake-up method based on a time-domain binary neural network, comprising:

[0006] Obtaining the audio file to be recognized, thereby obtaining the voice signal to be processed;

[0007] Extract acoustic features from the speech signal to be processed, and perform dimension transformation on it to obtain acoustic features after dimension transformation;

[0008] The acoustic features after dimension transformation are input into the pre-trained time-domain binary neural network (TBNN) model to obtain the probability output of keywords and non-keywords.

[0009] Whether to wake up the system is determined based on whether the highest probability among the probability outputs of keywords and non-keywords is the keyword.

[0010] Furthermore, the acoustic feature extraction is performed on the speech signal to be processed, and the dimensionality conversion is performed on the speech signal to obtain the acoustic features after the dimensionality conversion processing, including:

[0011] The speech signal to be processed is pre-emphasized, framed and windowed, fast Fourier transformed, Mel filtered, and logarithmized to obtain a 1*40*98 logarithmic Mel spectrum acoustic feature.

[0012] The logarithmic Mel spectrum acoustic features are transformed from 1*40*98 to 40*1*98, which is used as the input of the time domain binary neural network (TBNN) model.

[0013] Furthermore, the time-domain binary neural network TBNN model includes a time-domain convolution layer, a BN layer, a ReLU activation function, 6 time-domain binary convolution blocks, a global average pooling layer, and a fully connected layer.

[0014] Furthermore, the acoustic features after dimension transformation are input into a pre-trained time-domain binary neural network (TBNN) model to obtain the probability output of keywords and non-keywords, including:

[0015] The 40*1*98 logarithmic Mel spectrum acoustic features are input into the pre-trained time-domain binary neural network (TBNN) model, first entering the first time-domain convolution layer with a convolution kernel of 1*3 size.

[0016] The output of the first time-domain convolutional layer is regularized by the BN layer to reduce overfitting, and then passes through the ReLU activation function to increase the nonlinearity in the network;

[0017] The output of the ReLU activation function is input into the time-domain binary convolution block, where each time-domain binary convolution block consists of a time-domain binary convolution layer, a BN layer, and a Hardtanh activation function connected in sequence. In each time-domain binary convolution layer, a one-dimensional convolution kernel with a convolution kernel of 1*9 is used, and the parameters in the convolution kernel, the weight w, and the input a of the previous layer are all binarized values;

[0018] The output of the time domain binary convolution layer is obtained after passing through the BN layer and the Hardtanh activation function to obtain the output of the time domain binary convolution block;

[0019] The time-domain binary neural network TBNN model uses 6 time-domain binary convolution blocks, which are connected in sequence. The output of the last time-domain binary convolution block enters the global average pooling layer for downsampling, and is then input into the fully connected layer, which finally outputs continuous non-keyword and keyword probabilities.

[0020] Furthermore, the training method of the time-domain binary neural network TBNN model includes:

[0021] The time-domain binary neural network (TBNN) model is trained using a back-propagation algorithm to obtain the trained time-domain binary neural network (TBNN) model as a classifier for the voice wake-up system.

[0022] Among them, during the network training process, the periodic gradient approximation method is used to replace the Sign function for back propagation, specifically:

[0023] Skip the back propagation process of the Sign function and replace its gradient with the following formula:

[0024]

[0025] Where k = 0.01, x is the weight value or activation value, s is the gradient approximation, and s is a periodic value. The formula is as follows:

[0026]

[0027] Among them, Sstart, n, m are adjustable parameters, current_eopch is the number of cycles of current training, and cos is the cosine function;

[0028] The value of s changes periodically with the properties of the cosine function. At the beginning of training, s takes the largest value, and the network training speed is fast at this time.

[0029] As the number of training cycles changes, the value of s becomes smaller, which slows down the training speed of the network and plays a good role in fine-tuning the weight parameters in the network;

[0030] Since s changes periodically, the training of the weight parameters in the network will switch continuously between fast update and slow and precise update, making the network training more sufficient.

