Keyword detection method and related device

Abstract

The application provides a keyword detection method and related equipment. The method comprises: performing feature extraction on a to-be-processed voice signal to obtain acoustic features; and inputting the acoustic features into a trained neural network model to output a keyword detection result, wherein the trained neural network model is a dynamic binary neural network. According to the technical solution, the to-be-processed voice signal is first subjected to feature extraction to obtain acoustic features, and then the acoustic features are input into the trained dynamic binary neural network model to output a keyword detection result. In this way, the parameter quantity and the calculation quantity in the neural network model are greatly reduced, the data storage space is reduced, the wake-up speed is improved, and the power consumption of the detection system is effectively reduced, thereby reducing the difficulty of hardware implementation. Meanwhile, the recognition accuracy of the keyword detection result is effectively improved.

Description

Technical Field

[0001] This invention relates to the field of speech recognition technology, and more specifically, to a keyword detection method, a keyword detection device, an electronic device, and a storage medium. Background Technology

[0002] With the rise of neural networks in recent years, they have been applied to keyword recognition. Existing keyword detection systems typically run on mobile devices, which have limited memory and computing power. Therefore, keyword detection systems must simultaneously meet the requirements of high accuracy, low memory usage, and low computational load.

[0003] However, high-performance deep convolutional neural network models are highly complex, computationally intensive, and often require a large amount of memory, making them difficult to deploy on mobile devices with limited memory. In other words, existing keyword detection systems require significant computing power and memory, making hardware implementation challenging. This not only makes it difficult to guarantee the detection accuracy of the neural network but also increases detection costs and system power consumption.

[0004] Therefore, a new technical solution is urgently needed to solve the above-mentioned technical problems. Summary of the Invention

[0005] The summary section introduces a series of simplified concepts, which will be further explained in detail in the detailed description section. The summary section of this invention is not intended to limit the key features and essential technical features of the claimed technical solution, nor is it intended to determine the scope of protection of the claimed technical solution.

[0006] In a first aspect, the present invention proposes a keyword detection method, comprising:

[0007] Feature extraction is performed on the speech signal to be processed to obtain acoustic features;

[0008] Acoustic features are input into a trained neural network model to output keyword detection results. The trained neural network model is a dynamic binary neural network.

[0009] Optionally, the dynamic binary neural network includes a first convolutional layer, a dynamic binary layer, a max pooling layer, and a fully connected layer;

[0010] The acoustic features are input into the trained neural network model, including:

[0011] The acoustic features are input into the first convolutional layer to output activation data and its corresponding weights;

[0012] The activation data and its corresponding weights are input into the dynamic binary layer to output dynamic binary activation output data. The dynamic binary layer includes multiple dynamic binary modules.

[0013] The dynamic binary activation output data is input into the max pooling layer to output downsampled data;

[0014] The downsampled data is input into the fully connected layer to output the keyword detection results.

[0015] Optionally, the dynamic binary module includes a second convolutional layer, a second batch normalization layer, and a second nonlinear layer;

[0016] The activation data and its corresponding weights are input into the dynamic binary layer to output dynamic binary activation output data, including:

[0017] The activation data and its corresponding weights are dynamically binarized using a dynamic binarization function to output the binarized activation data and its corresponding weights.

[0018] The binarized activation data and its corresponding weights are input into the second convolutional layer to perform a convolution operation on the binarized activation data and its corresponding weights, and then output the binarized convolution operation result.

[0019] The binarized convolution result is input into the second batch normalization layer to output the second standardized data;

[0020] The second standardized data is input into the second nonlinear layer to output dynamic binary activated output data.

[0021] Optionally, the dynamic binary neural network further includes a first batch normalization layer and a first nonlinear layer;

[0022] Before inputting the activation data and its corresponding weights into the dynamic binary layer, the method also includes:

[0023] The activation data is input into the first batch normalization layer to output the first standardized data.

