False wake-up corpus acquisition method and related method, device, equipment and storage medium

Abstract

This application discloses a method for acquiring false wake-up corpus, as well as related methods, apparatus, devices, and storage media. The method includes: constructing model instructions based on the product requirements of a wake-up model; synthesizing speech from the false wake-up text output by a generative model in response to the model instructions to obtain false wake-up speech; obtaining a reward value generated by the generative model based on the wake-up model's response to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text; adjusting the network parameters of the generative model based on the reward value, and determining whether to select the false wake-up text and false wake-up speech as sample corpus for training the wake-up model based on the reward value; and returning to the step of constructing model instructions based on the product requirements of the wake-up model, until the termination condition is met. This approach can improve the efficiency and quality of acquiring false wake-up corpus.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a method for acquiring false wake-up corpus, as well as related methods, apparatus, devices and storage media. Background Technology

[0002] Voice wake-up technology is the "gatekeeper" of human-computer interaction, but it faces a fundamental challenge in the real world: the ambiguity and uncertainty of decision boundaries.

[0003] Currently, existing technologies typically employ methods such as manual collection and rule synthesis to acquire false wake-up data, aiming to minimize the false wake-up rate of wake-up models. However, these existing technologies essentially involve passive and sparse sampling within an infinite "adversarial space," resulting in low efficiency and difficulty in acquiring high-quality false wake-up data. Therefore, improving the efficiency and quality of false wake-up data acquisition has become an urgent problem to be solved. Summary of the Invention

[0004] The main technical problem addressed by this application is to provide a method for acquiring false wake-up corpus, as well as related methods, apparatus, devices, and storage media, which can improve the acquisition efficiency and data quality of false wake-up corpus.

[0005] To address the aforementioned technical problems, the first aspect of this application provides a method for acquiring false wake-up corpus, comprising: constructing model instructions based on the product requirements of a wake-up model; wherein the product requirements include the target wake-up word of the wake-up model, and the model instructions are used to instruct a generative model to generate false wake-up text; performing speech synthesis on the false wake-up text output by the generative model in response to the model instructions to obtain false wake-up speech; obtaining a reward value generated by the generative model this time based on the response result of the wake-up model to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text; wherein the reward value is positively correlated with the probability that the false wake-up speech successfully falsely wakes up the wake-up model; adjusting the network parameters of the generative model based on the reward value, and determining whether to select the false wake-up text and false wake-up speech as sample corpus for training the wake-up model based on the reward value, and returning to the step of constructing model instructions based on the product requirements of the wake-up model, until the termination condition is met.

[0006] To address the aforementioned technical problems, a second aspect of this application provides a method for training a wake-up model, comprising: acquiring sample corpus; wherein the sample corpus is obtained through the method for acquiring false wake-up corpus described in the first aspect; and training the wake-up model based on the sample corpus.

[0007] To address the aforementioned technical problems, a third aspect of this application provides a voice wake-up method, comprising: acquiring voice data collected by a voice device; determining whether to wake up the voice device based on the response result of a wake-up model to the collected voice data; wherein the wake-up model is obtained based on the training method of the wake-up model described in the second aspect above.

[0008] To address the aforementioned technical problems, a fourth aspect of this application provides a device for acquiring false wake-up corpus, comprising: an instruction construction module, a speech synthesis module, a reward determination module, and an adjustment loop module. The instruction construction module is used to construct model instructions based on the product requirements of the wake-up model; wherein the product requirements include the target wake-up word of the wake-up model, and the model instructions are used to instruct the generative model to generate false wake-up text. The speech synthesis module is used to perform speech synthesis based on the false wake-up text output by the generative model in response to the model instructions, obtaining false wake-up speech. The reward determination module is used to obtain a reward value generated by the generative model this time based on the wake-up model's response to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text; wherein the reward value is positively correlated with the probability that the false wake-up speech successfully falsely wakes up the wake-up model. The adjustment loop module is used to adjust the network parameters of the generative model based on the reward value, and based on the reward value, determine whether to select the false wake-up text and false wake-up speech as sample corpus for training the wake-up model, and return to the step of constructing model instructions based on the product requirements of the wake-up model, until the termination condition is met.

[0009] To address the aforementioned technical problems, a fifth aspect of this application provides a training apparatus for a wake-up model, comprising: a corpus acquisition module and a model training module. The corpus acquisition module is used to acquire sample corpus; wherein the sample corpus is obtained through the false wake-up corpus acquisition apparatus of the fourth aspect described above. The model training module is used to train the wake-up model based on the sample corpus.

[0010] To address the aforementioned technical problems, a sixth aspect of this application provides a voice wake-up device, comprising: a voice acquisition module and a wake-up determination module. The voice acquisition module is used to acquire voice data collected by a voice device; the wake-up determination module is used to determine whether to wake up the voice device based on the response result of a wake-up model to the acquired voice data; wherein the wake-up model is obtained based on the training device for the wake-up model described in the fifth aspect above.

[0011] To address the aforementioned technical problems, the seventh aspect of this application provides an electronic device comprising at least a memory and a processor coupled to each other. The memory stores at least program instructions, and the processor executes the program instructions to implement the false wake-up corpus acquisition method in the first aspect, or the wake-up model training method in the second aspect, or the voice wake-up method in the third aspect.

[0012] To address the aforementioned technical problems, the eighth aspect of this application provides a computer-readable storage medium storing program instructions executable by a processor. The program instructions are used to implement the method for acquiring false wake-up corpus of the first aspect, or to implement the method for training the wake-up model in the second aspect, or to implement the voice wake-up method in the third aspect.

[0013] The above scheme, based on the product requirements of the wake-up model, constructs model instructions. The product requirements include the target wake-up word for the wake-up model. The model instructions instruct the generative model to generate false wake-up text. Then, based on the false wake-up text output by the generative model in response to the model instructions, speech synthesis is performed to obtain false wake-up speech. Based on the wake-up model's response to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, the reward value generated by the generative model is obtained. The reward value is positively correlated with the probability that the false wake-up speech successfully falsely wakes up the wake-up model. Based on the reward value, the network parameters of the generative model are adjusted, and based on the reward value, it is determined whether to select the false wake-up text and false wake-up speech as sample data for training the wake-up model. The process then returns to the steps of constructing model instructions based on the product requirements of the wake-up model, until the termination condition is met. This is because the false wake-up data is mainly actively generated by the generative model. Compared to passive methods such as manual collection and rule synthesis, this generative approach improves the efficiency of acquiring false wake-up data. Furthermore, by combining the wake-up model's response to false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, a reward value is obtained for the generative model's current generation. This reward value is then used to adjust the network parameters of the generative model. This creates a real-time closed-loop feedback between generation and sample evaluation through reward values. The generation strategy is dynamically adjusted based on the weaknesses of the wake-up model, forcing it to generate content that yields higher rewards. The iterative process also helps to continuously uncover the decision boundary defects of the wake-up model, leading to the generation of more challenging false wake-up data. This avoids sample redundancy caused by blind generation, making the false wake-up data more targeted and effective, thus improving its data quality. Therefore, this approach improves both the efficiency and quality of acquiring false wake-up data. Attached Figure Description