[0031] Furthermore, the determining whether to wake up the system based on whether the highest probability among the probability outputs of keywords and non-keywords is a keyword includes:

[0032] Set multiple keywords as wake-up words in the voice wake-up system to wake up the device;

[0033] Among the obtained keyword and non-keyword probability outputs, the keyword or non-keyword with the largest probability value is selected as the label and used as the output of the voice wake-up system this time. If the label is a keyword and the time since the last wake-up has exceeded the set time limit, the device is woken up according to the keyword.

[0034] In a second aspect, the present invention provides a voice wake-up system based on a time-domain binary neural network, comprising:

[0035] An acquisition module is used to acquire the audio file to be recognized, thereby obtaining the voice signal to be processed;

[0036] The first processing module is used to extract acoustic features from the speech signal to be processed and perform dimensionality conversion on the signal to obtain acoustic features after dimensionality conversion;

[0037] The second processing module is used to input the acoustic features after dimension transformation into the pre-trained time-domain binary neural network (TBNN) model to obtain the probability output of keywords and non-keywords;

[0038] The judgment module is used to judge whether to wake up the system based on whether the highest probability among the probability outputs of keywords and non-keywords is the keyword.

[0039] Furthermore, the time-domain binary neural network TBNN model includes a time-domain convolution layer, a BN layer, a ReLU activation function, 6 time-domain binary convolution blocks, a global average pooling layer, and a fully connected layer.

[0040] In a third aspect, the present invention provides an electronic device including a processor and a storage medium;

[0041] The storage medium is used to store instructions;

[0042] The processor is configured to operate according to the instructions to execute the steps of any of the aforementioned methods.

[0043] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, which implements the steps of any of the aforementioned methods when executed by a processor.

[0044] Compared with the prior art, the present invention has the following beneficial effects:

[0045] The present invention provides a voice wake-up method and system based on a time-domain binary neural network. A new time-domain binary neural network (TBNN) is proposed as a classifier for the voice wake-up system. Binarization is used to quantize a large number of parameters in the network to 1 bit, significantly reducing the memory usage of the voice wake-up system. The traditional two-dimensional convolution is replaced by a one-dimensional time-domain binary convolution. The original two-dimensional input is converted into a one-dimensional time-domain input using a one-dimensional time-domain transformation. A one-dimensional convolution kernel is used to increase the attention to the time-domain information and compress the frequency-domain information, significantly reducing the amount of computation. In addition, the multiplication and addition operations required for the original convolution can be converted into lower-cost shift operations such as xnor and popcount due to binarization. Therefore, the time-domain binary neural network (TBNN) in the present invention can significantly reduce the memory usage and computational complexity of the voice wake-up system, thereby reducing system power consumption. The present invention provides a lightweight voice wake-up system implementation solution suitable for chip deployment and convenient for mobile use.

[0046] In addition, to address the problem of network difficulty or insufficient training caused by the non-differentiable Sign function in binary neural networks, a new gradient approximation method, the periodic gradient approximation method, is proposed. This method allows parameters such as weights in the neural network TBNN to switch continuously between fast update and slow and precise update during training, effectively avoiding the network training difficulties caused by the non-differentiable Sign function in the above-mentioned binary neural network, ultimately improving the system's voice wake-up accuracy and providing a new idea for the optimization of binary neural networks. BRIEF DESCRIPTION OF THE DRAWINGS

[0047] Figure 1 This is a flow chart of a voice wake-up method based on a time-domain binary neural network provided by an embodiment of the present invention;

[0048] Figure 2 This is a schematic diagram of the voice wake-up system provided by an embodiment of the present invention;

[0049] Figure 3 This is a structural diagram of a time-domain binary neural network (TBNN) provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further described below in conjunction with the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention.