[0024] The first standardized data is input into the first nonlinear layer to output the first enhanced activation data and its corresponding weights.

[0025] Optionally, the second nonlinear layer includes a dynamic activation function;

[0026] The method also includes:

[0027] During the training of a dynamic binary neural network, the adjustable biases in the dynamic binarization function and / or the dynamic activation function are adjusted to adjust the activation data and its corresponding weights, as well as the binarized activation data and its corresponding weights.

[0028] Alternatively, convolution operations can be implemented using shift operations.

[0029] Optionally, the dynamic binarization function includes full-precision weights after the first standardization process;

[0030] The method also includes:

[0031] For the full-precision weights after the first standardization process, the dynamic scaling factor calculated per channel is statistically analyzed.

[0032] During the training of a dynamic binary neural network, the dynamic scaling factor is adjusted to adjust the weights after binarization.

[0033] Secondly, a keyword detection device is also proposed, including:

[0034] The feature extraction module is used to extract features from the speech signal to be processed in order to obtain acoustic features;

[0035] The detection module is used to input acoustic features into a trained neural network model to output keyword detection results. The trained neural network model is a dynamic binary neural network.

[0036] Thirdly, an electronic device is also proposed, including a processor and a memory, wherein the memory stores computer program instructions, which are executed by the processor to perform the keyword detection method described above.

[0037] Fourthly, a storage medium is also proposed, on which program instructions are stored. When the program instructions are run, they are used to execute the keyword detection method described above.

[0038] According to the above technical solution, the speech signal to be processed is first subjected to feature extraction to obtain acoustic features. Then, the acoustic features are input into a trained dynamic binary neural network model to output keyword detection results. This significantly reduces the number of parameters and computational load in the neural network model, reduces data storage space, improves wake-up speed, and effectively reduces the power consumption of the detection system, thereby lowering the difficulty of hardware implementation. Simultaneously, it also effectively improves the recognition accuracy of keyword detection results.

[0039] The keyword detection method of the present invention, and other advantages, objectives and features of the present invention will be partly apparent from the following description, and partly understood by those skilled in the art through study and practice of the present invention. Attached Figure Description

[0040] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit this specification. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0041] Figure 1 A schematic flowchart of a keyword detection method according to an embodiment of the present invention is shown;

[0042] Figure 2 A schematic block diagram of a keyword detection device according to an embodiment of the present invention is shown; and

[0043] Figure 3 A schematic block diagram of an electronic device according to an embodiment of the present invention is shown. Detailed Implementation

[0044] The terms "first," "second," "third," "fourth," etc. (if present) in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus. The technical solutions of the embodiments of this application will now be clearly and completely described in conjunction with the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them.

[0045] According to a first aspect of the present invention, a keyword detection method is proposed. Figure 1 A schematic flowchart of a keyword detection method 100 according to an embodiment of the present invention is shown. Method 100 may include the following steps:

[0046] Step S110: Feature extraction is performed on the speech signal to be processed to obtain acoustic features.

[0047] For example, a sound acquisition device, such as a microphone, can be used to acquire speech to obtain a speech signal to be processed. In one embodiment, the speech signal to be processed can be subjected to the following processing operations in sequence to extract features from the speech signal to be processed, thereby obtaining acoustic features. The processing operations may include, for example: pre-emphasis, frame windowing, fast Fourier transform, Mel filtering, and logarithmic calculation. Correspondingly, in this embodiment, the final acoustic features obtained are log-Mel spectral features.

[0048] Step S130: Input the acoustic features into the trained neural network model to output the keyword detection results, wherein the trained neural network model is a dynamic binary neural network.

[0049] For example, after acquiring the acoustic features in step S110, the acoustic features can be input into a trained dynamic binary neural network to output the keyword detection result. In other words, in this application, a dynamic binary neural network is used as a classifier, and the original 32-bit activation values ​​and weight values ​​in the dynamic binary neural network are quantized to 1 bit. Compared with the traditional keyword detection method based on convolutional neural networks, the convolutional neural network is dynamically binarized, thereby compressing the neural network classifier and achieving a lightweight neural network model.