[0014] Figure 1 This is a flowchart illustrating an embodiment of the method for obtaining false wake-up corpus of this application;

[0015] Figure 2 This is a schematic diagram of a process of an embodiment of the method for obtaining false wake-up corpus of this application;

[0016] Figure 3 This is a flowchart illustrating an embodiment of the training method for the wake-up model in this application;

[0017] Figure 4 This is a schematic diagram illustrating an embodiment of the training method for the wake-up model in this application;

[0018] Figure 5 This is a flowchart illustrating an embodiment of the voice wake-up method of this application;

[0019] Figure 6 This is a schematic diagram of the framework of an embodiment of the false wake-up corpus acquisition device of this application;

[0020] Figure 7 This is a schematic diagram of the framework of an embodiment of the training device for the wake-up model of this application;

[0021] Figure 8 This is a schematic diagram of the framework of an embodiment of the voice wake-up device of this application;

[0022] Figure 9 This is a schematic diagram of the framework of an embodiment of the electronic device of this application;

[0023] Figure 10 This is a schematic diagram of a framework of an embodiment of the computer-readable storage medium of this application. Detailed Implementation

[0024] The embodiments of this application will now be described in detail with reference to the accompanying drawings.

[0025] In the following description, specific details such as particular system architectures, interfaces, and technologies are presented for illustrative purposes rather than for limiting purposes, in order to provide a thorough understanding of this application.

[0026] In this paper, the terms "system" and "network" are often used interchangeably. The term "and / or" describes the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A alone, A and B simultaneously, or B alone. Additionally, the slash " / " generally indicates that the preceding and following related objects have an "or" relationship. Furthermore, "many" in this paper indicates two or more objects.

[0027] Please see Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the method for obtaining false wake-up corpus of this application. It should be noted that the process operations in this embodiment can be executed by an electronic device with computing capabilities or by related equipment including such an electronic device. The specific structure of the electronic device or related equipment including the electronic device is not limited here. Specifically, this embodiment may include the following steps:

[0028] Step S11: Based on the product requirements of the wake-up model, construct model instructions.

[0029] In this embodiment of the disclosure, the product requirements may include the target wake-up word for the wake-up model, and the model instructions may be used to instruct the generative model to generate false wake-up text. It should be noted that the wake-up model can successfully wake up in response to the target wake-up word (i.e., it should wake up and is indeed woken up), while false wake-up text is the text that induces the wake-up model to wake up incorrectly (i.e., it should not wake up but is woken up).

[0030] In an implementation scenario, product requirements may also include at least one of the following: several interfering words with pronunciation similar to the target wake word, and several application scenarios for the wake-up model. For example, interfering words may have the same or similar initials and / or finals as the target wake word. For instance, taking the target wake word "Xiao Yi" as an example, interfering words may include, but are not limited to, "yili" (benefit), "xiaoyi" (aunt), "xiaoyi" (smile), etc. Furthermore, several application scenarios for the wake-up model may include, but are not limited to: television scenarios, mobile assistant scenarios, and story / animation scenarios that children's products may be interested in. Other application scenarios are not limited here, nor will they be listed one by one. In addition, product requirements may also include the range of pronunciation variations for the target wake word (e.g., for the target wake word "Xiao Fei Xiao Fei," whether to guarantee a wake-up rate for sounds like "Xiao Fei Xiao Fei" or "Xiao Fei Xiao Fei"), the acceptable range of elision (e.g., for the target wake word "Xiao Fei Xiao Fei," whether to accept elision to the degree of "Xiao Fei Yao Fei" or even "Xiao Fei Fei"), and the range of false wake-up from similar words (i.e., listing acceptable and unacceptable similar words, such as whether to control false wake-ups for "Xiao Du Xiao Du").

[0031] In one implementation scenario, product requirements can be organized into structured prompts as model instructions for the generation process. For example, model instructions can specifically instruct the generative model to generate false wake-up text based on the application scenario where the wake-up model needs to be deployed (e.g., a "television scenario"). The false wake-up text does not contain the target wake-up word (e.g., "Xiao Yi") but contains interfering words with similar pronunciations (e.g., "Xiao Yi"). Taking the above example, the model instructions could include, but are not limited to, the following: "Please generate text corpus for falsely waking up the wake-up model according to the following requirements: Application scenario: Television scenario; Target wake-up word: Xiao Yi; Interfering words: Xiao Yi." Of course, the above example is merely one possible example of model instructions in practical applications; the specific textual expression of model instructions is not limited here, nor will all examples be provided.

[0032] In one implementation scenario, a generative large model may include, but is not limited to, open-source large models such as Llama; or, a generative large model may be obtained by fine-tuning the parameters of an open-source large model based on a specific corpus; or, a generative large model may be a custom large model. The specific source of the generative large model is not limited here, nor will it be listed one by one.

[0033] Step S12: Speech synthesis is performed on the false wake-up text output by the generative model response model instruction to obtain the false wake-up speech.

[0034] Specifically, after constructing the model instructions, these instructions can be input into the generative model, causing it to respond and output false wake-up text. Based on this, speech synthesis can be performed on the false wake-up text to obtain false wake-up speech. As a possible example, using the aforementioned model instructions, false wake-up text can be obtained, including but not limited to: “Auntie (below), my mom (please watch) is calling you (movie) to come home for dinner tomorrow (XXXXX)”, etc. It should be noted that the text in parentheses in the above example represents background text related to the false wake-up text. Using the aforementioned example, after performing speech synthesis on the false wake-up text (e.g., using a multimodal large model), a false wake-up speech with the foreground speech “Auntie, my mom is calling you to come home for dinner tomorrow” and the background speech “below, please watch movie XXXXX” can be obtained. Alternatively, as another possible example, the false wake-up text generated by the model instructions may not contain background text. In this case, after performing speech synthesis on the false wake-up text, noise addition and perturbation operations can be performed on the synthesized speech to obtain false wake-up speech. Of course, the examples above are merely a few possible instances of mis-wake-up text and speech in practical applications. The specific content of the mis-wake-up text and speech is not limited here, nor will it be listed in detail. Furthermore, speech synthesis can be achieved through TTS (Text to Speech) systems, etc., and the specific methods of speech synthesis are not limited here.

[0035] As a possible example in practical applications, to improve sample diversity and realism, after obtaining the false wake-up speech through speech synthesis, enhancement can be performed on the false wake-up speech (e.g., introducing acoustic disturbances such as noise and reverberation) to simulate the acoustic conditions in the real environment as closely as possible. Of course, the above enhancement operations are not necessary in practical applications, but can be selectively configured according to the needs of the actual application. For example, in practical applications, enhancement operations can be implemented using noise addition, time-domain masking, and frequency-domain masking. Please refer to Table 1, which is a schematic table of one embodiment of enhancement operations. As shown in Table 1, different enhancement operations can employ different enhancement breadths, and under different enhancement breadths, the content invariance rate (i.e., the ratio of the false wake-up speech after the enhancement operation to the original false wake-up speech that has not changed in speech content) will also change, and correspondingly, the augmentation type (i.e., strong augmentation and weak augmentation) will also be different.