[0051] Example 1

[0052] This embodiment introduces a voice wake-up method based on a time-domain binary neural network, including:

[0053] Obtaining the audio file to be recognized, thereby obtaining the voice signal to be processed;

[0054] Extract acoustic features from the speech signal to be processed, and perform dimension transformation on it to obtain acoustic features after dimension transformation;

[0055] The acoustic features after dimension transformation are input into the pre-trained time-domain binary neural network (TBNN) model to obtain the probability output of keywords and non-keywords.

[0056] Whether to wake up the system is determined based on whether the highest probability among the probability outputs of keywords and non-keywords is the keyword.

[0057] The application process of the voice wake-up method based on the time-domain binary neural network provided in this embodiment specifically involves the following steps:

[0058] Step 1: Obtain the audio file to be recognized, thereby obtaining the speech signal to be processed.

[0059] Step 2: Perform acoustic feature extraction on the speech signal to be processed, outputting a logarithmic Mel spectrum feature frame, and performing dimensionality conversion. The resulting logarithmic Mel spectrum serves as the input for the neural network model in step 3. The main steps of feature extraction include pre-emphasis, framing and windowing, fast Fourier transform (FFT), Mel filtering, and logarithm extraction. Specifically, the frame length obtained during framing and windowing is 30ms, the frame shift is 10ms, the speech length for each feature processing is 1s (98 frames per second, 98 = (1000-30) / 10), and the number of Mel filters selected during Mel filtering is 40, resulting in a 40*98 logarithmic Mel spectrum feature. Furthermore, since the speech signal input is single-channel, the final logarithmic Mel spectrum acoustic feature obtained is 1*40*98. The logarithmic Mel spectrum is transformed from 1*40*98 to 40*1*98, that is, it is transformed from a single-channel 40*98 two-dimensional feature to a 40-channel 1*98 time domain feature, and used as the input of the neural network TBNN.

[0060] Step 3: Build a neural network time-domain binary neural network (TBNN) model, with the input being the 40*1*98 log-Mel spectrum acoustic features and the output being the continuous non-keyword and keyword probabilities. TBNN consists of a time-domain convolution layer, a BN layer, a ReLU activation function, 6 time-domain binary convolution blocks (TB-Conv Block), a global average pooling layer, and a fully connected layer, connected in sequence. Specifically, the 40*1*98 log-Mel spectrum features obtained in step 2 are first input to the first time-domain convolution layer, whose convolution kernel is 1*3 (a 3*3 two-dimensional convolution kernel is generally used in traditional convolution). Next, the output of the first time-domain convolution layer is regularized by the BN layer to reduce overfitting. It then passes through the ReLU activation function to increase the nonlinearity in the network and obtain the input of the time-domain binary convolution block (TB-Conv Block). Specifically, each time-domain binary convolution block is composed of a time-domain binary convolution layer (TB-Conv), a BN layer, and a Hardtanh activation function (similar to the ReLU activation function, a commonly used activation function) connected in sequence. Among them, in each time-domain binary convolution layer, a one-dimensional convolution kernel with a convolution kernel of 1*9 is used, and the parameters in the convolution kernel, that is, the weight w, and the input a of the previous layer are all binarized values ​​(only 1 or -1). The binarization formula is xb=Sign(x) (x is w / a, xb is the binarized w / a). Therefore, in the time-domain binary convolution layer of the present invention, the traditional convolution multiplication and addition operation will be replaced by a simple exclusive-or operation (xnor) and a counter operation (popcount), that is, the output y of this layer r =popcount(xnor(a b , w b)), thereby greatly reducing the computational complexity. In addition, the time domain convolution operation brought by the 1*9 convolution kernel adds more attention to the time domain information, and the frequency domain information is compressed, which can further reduce the amount of calculation. After that, the output of the time domain binary convolution layer is passed through the BN layer and the Hardtanh activation function to obtain the output of the time domain binary convolution block. The TBNN model in the present invention uses a total of 6 time domain binary convolution blocks, which are connected in sequence. The output of the last time domain binary convolution block enters the global average pooling layer for downsampling, and is then input into the fully connected layer, and finally outputs continuous non-keyword and keyword probabilities.