[0050] According to the above technical solution, the speech signal to be processed is first subjected to feature extraction to obtain acoustic features. Then, the acoustic features are input into a trained dynamic binary neural network model to output keyword detection results. This significantly reduces the number of parameters and computational load in the neural network model, reduces data storage space, improves wake-up speed, and effectively reduces the power consumption of the detection system, thereby lowering the difficulty of hardware implementation. Simultaneously, it also effectively improves the recognition accuracy of keyword detection results.

[0051] Optionally, the dynamic binary neural network includes a first convolutional layer, a dynamic binary layer, a max-pooling layer, and a fully connected layer. In this embodiment, step S130, inputting acoustic features into the trained neural network model, may include the following steps.

[0052] Step S131: Input the acoustic features into the first convolutional layer to output activation data and its corresponding weights.

[0053] Taking the Mel-spectrum acoustic feature as an example, the following explanation will illustrate this further. For instance, the Mel-spectrum acoustic feature frame can be used as the activation value input to the first convolutional layer for convolution operations, followed by activation data and corresponding weights.

[0054] Step S132: Input the activation data and its corresponding weights into the dynamic binary layer to output dynamic binary activation output data. The dynamic binary layer includes multiple dynamic binary modules.

[0055] After obtaining the activation data and its corresponding weights, they can be input into a dynamic binary layer. Multiple dynamic binary modules within the layer perform dynamic binary quantization on the activation data and weights, ultimately outputting dynamic binary activation data. In essence, multiple dynamic binary modules are connected in series, continuously performing dynamic binary quantization on the activation data from the previous module until the activation output data from the last module is obtained.

[0056] Step S133: Input the dynamic binary activation output data into the max pooling layer to output downsampled data.

[0057] The activation output data from the last dynamic binary module is input into the max pooling layer for downsampling, thereby outputting downsampled data.

[0058] Step S134: Input the downsampled data into the fully connected layer to output the keyword detection results.

[0059] Inputting downsampled data into a fully connected layer outputs continuous probabilities for both keywords and non-keywords. Essentially, the keyword and non-keyword probabilities represent the keyword detection results. Specifically, ten keywords can be set. After obtaining the keyword detection results, the keyword or non-keyword with the highest probability value can be selected as the label for this detection output. If the label is a keyword, the result is output and displayed; if the label is a non-keyword, it is not displayed.

[0060] This significantly improves the accuracy of keyword detection results and greatly enhances the feasibility of applying binarization networks to keyword detection methods.

[0061] Preferably, the dynamic binary neural network may further include a first batch normalization layer and a first nonlinear layer. Before inputting the activation data and its corresponding weights into the dynamic binary layer in step S132, the method may further include the following steps.

[0062] Step S121: Input the activation data into the first batch normalization layer to output the first standardized processing data.

[0063] For example, before binarizing the weights w, they can be standardized first, i.e., w can be normalized. The purpose is to balance the weight distribution while reducing the quantization error when the weights w are binarized to 1 and -1. Specifically, the activation data can be input into the first batch normalization layer for standardization to output the first standardized data.

[0064] Step S122: Input the first normalized data into the first nonlinear layer to output the first enhanced activation data and its corresponding weights.

[0065] After standardization, the data is input into the first nonlinear layer. The ReLU activation function is then used to increase the nonlinearity of the activation values, reduce overfitting, and finally obtain the first enhanced activation data and its corresponding weights.

[0066] Optionally, the dynamic binary module may include a second convolutional layer, a second batch normalization layer, and a second nonlinear layer. In this embodiment, step S132, which inputs activation data and its corresponding weights into the dynamic binary layer to output dynamic binary activation output data, may include the following steps.