[0036] Table 1. Schematic diagram of an embodiment of enhanced operation

[0037]

[0038] It should be noted that ms in Table 1 represents the time unit millisecond, and the default unit for the augmentation strength of the frequency domain mask is Hertz (Hz). Of course, Table 1 only shows a few possible scenarios of augmentation operations in practical applications. Other possible scenarios of augmentation operations are not limited here, nor will they be listed one by one.

[0039] Step S13: Based on the wake-up model's response to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, obtain the reward value generated by the generative model this time.

[0040] In this embodiment, the reward value is positively correlated with the probability that a false wake-up voice message successfully wakes up the wake-up model. That is, the higher the probability of a false wake-up voice message successfully wakes up the wake-up model, the larger the reward value, and vice versa. For example, the reward value can be minimized when a false wake-up voice message successfully wakes up the wake-up model, while the reward value can be maximized when a false wake-up voice message successfully wakes up the wake-up model.

[0041] In one implementation scenario, the wake-up model can be designed based on deep neural networks, such as, but not limited to, convolutional neural networks, recurrent neural networks, and Transformers. The network structure of the wake-up model is not limited here. Furthermore, the wake-up model can also be implemented based on dynamic time warping, Gaussian mixture models-hidden Markov models, etc. Implementation methods for wake-up models will not be listed here.

[0042] In one implementation scenario, the wake-up model's response to a false wake-up voice can include either a successful false wake-up or a failed false wake-up. As a possible example, after the false wake-up voice is input to the wake-up model, the model can output a wake-up score. This wake-up score can represent the probability, as predicted by the wake-up model, that the false wake-up voice contains the target wake-up word. For example, a higher wake-up score indicates a higher probability, and vice versa. Furthermore, the wake-up score can range from 0 to 1 (in this case, the wake-up score can be considered the predicted probability that the false wake-up voice contains the target wake-up word). Of course, the above example is merely one possible illustration of the wake-up score; the specific value and range of the wake-up score are not limited, and further examples will not be provided. Based on this, if the wake-up score meets the relevant conditions regarding the wake-up score threshold (e.g., the wake-up score is not lower than the wake-up score threshold, or the wake-up score is higher than the wake-up score threshold), the response result can be determined to include successful false wake-up; conversely, if the wake-up score does not meet the wake-up score threshold, the response result can be determined to include failed false wake-up. It should be noted that although the response result is interpreted as successful or failed false wake-up, it does not mean that the false wake-up speech is necessarily a successful or failed false wake-up of the wake-up model. In practical applications, false wake-up text and its speech have a certain probability of containing the target wake-up word. If the false wake-up speech successfully wakes the wake-up model, it is a normal wake-up; conversely, if the false wake-up speech fails to wake the wake-up model, it is an abnormal wake-up. However, considering that in practical applications, as the performance of generative models gradually improves, the probability of false wake-up text and its speech containing the target wake-up word will be suppressed. Therefore, in most cases, false wake-up text and its speech do not contain the target wake-up word. To align with collecting false wake-up data for the model, the response results in the above two scenarios are uniformly referred to as successful or failed false wake-up.

[0043] In one implementation scenario, the text to be recognized from a false wake-up speech can be obtained using neural network models such as convolutional neural networks, recurrent neural networks, and long short-term memory networks, or it can be recognized using methods such as hidden Markov models. The specific methods for implementing speech recognition are not limited here, nor will they be listed in detail.

[0044] In one implementation scenario, to obtain the reward value for the generative model's current generation task, it's possible to detect whether the response result indicates a successful false wake-up and whether the identified text contains the target wake-up word. Based on this, for a response result indicating a successful false wake-up and the identified text not containing the target wake-up word, an upper reward limit can be selected as the reward value. For a response result indicating a failed false wake-up and the identified text not containing the target wake-up word, the reward value can be obtained based on the phoneme similarity between the phoneme sequence and the target wake-up word, and the wake-up score of the model triggered by the false wake-up speech. For a response indicating the identified text contains the target wake-up word, a lower reward limit can be selected as the reward value. It should be noted that the upper reward limit is a positive value, and the lower reward limit is a negative value. This approach, by assigning different reward values ​​to different cases, guides the generative model to generate corpora that can successfully trigger false wake-ups without containing the target wake-up word, and minimizes the inclusion of the target wake-up word or related interfering words in the generated false wake-up text.

[0045] In a specific implementation scenario, as mentioned earlier, when the response result indicates a successful false wake-up and the identified text does not contain the target wake-up word, the reward cap can be selected as the reward value. It should be noted that if the model is woken up even when the identified text does not contain the target wake-up word, this can be considered a true "false wake-up" as described above. Therefore, the generative model can be given the highest reward, i.e., the reward cap, to encourage it to continue generating content in this direction. Furthermore, the reward cap can be set according to the actual application needs. For example, reward caps and lower limits can be preset, and the reward value can be constrained to the range formed by the reward cap and lower limit. For example, if the reward value range is -1 to 1, the reward lower limit can be -1, and the reward upper limit can be 1. Of course, the above example is only one possible example of the reward cap and lower limit; the specific values ​​of the reward cap and lower limit are not limited here, nor will they be listed in detail.

[0046] In a specific implementation scenario, as mentioned earlier, when the identified text contains the target wake-up word, a reward lower bound can be chosen as the reward value. It should be noted that regardless of whether the wake-up model successfully triggers a false wake-up, it is undesirable for the identified text to contain the target wake-up word, as this indicates a high probability that the generative model did not correctly generate the false wake-up text according to the model's instructions (i.e., the generative model did not correctly understand the model's instructions). Of course, this situation could also occur for other reasons (such as changes in the speech content due to the aforementioned augmentation operation on the false wake-up speech). In this case, regardless of the reason why the identified text contains the target wake-up word, the generative model can be given a minimum reward, i.e., a reward lower bound, to penalize the generative model to avoid generating content in this direction in the future.

[0047] In a specific implementation scenario, in practical applications, besides recognizing text containing the target wake word, other undesirable situations may occur. For example, if the phoneme sequence contains the complete initial consonant sequence and / or final vowel sequence of the target wake word, a lower reward limit can be chosen as the reward value to penalize the generative model. Another example is if the similarity between the phoneme sequence and the target wake word's pronunciation sequence in the elision state meets similarity conditions (e.g., similarity higher than a similarity threshold, similarity not lower than a similarity threshold, etc.). In this case, a negative value between the lower and upper reward limits can be chosen as the reward value to penalize the generative model. Of course, the above examples are merely a few possible examples of other undesirable situations in practical applications; other possible scenarios are not limited here, nor will they be listed one by one. It should be noted that the priorities of various undesirable situations can also be set. When multiple undesirable situations occur in practical applications, the least desirable situation can be selected based on its priority, and the reward value can be determined according to the reward value calculation method under this situation. Of course, the above examples are only one possible way to determine the reward value when multiple undesirable situations occur in practical applications; settings can also be made according to product requirements in practical applications, and will not be listed one by one here. The above approach selects a lower reward limit as the reward value in response to the phoneme sequence containing a complete initial consonant sequence and / or a complete final vowel sequence of the target wake word. In response to the similarity between the phoneme sequence and the pronunciation sequence of the target wake word in the elision state satisfying the similarity condition, a negative value between the lower reward limit and the upper reward limit is selected as the reward value. This approach can provide different negative incentives based on the depth of involvement when the phoneme sequence involves the target wake word, and can force the generative model to stop generating content in this direction as much as possible.