[0061] Step 4: Train the neural network TBNN. Use the commonly used back-propagation algorithm to train the TBNN, and use it to obtain the trained TBNN as a classifier for the voice wake-up system (that is, to obtain the weights of each layer in the network). During the back-propagation process, it is necessary to differentiate each calculation in step 3 (the calculation process in step 3 is the forward propagation), use the chain derivative rule to backpropagate the gradient, and update the weight parameters of each layer. However, the binarization function Sigh used in step 3 is not differentiable, so normal back-propagation cannot be performed, that is, the network cannot be trained. Therefore, in the present invention, a new gradient approximation method is also proposed, the periodic gradient approximation method, which is used to replace the Sign function for back-propagation, and the effect is significantly superior to the traditional STE method. Specifically, during the network training process, the back-propagation process of the Sign function is skipped, and its gradient (derivative) is replaced by the following formula:

[0062] Where k = 0.01, x is the weight value or activation value (because both weight and activation value need to be binarized), and s is the periodic value: Among them, Sstart, n, and m are adjustable parameters. In the TBNN of the present invention, the optimal values ​​after multiple experiments are S_start = 1.8, n = 8, and m = 1.2; current_eopch is the number of cycles in the current training cycle (i.e., the number of cycles), and cos is the cosine function. In this way, the value of s can be periodically adjusted according to the properties of the cosine function. At the beginning of training, s is at its maximum value, and almost all weights / activations are within the range [-s, s]. Their corresponding gradients are approximated to a larger value s. This ensures that the network training speed is relatively fast at this time, and can effectively initialize the network weight parameters, but there will be a large error. As the number of training cycles increases, the value of s decreases, the number of weights / activations within the range [-s, s] decreases, and their gradient approximation s is also smaller. This can slow the network training speed and effectively fine-tune the network weight parameters, thereby reducing error. (For weights / activations outside the range [-s, s], their gradients are approximated to k = 0.01, which is very small and has little impact on network training.) Ultimately, since s changes periodically, the training of the weight parameters in the network will switch continuously between fast update and slow and precise update, making the network training more sufficient and improving the wake-up accuracy.

[0063] Step 5: After training the TBNN neural network from Step 3 using Step 4, the trained TBNN is used as the classifier for the voice wake-up system. The voice wake-up system uses ten keywords as wake-up words to wake the device. For each new non-keyword and keyword probability output from the voice wake-up system neural network, the keyword or non-keyword with the highest probability is selected as the label and output as the current voice wake-up system output. If the label is a keyword and the time since the last wake-up has exceeded the set time limit, the device is woken up according to the keyword.

[0064] The following describes the contents designed in the above embodiment in conjunction with a preferred embodiment.

[0065] Figure 2 This is a block diagram of the voice wake-up system and a flowchart of the voice wake-up program design. The implemented voice wake-up system is primarily divided into training and testing phases. During the training phase, features are extracted from the training voice data to obtain logarithmic Mel-spectrogram feature inputs, which are then fed into the voice wake-up model (classifier), namely the time-domain binary neural network (TBNN) in this invention. The model is trained using the labels in the training voice data, thereby updating the network weights and ultimately producing a trained voice wake-up model (classifier).

[0066] During the training phase, the commonly used back-propagation algorithm is used to train the TBNN. Since the Sign function is not differentiable during the back-propagation process of the TBNN, the new periodic gradient approximation method proposed in this invention is used to replace the Sign function for back-propagation, and the effect is significantly superior to the traditional STE method. The specific implementation is as follows:

[0067]

[0068]

[0069] Where grad_output is the gradient used for backpropagation instead of the Sign function, i.e., the approximate gradient; k = 0.01, x is the weight value or activation value, s is the periodic value, S_start = 1.8, n = 8, m = 1.2; current_eopch is the number of cycles in the current training, and cos is the cosine function.