[0067] Step S132a: Dynamically binarize the activation data and its corresponding weights using a dynamic binarization function to output the binarized activation data and its corresponding weights.

[0068] For example, the dynamic binarization function can be the Dy-Sign function, as shown in the following equation: Among them, a b w b Let a represent the binarized activation data and weights, respectively. r w and w represent the full-precision activation data and weights, respectively. r This represents the standardized full-precision weights, and bias is the dynamically adjustable bias during training.

[0069] Step S132b: Input the binarized activation data and its corresponding weights into the second convolutional layer to perform a convolution operation on the binarized activation data and its corresponding weights, and then output the binarized convolution operation result.

[0070] For example, the binarized activation data a obtained from the above steps b With weight w b The values ​​are 1 and -1. When input into the second convolutional layer, Dy-B-Conv, a convolution operation is performed, the output is a binarized convolution result y. r As shown in the following formula: y r =Ds*(popcount[xnor(a b ,w b )]).

[0071] Step S132c: Input the binarized convolution operation result into the second batch normalization layer to output the second standardized data.

[0072] Similarly, after obtaining the binarized convolution result, it can be input into the second batch normalization layer for standardization, and then the second standardized data can be obtained.

[0073] In step S132d, the second normalized processing data is input into the second nonlinear layer to output dynamic binary activation output data.

[0074] The second standardized data is then input into the second nonlinear layer and passed through the dynamic activation function Dy-ReLU to increase the nonlinearity of the second standardized data, reduce overfitting, and further optimize the activation distribution to reduce quantization error.

[0075] In one specific embodiment, the second nonlinear layer may include a dynamic activation function. Specifically, the dynamic activation function is as follows: DyReLU(x) = max(0, x + bias), where bias is a dynamically adjustable bias during training, i.e., an adjustable bias.

[0076] In this embodiment, the method may further include: step S140, during the training process of the dynamic binary neural network, adjusting the adjustable bias in the dynamic binarization function and / or the dynamic activation function to adjust the activation data and its corresponding weights and the binarized activation data and its corresponding weights.

[0077] For example, in the training of a dynamic binary neural network model, the adjustable bias in the dynamic binarization function Dy-Sign and the dynamic activation function Dy-ReLU can be fine-tuned. That is, the weights and activation data before and after binarization can be adjusted using the adjustable bias, thereby reducing the quantization error of the weight w or activation value a caused by binarization. This greatly alleviates the defect of excessive information loss in the binarization network and improves the recognition accuracy of the network.

[0078] Alternatively, convolution operations can be implemented using shift operations.

[0079] Specifically, after binarizing the weights and activation data, the convolutional operations of the Dy-B-Conv layer can be implemented through simple shift operations, such as XNOR and Popcount operations, instead of the complex multiplication and addition operations used in traditional convolutional neural networks.

[0080] This not only greatly reduces the memory footprint and computational cost of the network model, but also saves on the overall overhead of implementing keyword detection methods at the hardware level.

[0081] Optionally, the dynamic binarization function may include full-precision weights after a first normalization process. In this embodiment, the method may further include the following steps.

[0082] Step S151: For the full-precision weights after the first standardization process, calculate the dynamic scaling factor per channel.

[0083] As shown above, the full-precision weight after the first standardization process is w, and the dynamic scaling factor Dy-scale, i.e. Ds, is calculated by channel for this weight.

[0084] Step S152: During the training process of the dynamic binary neural network, the dynamic scaling factor is adjusted to adjust the weights after binarization.

[0085] After obtaining the dynamic scaling factor statistically, this factor is dynamically adjusted as the training process progresses. This adjustment can include increasing or decreasing the scaling factor. The dynamic scaling factor Dy-scale can be used to scale the binarized weights, thereby reducing the quantization error of the weight values ​​and ultimately improving the accuracy of keyword detection results.