[0048] In a specific implementation scenario, as mentioned earlier, when the response result indicates a false wake-up failure and the recognized text does not contain the target wake-up word, it can be considered that although the false wake-up speech did not contain the target wake-up word as desired, it did not achieve the expected goal of triggering a false wake-up for the wake-up model. That is, the false wake-up speech is insufficient to "induce" the wake-up model. In this case, a reward cap cannot be simply given to the generative model. Instead, a more targeted reward value needs to be obtained based on the phoneme similarity between the phoneme sequence and the target wake-up word, and the wake-up score value triggered by the false wake-up speech. Specifically, the reward value can be obtained by fusing phoneme similarity and wake-up score values ​​(e.g., averaging, weighting, etc.). It should be noted that the reward value can be positively correlated with both phoneme similarity and wake-up score value. That is, the higher the phoneme similarity and the higher the wake-up score value, the larger the reward value; conversely, the lower the phoneme similarity and the lower the wake-up score value, the smaller the reward value. Furthermore, as mentioned earlier, the wake-up model triggers wake-up when the wake-up score value meets the triggering condition regarding the wake-up score threshold. For details, please refer to the aforementioned descriptions, which will not be repeated here. The above method fuses phoneme similarity and wake-up score to obtain a reward value, which is positively correlated with both phoneme similarity and wake-up score. When the wake-up score meets the triggering condition of the wake-up score threshold, the wake-up model triggers wake-up, which can encourage generative models to generate false wake-up text that is as similar as possible to the pronunciation of the target wake-up word and can "induce" false wake-up.

[0049] Step S14: Based on the reward value, adjust the network parameters of the generative model, and based on the reward value, determine whether to select the false wake-up text and false wake-up speech as sample corpus for training the wake-up model, and return the product requirements based on the wake-up model to construct model instructions, until the termination condition is met.

[0050] In one implementation scenario, as a possible approach, generative models can be trained using reinforcement learning based on reward values ​​to incentivize them to generate the most effective negative examples possible (i.e., false wake-up corpora that do not contain the target wake-up word but can successfully trigger false wake-ups). Specifically, parameters can be adjusted based on the reward value using Proximal Policy Optimization (PPO). That is, the aforementioned model instruction can be considered as state s, the aforementioned false wake-up text as action a, and the aforementioned reward value A, combined with Proximal Policy Optimization, to perform gradient updates on the adjustable parameters (e.g., LoRA, PEFT modules) in the generative model. It should be noted that PEFT (Parameter-Efficient Fine-Tuning) is a technical framework that adapts to downstream tasks by fine-tuning only a small number of parameters while freezing most of the model's parameters; LoRA (Low-Rank Adaptation) is inspired by the fact that the rank of the parameter update matrix is ​​very low after large model pre-training. For the specific implementation processes of both, please refer to their technical details, which will not be elaborated here. Figure 2 , Figure 2 This is a schematic diagram illustrating the process of one embodiment of the method for obtaining false wake-up corpus of this application. Figure 2 As shown, after the generative model generates false wake-up text under the guidance of model instructions, it can perform speech synthesis based on the false wake-up text to obtain false wake-up speech. The false wake-up speech can be augmented through operations such as noise addition, temporal masking, frequency masking, speech rate perturbation, and reverberation simulation. Combined with the recognized text of the false wake-up speech after speech recognition, the response result of the false wake-up speech after being input into the wake-up model, and the phoneme sequence of the false wake-up text, the reward value of the generative model is obtained. Based on the reward value of the generative model, the generative model is optimized and trained through optimization methods such as PPO. In this way, the network parameters of the generative model can be continuously adjusted, forcing the generative model to gradually tend towards high-reward behavior, that is, to generate more effective false wake-up text.

[0051] In one implementation scenario, after obtaining the reward value, it can be determined whether to select the false wake-up text and its false wake-up speech as sample corpus for training the wake-up model. As mentioned earlier, the upper limit of the reward is a positive value and the lower limit of the reward is a negative value. In this case, false wake-up text and its false wake-up speech with positive rewards (i.e., reward values ​​are integer values) can be selected as sample corpus.

[0052] In one implementation scenario, the steps of adjusting the parameters of the generative model based on the reward value and determining whether to select the false wake-up text and its false wake-up speech as sample corpus based on the reward value can be executed in parallel, or the former can be executed before the latter, or the latter can be executed before the former. No limitation is made here.

[0053] In one implementation scenario, as one possible approach, before returning to the product requirements based on the wake-up model and constructing model instructions, the wake-up model can be trained on sample corpora; that is, the wake-up model can be trained synchronously with the generative model. Alternatively, as another possible approach, when the sample corpora have accumulated to a target number (e.g., 1000, 5000, etc.), the wake-up model can then be trained on the sample corpora; that is, the wake-up model can be trained later than the generative model. It should be noted that the above examples are merely two possible training methods in practical applications. The specific method can be chosen adaptively based on the actual situation (e.g., the former can be chosen when computing resources are relatively abundant; the latter can be chosen when computing resources are not abundant). No further limitations are imposed here, nor will further examples be provided.

[0054] In one implementation scenario, as a possible approach, the termination condition can be pre-set according to the actual application. For example, the termination condition can be set to the number of iterations of model training reaching a threshold (e.g., 5000, 10000, etc.). Or, for another example, the termination condition can be set to the reward value converging to a stable interval. Of course, the above examples are just a few possible ways to set the termination condition in practical applications, and the specific setting of the termination condition is not limited here (e.g., in the case of synchronous training of the wake-up model and the generative model, the termination condition can also be set to the model performance of the wake-up model meeting a predetermined index), and will not be listed one by one here.

[0055] In another implementation scenario, as a possible alternative, the termination condition can be set based on the dynamic changes of the accumulated sample corpus during the iterative process, unlike the aforementioned implementation. For example, in each iteration, clustering can be performed based on the various corpus features of the accumulated sample corpus to obtain several cluster sets. Then, based on the corpus features of the currently acquired sample corpus, it can be determined whether the currently acquired sample corpus can be assigned to any cluster set. It should be noted that the termination condition can be set as follows: a consecutive target number of times (e.g., 5 times, 10 times, 20 times, etc.) fails to generate sample corpus that cannot be assigned to any cluster set. If the corpus features of the currently acquired sample corpus cannot be assigned to any cluster set, it indicates that the currently acquired sample corpus does not belong to any cluster set in the feature space. That is, the key information for achieving false wake-up in the currently acquired sample corpus differs from the accumulated sample corpus. Therefore, when a consecutive target number of times fails to generate sample corpus that cannot be assigned to any cluster set, it can be considered that the key information for achieving false wake-up has been sufficiently explored in the feature space.