[0070] During the testing phase, the trained voice wake-up model can be used to infer and judge the features extracted from the test voice data. If it is a wake-up word, the system will be woken up, otherwise it will not be woken up. Figure 3 This is the structural diagram of the time-domain binary neural network TBNN proposed in the present invention.

[0071] Using the above Figure 2 After training and testing the voice wake-up system using the following process, the final voice wake-up accuracy for 12 categories (non-keyword, silence, and 10 keywords), as well as the number of parameters and floating-point operations (FLOPs) used by the neural network model in the system, are shown in Table 1. In Table 1, DSCNN, ResNet, and TC-ResNet are full-precision lightweight neural networks commonly used in current voice wake-up systems, and TBNN is the time-domain binary neural network in the present invention.

[0072]

[0073] Table 1 Performance comparison between full-precision network and B-ResNet8

[0074] As shown in Table 1, the voice wake-up system based on the time-domain binary neural network (TBNN) in the present invention achieves a wake-up accuracy of 95.3%. Compared with traditional full-precision network-based voice wake-up systems, such as the best-performing TC-ResNet, the accuracy only decreases by 0.8%, but the number of parameters is reduced by 81.5% and the amount of computation is reduced by 5.8%. Furthermore, compared with DSCNN and ResNet, the time-domain binary neural network (TBNN) in the present invention achieves higher wake-up accuracy while using fewer parameters and computation, demonstrating the superiority of the time-domain binary neural network (TBNN) in the voice wake-up system. The time-domain binary neural network (TBNN) proposed in the present invention significantly reduces the memory usage and computational overhead of the neural network model in the voice wake-up system, providing a lightweight voice wake-up system implementation suitable for chip deployment and mobile use.

[0075] Example 2

[0076] This embodiment provides a voice wake-up system based on a time-domain binary neural network, including:

[0077] An acquisition module is used to acquire the audio file to be recognized, thereby obtaining the voice signal to be processed;

[0078] The first processing module is used to extract acoustic features from the speech signal to be processed and perform dimensionality conversion on the signal to obtain acoustic features after dimensionality conversion;

[0079] The second processing module is used to input the acoustic features after dimension transformation into the pre-trained time-domain binary neural network (TBNN) model to obtain the probability output of keywords and non-keywords;

[0080] The judgment module is used to judge whether to wake up the system based on whether the highest probability among the probability outputs of keywords and non-keywords is the keyword.

[0081] Furthermore, the time-domain binary neural network TBNN model includes a time-domain convolution layer, a BN layer, a ReLU activation function, 6 time-domain binary convolution blocks, a global average pooling layer, and a fully connected layer.

[0082] Example 3

[0083] This embodiment provides an electronic device, including a processor and a storage medium;

[0084] The storage medium is used to store instructions;

[0085] The processor is configured to operate according to the instructions to execute the steps of the method according to any one of the first embodiments.

[0086] Example 4

[0087] This embodiment provides a computer-readable storage medium having a computer program stored thereon. When the program is executed by a processor, the steps of any one of the methods described in Embodiment 1 are implemented.