[0086] It is understood that steps S140 and S150 are merely different steps in different embodiments. In the actual execution of the method, there is no order of execution, and the above sequence numbers only serve to distinguish them.

[0087] According to a second aspect of the present invention, a keyword detection device is also proposed. Figure 2 A schematic block diagram of a keyword detection device 200 according to an embodiment of the present invention is shown. Figure 2 As shown, the device 200 may include a feature extraction module 210 and a detection module 220.

[0088] Feature extraction module 210 is used to extract features from the speech signal to be processed in order to obtain acoustic features;

[0089] The detection module 220 is used to input acoustic features into a trained neural network model to output keyword detection results, wherein the trained neural network model is a dynamic binary neural network.

[0090] According to a third aspect of the present invention, an electronic device is also provided. Figure 3 A schematic block diagram of an electronic device 300 according to an embodiment of the present invention is shown. Figure 3 As shown, the electronic device 300 includes a processor 310 and a memory 320, wherein the memory 320 stores computer program instructions, which are executed by the processor 310 to perform the keyword detection method described above.

[0091] According to a fourth aspect of the present invention, a storage medium is also provided, on which program instructions are stored, which, when executed, are used to perform the keyword detection method described above. The storage medium may, for example, include a storage component of a tablet computer, a hard disk of a computer, a read-only memory (ROM), an erasable programmable read-only memory (EPROM), a portable compact disc read-only memory (CD-ROM), a USB memory, or any combination of the above storage media. The computer-readable storage medium may be any combination of one or more computer-readable storage media.

[0092] Those skilled in the art can understand the specific details and beneficial effects of the keyword detection device, electronic equipment, and storage medium by reading the above description of the keyword detection method, and will not be repeated here for the sake of brevity.

[0093] In the several embodiments provided in this application, it should be understood that the disclosed apparatus and / or device can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0094] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0095] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0096] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0097] The above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit it. Although this application has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of this application.

Claims

1. A keyword detection method, characterized in that, include: Feature extraction is performed on the speech signal to be processed to obtain acoustic features; The following processing operations are performed sequentially on the speech signal to be processed in order to extract features from the speech signal and obtain acoustic features. The processing operations include: pre-emphasis, frame windowing, fast Fourier transform, Mel filtering, and logarithmic calculation. The final acoustic features obtained are log-Mel spectrum features. The acoustic features are input into a trained neural network model to output keyword detection results, wherein the trained neural network model is a dynamic binary neural network. The dynamic binary neural network includes a first convolutional layer, a dynamic binary layer, a max pooling layer, and a fully connected layer; The step of inputting the acoustic features into the trained neural network model includes: The acoustic features are input into the first convolutional layer to output activation data and its corresponding weights; The activation data and its corresponding weights are input into the dynamic binary layer to output dynamic binary activation output data, wherein the dynamic binary layer includes multiple dynamic binary modules. The dynamic binary activation output data is input into the max pooling layer to output downsampled data; The downsampled data is input into the fully connected layer to output the keyword detection result; The dynamic binary module includes a second convolutional layer, a second batch normalization layer, and a second nonlinear layer. The step of inputting the activation data and its corresponding weights into the dynamic binary layer to output dynamic binary activation output data includes: The activation data and its corresponding weights are dynamically binarized using a dynamic binarization function to output the binarized activation data and its corresponding weights. The dynamic binarization function is the Dy-Sign function, as shown in the following equation: , , ,in, , These represent the binarized activation data and weights, respectively. , These represent the full-precision activation data and the weights, respectively. This represents the standardized full-precision weights. This is a dynamically adjustable bias during training. This represents the mean of the full-precision weights; The standard deviation of the full-precision weights; The binarized activation data and its corresponding weights are input into the second convolutional layer to perform a convolution operation on the binarized activation data and its corresponding weights, and then output the binarized convolution operation result. The binarized convolution operation result is input into the second batch normalization layer to output the second standardized data; The second standardized data is input into the second nonlinear layer to output the dynamic binary activated output data; The second nonlinear layer includes a dynamic activation function; The method further includes: During the training process of the dynamic binary neural network, the adjustable biases in the dynamic binarization function and / or the dynamic activation function are adjusted to adjust the activation data and its corresponding weights, and the binarized activation data and its corresponding weights.