[0056] In a specific implementation scenario, as one possible example, semantic features can be extracted from the false wake-up text in the sample corpus, and these semantic features can be used as the corpus features of the sample corpus. Alternatively, as another possible example, features can be extracted from the false wake-up speech in the sample corpus, and these speech features can be used as the corpus features of the sample corpus. Or, semantic features can be extracted from the false wake-up text in the sample corpus, and then features can be extracted from the false wake-up speech in the sample corpus, followed by a fusion of the semantic and speech features to obtain the corpus features of the sample corpus. Of course, the above examples are merely a few possible examples of extracting corpus features in practical applications; the methods for extracting corpus features are not limited here, nor will they be listed in detail.

[0057] In a specific implementation scenario, clustering can be performed after extracting the corpus features of the accumulated sample corpora. Specifically, clustering algorithms such as K-Means and DBSCAN (Density-Based Spatial Clustering of Applications with Noise) can be used. Here, the specific method of implementing feature clustering is not limited, nor will it be listed one by one.

[0058] In a specific implementation scenario, after clustering the corpus features of each accumulated sample corpus to obtain several cluster sets, it can be determined whether the currently acquired sample corpus can be assigned to any cluster set based on its corpus features. For example, clustering algorithms such as K-Means and DBSCAN can be used to implement feature clustering, and the assignment of the currently acquired sample corpus to any cluster set can be determined based on its corpus features. Alternatively, for another example, the feature distance between the corpus features of the currently acquired sample corpus and the central features of each cluster set can be calculated. Based on this, if there is a feature distance that satisfies the filtering condition of a distance threshold (e.g., feature distance is below the distance threshold, feature distance is not longer than the distance threshold), it can be determined that the currently acquired sample corpus can be assigned to the corresponding cluster set. Conversely, if there is no feature distance that satisfies the filtering condition of a distance threshold, it can be determined that the currently acquired sample corpus cannot be assigned to any cluster set. Furthermore, it should be noted that if the feature distances between the central features of multiple cluster sets and the corpus features of the sample corpus obtained in this study all meet the aforementioned screening conditions, then it can be determined that the sample corpus obtained in this study can be assigned to any of these cluster sets. For example, the sample corpus obtained in this study can be assigned to the cluster set with the smallest feature distance.

[0059] In a specific implementation scenario, if the acquired sample corpus cannot be assigned to any cluster set, a new cluster set can be created based on the acquired sample corpus. That is, a new cluster set can be created, and the acquired sample corpus (and / or its corpus features) can be placed into the new cluster set. In addition, interference words that replace the target wake-up word and trigger false wake-ups can be extracted from the acquired sample corpus and used as target interference words. For example, still using the target wake-up word "Xiao Yi" as an example, if "smile" replaces the target wake-up word and triggers false wake-ups in the acquired sample corpus, this can be used as the target interference word. Of course, the above example is only one possible example in practical applications; the specific content of the target interference words is not limited here, nor will all examples be provided. Based on this, when returning to the step of constructing model instructions based on the product requirements of the wake-up model, new model instructions can be constructed based on the product requirements and the target interference words. Taking the previous example, the model instruction can be adjusted to include, but is not limited to, the following: "Please generate text corpus for falsely waking up the wake-up model according to the following requirements. Application scenario: TV scenario; target wake-up word: Xiao Yi; interference words: Xiao Yi, smile." This allows us to obtain the false wake-up text output by the generative model in response to the model instruction, such as, but not limited to, "Xiao Yi (next) greets relatives with a smile (please watch the movie) (XXXX)" (the specific meaning of the parentheses and their contents can be found in the aforementioned explanation, and will not be repeated here). The specific content of the false wake-up text is not limited here. In response to the fact that the sample corpus obtained this time cannot be assigned to any cluster set, a new cluster set is created based on the sample corpus obtained this time, and interference words that replace the target wake-up word to trigger false wake-ups are extracted from the sample corpus obtained this time as target interference words. Furthermore, when returning to the step of constructing model instructions based on the product requirements of the wake-up model, new model instructions are constructed based on the product requirements and target interference words. This allows the generative model to be guided in its generation direction during subsequent generation processes based on the newly generated target interference words.

[0060] In a specific implementation scenario, to guide the generative model's generation direction, the feature distances between the acquired sample corpus and each cluster can be obtained based on the corpus features and the central features of each cluster (e.g., measured using cosine similarity, Euclidean distance, etc.). The reward value is then updated based on the minimum feature distance to obtain a new reward value. It should be noted that the generative model can adjust its parameters based on this new reward value. That is, after obtaining the reward value generated by the generative model based on the wake-up model's response to false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, the reward value can be used to determine whether to select the false wake-up text and false wake-up speech as sample corpus for training the wake-up model. However, the network parameters of the generative model are not adjusted based on the reward value at this stage. Instead, the reward value is first updated based on the minimum feature distance to obtain a new reward value, and then the network parameters of the generative model are adjusted based on the new reward value (the specific method can be found in the aforementioned description, which will not be repeated here). Furthermore, the update operation performed on the reward value based on the minimum feature distance can specifically be a forward update. For example, the magnitude of the minimum feature distance can be positively correlated with the update magnitude of the reward value. That is, the larger the minimum feature distance, the larger the update magnitude of the reward value (e.g., by adding a large positive value), and conversely, the smaller the minimum feature distance, the smaller the update magnitude of the reward value (e.g., by adding a small positive value). The above method, based on the corpus features of the acquired sample corpus and the central features of each cluster set, obtains the feature distances between the acquired sample corpus and each cluster set. The reward value is then updated based on the minimum feature distance to obtain a new reward value. This method, by incorporating the minimum feature distance to adjust the reward value, encourages the generative model to fully explore the feature space to achieve false wake-up interference words.

[0061] The above scheme, based on the product requirements of the wake-up model, constructs model instructions. The product requirements include the target wake-up word for the wake-up model. The model instructions instruct the generative model to generate false wake-up text. Then, based on the false wake-up text output by the generative model in response to the model instructions, speech synthesis is performed to obtain false wake-up speech. Based on the wake-up model's response to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, the reward value generated by the generative model for this time is obtained. Then, based on the reward value, the network parameters of the generative model are adjusted, and based on the reward value, it is determined whether to select the false wake-up text and false wake-up speech as sample data for training the wake-up model. The process of returning to the product requirements of the wake-up model and constructing model instructions continues until the termination condition is met. This approach is advantageous because the false wake-up data is mainly obtained actively by the generative model, compared to manual collection and standardized processing. In contrast to passive methods like synthesis, this approach can improve the efficiency of acquiring false wake-up data. Furthermore, by combining the wake-up model's response to false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, the reward value generated by the generative model is obtained. This reward value is then used to adjust the network parameters of the generative model. This creates a real-time closed-loop feedback between generation and sample evaluation through reward values. This allows for dynamic adjustments to the generation strategy based on the weaknesses of the wake-up model, forcing it to generate content that yields higher rewards. The iterative process also helps to continuously uncover the decision boundary defects of the wake-up model, leading to the generation of more challenging false wake-up data. This avoids sample redundancy caused by blind generation, making the false wake-up data more targeted and effective, thus improving its data quality. Therefore, this approach improves both the efficiency of acquiring false wake-up data and its data quality.