[0088] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A voice wake-up method based on time-domain binary neural network, characterized in that: include: Obtaining the audio file to be recognized, thereby obtaining the voice signal to be processed; Extract acoustic features from the speech signal to be processed and perform dimensionality transformation on it to obtain acoustic features after dimensionality transformation; including: The speech signal to be processed is pre-emphasized, framed and windowed, fast Fourier transformed, Mel filtered, and logarithmized to obtain a 1*40*98 logarithmic Mel spectrum acoustic feature. The logarithmic Mel spectrum acoustic features are transformed from 1*40*98 to 40*1*98, which is used as the input of the time domain binary neural network model; The acoustic features after dimension transformation are input into a pre-trained time-domain binary neural network (TBNN) model to obtain the probability output of keywords and non-keywords; the time-domain binary neural network (TBNN) model includes a time-domain convolution layer, a BN layer, a ReLU activation function, 6 time-domain binary convolution blocks, a global average pooling layer, and a fully connected layer; The training method of the time domain binary neural network TBNN model includes: The time-domain binary neural network (TBNN) model is trained using a back-propagation algorithm to obtain the trained time-domain binary neural network (TBNN) model as a classifier for the voice wake-up system. Among them, during the network training process, the periodic gradient approximation method is used to replace the Sign function for back propagation, specifically: Skip the back propagation process of the Sign function and replace its gradient with the following formula: ； Where k=0.01, x is the weight value or activation value, s is the gradient approximation, and s is a periodic value. The formula is as follows: ； Among them, Sstart, n, m are adjustable parameters, Sstart=1.8, n=8, m=1.2, current_eopch is the number of cycles of current training, and cos is the cosine function; The value of s changes periodically with the properties of the cosine function. At the beginning of training, s takes the largest value, and as the number of training cycles changes, the value of s becomes smaller. Whether to wake up the system is determined based on whether the highest probability among the probability outputs of keywords and non-keywords is the keyword.

2. The voice wake-up method based on time-domain binary neural network according to claim 1, characterized in that: The acoustic features after dimension transformation are input into a pre-trained time-domain binary neural network (TBNN) model to obtain the probability output of keywords and non-keywords, including: The 40*1*98 logarithmic Mel spectrum acoustic features are input into the pre-trained time-domain binary neural network (TBNN) model, first entering the first time-domain convolution layer with a convolution kernel of 1*3 size. The output of the first time-domain convolutional layer is regularized by the BN layer to reduce overfitting, and then activated by the ReLU function to increase the nonlinearity of the network. The output of the ReLU activation function is input into the time-domain binary convolution block, where each time-domain binary convolution block consists of a time-domain binary convolution layer, a BN layer, and a Hardtanh activation function connected in sequence. In each time-domain binary convolution layer, a one-dimensional convolution kernel with a convolution kernel of 1*9 is used, and the parameters in the convolution kernel, the weight w, and the input a of the previous layer are all binarized values; The output of the time domain binary convolution layer is obtained after passing through the BN layer and the Hardtanh activation function to obtain the output of the time domain binary convolution block; The time-domain binary neural network TBNN model uses 6 time-domain binary convolution blocks, which are connected in sequence. The output of the last time-domain binary convolution block enters the global average pooling layer for downsampling, and is then input into the fully connected layer, which finally outputs continuous non-keyword and keyword probabilities.

3. The voice wake-up method based on time-domain binary neural network according to claim 1, characterized in that: The determining whether to wake up the system based on whether the maximum probability among the probability outputs of keywords and non-keywords is the keyword includes: Set multiple keywords as wake-up words in the voice wake-up system to wake up the device; Among the obtained keyword and non-keyword probability outputs, the keyword or non-keyword with the largest probability value is selected as the label and used as the output of the voice wake-up system this time. If the label is a keyword and the time since the last wake-up has exceeded the set time limit, the device is woken up according to the keyword.

4. A voice wake-up system based on a time-domain binary neural network, using the voice wake-up method based on a time-domain binary neural network according to claim 1, characterized in that: include: An acquisition module is used to acquire the audio file to be recognized, thereby obtaining the voice signal to be processed; The first processing module is used to extract acoustic features from the speech signal to be processed and perform dimensionality conversion on the signal to obtain acoustic features after dimensionality conversion; The second processing module is used to input the acoustic features after dimension transformation into the pre-trained time-domain binary neural network (TBNN) model to obtain the probability output of keywords and non-keywords; The judgment module is used to judge whether to wake up the system based on whether the highest probability among the probability outputs of keywords and non-keywords is the keyword.

5. An electronic device, characterized in that: including processors and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to execute the steps of the method according to any one of claims 1 to 3.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the program is executed by a processor, the steps of the method according to any one of claims 1 to 3 are implemented.

Images