2. The keyword detection method as described in claim 1, characterized in that, The dynamic binary neural network further includes a first batch normalization layer and a first nonlinear layer; Before inputting the activation data and its corresponding weights into the dynamic binary layer, the method further includes: The activation data is input into the first batch normalization layer to output the first standardized processing data; The first standardized data is input into the first nonlinear layer to output the first enhanced activation data and its corresponding weights.

3. The keyword detection method as described in claim 1, characterized in that, The convolution operation is implemented through a shift operation.

4. The keyword detection method as described in claim 3, characterized in that, The dynamic binarization function includes full-precision weights after the first standardization process; The method further includes: For the full-precision weights after the first standardization process, the dynamic scaling factor calculated per channel is statistically analyzed. During the training process of the dynamic binary neural network, the dynamic scaling factor is adjusted to adjust the weights after binarization.

5. A keyword detection device, characterized in that, include: The feature extraction module is used to extract features from the speech signal to be processed in order to obtain acoustic features; The following processing operations are performed sequentially on the speech signal to be processed in order to extract features from the speech signal and obtain acoustic features. The processing operations include: pre-emphasis, frame windowing, fast Fourier transform, Mel filtering, and logarithmic calculation. The final acoustic features obtained are log-Mel spectrum features. The detection module is used to input the acoustic features into a trained neural network model to output keyword detection results, wherein the trained neural network model is a dynamic binary neural network. The dynamic binary neural network includes a first convolutional layer, a dynamic binary layer, a max pooling layer, and a fully connected layer; The step of inputting the acoustic features into the trained neural network model includes: The acoustic features are input into the first convolutional layer to output activation data and its corresponding weights; The activation data and its corresponding weights are input into the dynamic binary layer to output dynamic binary activation output data, wherein the dynamic binary layer includes multiple dynamic binary modules. The dynamic binary activation output data is input into the max pooling layer to output downsampled data; The downsampled data is input into the fully connected layer to output the keyword detection result; The dynamic binary module includes a second convolutional layer, a second batch normalization layer, and a second nonlinear layer. The step of inputting the activation data and its corresponding weights into the dynamic binary layer to output dynamic binary activation output data includes: The activation data and its corresponding weights are dynamically binarized using a dynamic binarization function to output the binarized activation data and its corresponding weights. The dynamic binarization function is the Dy-Sign function, as shown in the following equation: , , ,in, , These represent the binarized activation data and weights, respectively. , These represent the full-precision activation data and the weights, respectively. This represents the standardized full-precision weights. This is a dynamically adjustable bias during training. This represents the mean of the full-precision weights; The standard deviation of the full-precision weights; The binarized activation data and its corresponding weights are input into the second convolutional layer to perform a convolution operation on the binarized activation data and its corresponding weights, and then output the binarized convolution operation result. The binarized convolution operation result is input into the second batch normalization layer to output the second standardized data; The second standardized data is input into the second nonlinear layer to output the dynamic binary activated output data; The second nonlinear layer includes a dynamic activation function; The device further includes: During the training process of the dynamic binary neural network, the adjustable biases in the dynamic binarization function and / or the dynamic activation function are adjusted to adjust the activation data and its corresponding weights, and the binarized activation data and its corresponding weights.

6. An electronic device, characterized in that, It includes a processor and a memory, wherein the memory stores computer program instructions, which, when executed by the processor, are used to perform the keyword detection method as described in any one of claims 1 to 4.

7. A storage medium storing program instructions that, when executed, perform the keyword detection method as described in any one of claims 1 to 4.

Images