[0062] Please see Figure 3 , Figure 3 This is a flowchart illustrating an embodiment of the training method for the wake-up model of this application. It should be noted that the process operations in this embodiment can be executed by an electronic device with computing capabilities or by related equipment containing such an electronic device. The specific structure of the electronic device or related equipment containing the electronic device is not limited herein. Specifically, this embodiment may include the following steps:

[0063] Step S31: Obtain sample corpus.

[0064] In this embodiment of the disclosure, the sample corpus can be obtained through the process steps in the above-described method for obtaining false wake-up corpus, as detailed in the foregoing embodiments, which will not be repeated here. As one possible example, during the training of the wake-up model, the sample corpus can be obtained solely through the process steps in the aforementioned method for obtaining false wake-up corpus, and the wake-up model can be trained based on this sample corpus. As another possible example, during the training of the wake-up model, in addition to obtaining the sample corpus through the process steps in the aforementioned method for obtaining false wake-up corpus, supervised sample corpus collected from the real environment can also be collected, and the two can be combined to train the wake-up model. As yet another possible example, during the training of the wake-up model, in addition to obtaining the sample corpus through the process steps in the aforementioned method for obtaining false wake-up corpus, unsupervised sample corpus collected from the real environment can also be collected, and the two can be combined to train the wake-up model. As one possible example, during the training of the wake-up model, in addition to obtaining sample data through the steps in the aforementioned method embodiment for acquiring false wake-up data, supervised sample data collected from the real environment and unsupervised sample data collected from the real environment can also be collected to combine the three for training the wake-up model. It should be noted that the aforementioned supervised and unsupervised sample data can include false wake-up data as negative samples (i.e., data without the target wake-up word and used to "induce" the wake-up model to falsely wake up) and positive samples (i.e., data containing the target wake-up word and used to normally wake up the wake-up model). Furthermore, supervised sample data can be annotated with the real text corresponding to the speech data, while unsupervised sample data may not be annotated with the real text corresponding to the speech data, meaning it is impossible to determine which are positive samples and which are negative samples. Of course, the above examples are merely a few possible examples of sample data in practical applications; the specific situation of the sample data is not limited here, nor will it be listed in detail.

[0065] Step S32: Train the wake-up model based on the sample corpus.

[0066] In one implementation scenario, as a possible approach, the training loss can be obtained by measuring the predicted wake-up score of the wake-up model on the sample corpus and whether the sample corpus contains the target wake-up word. The network parameters of the wake-up model can then be adjusted based on this training loss. For example, the specific meaning of the predicted wake-up score can be found in the description of the wake-up score value in the aforementioned disclosed embodiments, and will not be repeated here. Taking the predicted wake-up score as a predicted wake-up probability with a numerical range of 0 to 1 as an example, if the sample corpus contains the target wake-up word, the ground truth of the sample corpus can be considered as 1; conversely, if the sample corpus does not contain the target wake-up word, the ground truth of the sample corpus can be considered as 0. Based on this, a loss function such as cross-entropy can be used to calculate the loss on the predicted wake-up probability and the ground truth, thus obtaining the training loss. Of course, the above example is merely one possible example of training the wake-up model in practical applications; other possible loss measurement methods are not limited here, nor will they be listed one by one.

[0067] In another implementation scenario, as an alternative possible implementation method, distinct from the aforementioned implementation methods, please refer to the relevant documentation. Figure 4 , Figure 4 This is a schematic diagram illustrating an embodiment of the training method for the wake-up model of this application. For example... Figure 4As shown, multiple losses can be combined to jointly measure the training loss of the wake-up model. For example, after augmentation, the training loss can include consistency regularization loss. It should be noted that the consistency regularization loss can be obtained by measuring the mean squared error between the feature vectors of the speech data before and after augmentation. Another example is that the training loss can also include multi-scale pseudo-label classification loss. It should be noted that the data for pseudo-label classification can include three categories: supervised data after augmentation (see the aforementioned supervised sample corpus for details), all unsupervised data (see the aforementioned unsupervised sample corpus for details), and all false wake-up data obtained through the aforementioned disclosed embodiments. Furthermore, the multi-scale aspect is reflected in the frame-level pseudo-labels and word-level pseudo-labels. Frame-level pseudo-labels represent the pronunciation information of the speech frame (such as phonemes), and word-level pseudo-labels indicate whether a word belongs to the target wake-up word. The multi-scale pseudo-label classification loss can be obtained by measuring the cross-entropy loss based on the multi-level pseudo-labels. Yet another example is that the training loss can also include supervised training loss. It should be noted that the supervised training loss can be obtained by measuring cross-entropy loss (i.e., the difference between the predicted text and the real text of the speech data) based on supervised data that has undergone weak augmentation (see Table 1 in the aforementioned public embodiments) or no augmentation. For example, the training loss can also include domain adaptability loss. It should be noted that domain adaptability loss is mainly used for gradient adversarial training between real-world collected data (which can be called non-synthetic data) and false wake-up corpora obtained through the aforementioned public embodiments (which can be called synthetic data). Gradient adversarial training uses a gradient inversion layer. Specifically, a new binary classification task head for synthetic and non-synthetic data is added for cross-entropy training. A gradient inversion layer can be added at the connection between the task head and the backbone network, allowing features to be updated in the opposite direction of classification, thus preventing the model from learning the difference between the synthetic and non-synthetic domains and obtaining domain-invariant features. Furthermore, to enhance the training benefits of synthetic data, a portion of the synthetic data can be extracted for simulations such as noise addition, reverberation, and speed variation. The noise addition part will superimpose noise recorded in the real scene to further confuse the synthetic domain and the real domain (i.e., the non-synthetic domain). Of course, in practical applications, the various losses mentioned above can be combined as needed, and will not be listed one by one here.

[0068] The above scheme obtains sample corpus, which is obtained through the process steps in the above-described method for obtaining false wake-up corpus. Based on the sample corpus, the wake-up model is trained, which can continuously explore the decision boundary defects of the wake-up model and help to reduce the false wake-up rate of the wake-up model as much as possible.

[0069] Please see Figure 5 , Figure 5This is a flowchart illustrating an embodiment of the voice wake-up method of this application. It should be noted that the process operations in this embodiment can be executed by an electronic device with computing capabilities or by related equipment including such an electronic device. The specific structure of the electronic device or related equipment including the electronic device is not limited here. Specifically, this embodiment may include the following steps:

[0070] Step S51: Acquire the voice recordings from the voice device.

[0071] Specifically, voice devices can include, but are not limited to, smartphones, tablets, in-vehicle systems, and smart speakers. No specific type of voice device is limited here, nor will a list be provided. It should be noted that voice devices can have integrated or external microphones to acquire and collect voice data.

[0072] Step S52: Based on the response results of the wake-up model to the collected voice, determine whether to wake up the voice device.

[0073] In this embodiment of the disclosure, the wake-up model can be obtained based on the process steps in the above-described wake-up model training method embodiment, which can be referred to in the foregoing disclosed embodiments, and will not be repeated here. As a possible example, the collected voice can be input into the wake-up model to obtain a wake-up score value (as described in the foregoing disclosed embodiments, when the wake-up score value is constrained to a numerical range of 0 to 1, it can also be regarded as a predicted wake-up probability). At this time, the response result can be determined as triggering wake-up or not triggering wake-up based on whether the wake-up score value meets the relevant conditions regarding the wake-up threshold. For example, if the wake-up score value meets the relevant conditions (e.g., the wake-up score value is not lower than the wake-up threshold, or the wake-up score value is higher than the wake-up threshold), the response result can be determined as triggering wake-up, and the voice device can be determined to be woken up; conversely, if the wake-up score value does not meet the relevant conditions (e.g., the wake-up score value is lower than the wake-up threshold, or the wake-up score value is not higher than the wake-up threshold), the response result can be determined as not triggering wake-up, and the voice device can be determined not to be woken up.

[0074] The above solution acquires the voice recordings from the voice device, and then determines whether to wake up the voice device based on the response of the wake-up model to the voice recordings. The wake-up model is obtained based on the process steps in the above wake-up model training method embodiment, which can minimize the false wake-up rate of the voice device.

[0075] Please see Figure 6 , Figure 6This is a schematic diagram of the framework of an embodiment of the device for acquiring false wake-up corpus of this application. The device 60 for acquiring false wake-up corpus includes: an instruction construction module 61, a speech synthesis module 62, a reward determination module 63, and an adjustment loop module 64. The instruction construction module 61 is used to construct model instructions based on the product requirements of the wake-up model; wherein, the product requirements include the target wake-up word of the wake-up model, and the model instructions are used to instruct the generative model to generate false wake-up text. The speech synthesis module 62 is used to perform speech synthesis based on the false wake-up text output by the generative model in response to the model instructions to obtain false wake-up speech. The reward determination module 63 is used to obtain the reward value generated by the generative model this time based on the response result of the wake-up model to the false wake-up speech, the recognition text of the false wake-up speech, and the phoneme sequence of the false wake-up text; wherein, the reward value is positively correlated with the probability that the false wake-up speech successfully falsely wakes up the wake-up model. The adjustment loop module 64 is used to adjust the network parameters of the generative model based on the reward value, and based on the reward value, determine whether to select the false wake-up text and the false wake-up speech as sample corpus for training the wake-up model, and return to the step of constructing model instructions based on the product requirements of the wake-up model, until the termination condition is met.

[0076] In the above scheme, the false wake-up corpus acquisition device 60 constructs model instructions based on the product requirements of the wake-up model. The product requirements include the target wake-up word of the wake-up model. The model instructions are used to instruct the generative model to generate false wake-up text. Then, based on the false wake-up text output by the generative model in response to the model instructions, speech synthesis is performed to obtain false wake-up speech. Based on the wake-up model's response to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, the reward value generated by the generative model this time is obtained. The reward value is positively correlated with the probability that the false wake-up speech successfully falsely wakes up the wake-up model. Then, based on the reward value, the network parameters of the generative model are adjusted, and based on the reward value, it is determined whether to select the false wake-up text and false wake-up speech as sample corpus for training the wake-up model, and the steps of constructing model instructions based on the product requirements of the wake-up model are returned until the termination condition is met. On the one hand, since the false wake-up corpus is mainly obtained through generative modeling... Generative models actively generate data, which, compared to "passive" methods such as manual collection and rule synthesis, improves the efficiency of acquiring false wake-up corpora. Furthermore, by combining the wake-up model's response to false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, a reward value is obtained for the generative model's current generation. This reward value is then used to adjust the network parameters of the generative model. This creates a real-time closed-loop feedback between generation and sample evaluation through reward values, dynamically adjusting the generation strategy based on the weaknesses of the wake-up model. This helps force the generative model to generate content that yields higher rewards. Moreover, the continuous iteration process helps to continuously uncover the decision boundary defects of the wake-up model, leading to the generation of more challenging false wake-up corpora. This avoids sample redundancy caused by blind generation, making the false wake-up corpora more targeted and effective, thus improving the data quality of the false wake-up corpora. Therefore, this method improves both the efficiency and quality of acquiring false wake-up corpora.

[0077] In some disclosed embodiments, the reward determination module 63 includes a content detection submodule for detecting whether the response result indicates a successful false wake-up and whether the identified text contains the target wake-up word; the reward determination module 63 includes a first response submodule for selecting an upper reward limit as the reward value in response to a response result indicating a successful false wake-up and the identified text not containing the target wake-up word; the reward determination module 63 includes a second response submodule for obtaining a reward value based on the phoneme similarity between the phoneme sequence and the target wake-up word and the wake-up score of the false wake-up speech-triggered wake-up model in response to a response result indicating a failed false wake-up and the identified text not containing the target wake-up word; and the reward determination module 63 includes a third response submodule for selecting a lower reward limit as the reward value in response to the identified text containing the target wake-up word; wherein, the upper reward limit is a positive value and the lower reward limit is a negative value.

[0078] In some disclosed embodiments, the reward determination module 63 includes a fourth response submodule, used to select a lower reward limit as a reward value in response to the phoneme sequence containing a complete initial consonant sequence and / or a complete final vowel sequence of the target wake word; the reward determination module 63 includes a fifth response submodule, used to select a negative value between the lower reward limit and the upper reward limit as a reward value in response to the similarity between the phoneme sequence and the pronunciation sequence of the target wake word in the elision state satisfying the similarity condition.

[0079] In some disclosed embodiments, the second response submodule is specifically used to fuse phoneme similarity and wake-up score to obtain a reward value; wherein the reward value is positively correlated with phoneme similarity and wake-up score, and the wake-up model triggers wake-up when the wake-up score meets the triggering condition regarding the wake-up score threshold.

[0080] In some disclosed embodiments, the device 60 for acquiring false wake-up corpus further includes a corpus clustering module, used to cluster based on the corpus features of each accumulated sample corpus to obtain several cluster sets; the device 60 for acquiring false wake-up corpus also includes an allocation determination module, used to determine whether the sample corpus acquired this time can be allocated to any cluster set based on the corpus features of the sample corpus acquired this time; wherein, the termination condition includes: the target number of consecutive attempts fails to generate sample corpus that cannot be allocated to any cluster set.

[0081] In some disclosed embodiments, the false wake-up corpus acquisition device 60 further includes a reconstruction module, which is used to create a new cluster set based on the acquired sample corpus in response to the fact that the acquired sample corpus cannot be assigned to any cluster set, and to extract interference words from the acquired sample corpus that replace the target wake-up word to trigger false wake-ups as target interference words, and to construct new model instructions based on product requirements and target interference words when returning to the step of constructing model instructions based on wake-up model product requirements.

[0082] In some disclosed embodiments, the device 60 for acquiring false wake-up corpus further includes a distance measurement module, used to obtain the feature distance between the acquired sample corpus and each cluster set based on the corpus features of the acquired sample corpus and the central features of each cluster set; the device 60 for acquiring false wake-up corpus also includes a reward update module, used to update the reward value based on the minimum feature distance to obtain a new reward value; wherein, the generative model adjusts its parameters based on the new reward value, and the magnitude of the minimum feature distance is positively correlated with the update magnitude of the reward value.

[0083] In some disclosed embodiments, before returning to the product requirements based on the wake-up model and constructing model instructions, the wake-up model is first trained based on sample corpora, or, when the sample corpora have accumulated to a target quantity, the wake-up model is then trained based on the sample corpora; and / or, the product requirements also include at least one of the following: several interfering words with pronunciations similar to the target wake-up word, several application scenarios for waking up the target; and / or, the model instructions are specifically used to instruct the generative model to generate false wake-up text based on the application scenarios that the wake-up model needs to deploy, the false wake-up text does not contain the target wake-up word but contains interfering words with pronunciations similar to the target wake-up word; and / or, the generative model adjusts its parameters based on reward values ​​through proximal policy optimization.

[0084] Please see Figure 7 , Figure 7 This is a schematic diagram of the framework of an embodiment of the training device for the wake-up model of this application. The training device 70 for the wake-up model includes a corpus acquisition module 71 and a model training module 72. The corpus acquisition module 71 is used to acquire sample corpus; wherein, the sample corpus is obtained through the false wake-up corpus acquisition device in the above embodiment of the false wake-up corpus acquisition device; the model training module 72 is used to train the wake-up model based on the sample corpus.

[0085] In the above scheme, the training device 70 of the wake-up model acquires sample corpus, and the sample corpus is obtained through the false wake-up corpus acquisition device in the above-mentioned false wake-up corpus acquisition device embodiment. Based on the sample corpus, the wake-up model is trained, which can continuously explore the decision boundary defects of the wake-up model and help to reduce the false wake-up rate of the wake-up model as much as possible.

[0086] Please see Figure 8 , Figure 8 This is a schematic diagram of the framework of an embodiment of the voice wake-up device of this application. The voice wake-up device 80 includes a voice acquisition module 81 and a wake-up determination module 82. The voice acquisition module 81 is used to acquire the voice collected by the voice device; the wake-up determination module 82 is used to determine whether to wake up the voice device based on the response result of the wake-up model to the acquired voice; wherein, the wake-up model is obtained based on the wake-up model training device in the above-described wake-up model training device embodiment.

[0087] In the above scheme, the voice wake-up device 80 acquires the voice collected by the voice device, and then determines whether to wake up the voice device based on the response result of the wake-up model to the voice collected. The wake-up model is obtained based on the training device of the wake-up model in the above-mentioned wake-up model training device embodiment, which can minimize the false wake-up rate of the voice device.

[0088] Please see Figure 9 , Figure 9This is a schematic diagram of an embodiment of the electronic device of this application. The electronic device 90 includes at least a memory 91 and a processor 92 coupled to each other. The memory 91 stores at least program instructions, and the processor 92 is used to execute the program instructions to implement the steps in any of the above-described embodiments of the false wake-up corpus acquisition method, or to implement the steps in any of the above-described embodiments of the wake-up model training method, or to implement the steps in any of the above-described embodiments of the voice wake-up method. For details, please refer to the foregoing disclosed embodiments, which will not be repeated here. It should be noted that, taking the implementation of the steps in any of the above-described embodiments of the false wake-up corpus acquisition method by the electronic device 90 as an example, the electronic device 90 may include, but is not limited to: a server, a microcomputer, etc.; or, taking the implementation of the steps in any of the above-described embodiments of the wake-up model training method by the electronic device 90 as an example, the electronic device 90 may include, but is not limited to: a server, a microcomputer, etc.; or, taking the implementation of the steps in any of the above-described embodiments of the voice wake-up method by the electronic device 90 as an example, the electronic device may include, but is not limited to: a smartphone, a tablet computer, etc. The specific type of the electronic device 90 is not limited here.

[0089] Specifically, processor 92 controls itself and memory 91 to implement the steps in any of the above-described embodiments of the false wake-up corpus acquisition method, or the steps in any of the above-described embodiments of the wake-up model training method, or the steps in any of the above-described embodiments of the voice wake-up method. Processor 92 can also be called a CPU (Central Processing Unit). Processor 92 may be an integrated circuit chip with signal processing capabilities. Processor 92 can also be a general-purpose processor, digital signal processor (DSP), application-specific integrated circuit (ASIC), field-programmable gate array (FPGA), or other programmable logic device, discrete gate or transistor logic device, or discrete hardware component. A general-purpose processor can be a microprocessor or any conventional processor. Furthermore, processor 92 can be implemented using integrated circuit chips.

[0090] In the above scheme, the electronic device 90 constructs model instructions based on the product requirements of the wake-up model. The product requirements include the target wake-up word of the wake-up model. The model instructions are used to instruct the generative model to generate false wake-up text. Then, based on the false wake-up text output by the generative model in response to the model instructions, speech synthesis is performed to obtain false wake-up speech. Based on the wake-up model's response to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, the reward value generated by the generative model is obtained. The reward value is positively correlated with the probability that the false wake-up speech successfully falsely wakes up the wake-up model. Based on the reward value, the network parameters of the generative model are adjusted, and based on the reward value, it is determined whether to select the false wake-up text and false wake-up speech as sample corpus for training the wake-up model. The process of constructing model instructions based on the product requirements of the wake-up model is then repeated until the termination condition is met. On the one hand, since the false wake-up corpus is mainly obtained through the generative model... Active generation of data, compared to "passive" methods such as manual collection and rule synthesis, can improve the efficiency of acquiring false wake-up corpora. Furthermore, by combining the wake-up model's response to false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, the reward value generated by the generative model is obtained. This is used to adjust the network parameters of the generative model, thus forming a real-time closed-loop feedback between generation and sample evaluation through reward values. This allows for dynamic adjustment of the generation strategy based on the weaknesses of the wake-up model, helping to force the generative model to generate content in a direction that yields higher rewards. In the continuous iterative process, it also helps to continuously uncover the decision boundary defects of the wake-up model, helping to generate more challenging false wake-up corpora, minimizing sample redundancy caused by blind generation, and making the false wake-up corpora more targeted and effective, thereby improving the data quality of the false wake-up corpora. Therefore, it improves the efficiency and quality of acquiring false wake-up corpora. In addition, it can also minimize the false wake-up rate.

[0091] Please see Figure 10 , Figure 10 This is a schematic diagram of a framework of an embodiment of the computer-readable storage medium of this application. The computer-readable storage medium 100 stores program instructions 101 that can be executed by a processor. The program instructions 101 are used to implement the steps in any of the above embodiments of the false wake-up corpus acquisition method, or to implement the steps in any of the above embodiments of the wake-up model training method, or to implement the steps in any of the above embodiments of the voice wake-up method.

[0092] The above scheme involves a computer-readable storage medium 100 constructing model instructions based on the product requirements of the wake-up model. The product requirements include the target wake-up word for the wake-up model. The model instructions instruct the generative model to generate false wake-up text. Then, based on the false wake-up text output by the generative model in response to the model instructions, speech synthesis is performed to obtain false wake-up speech. Based on the wake-up model's response to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, a reward value is obtained for the generative model's current generation. This reward value is positively correlated with the probability that the false wake-up speech successfully falsely wakes up the wake-up model. Based on the reward value, the network parameters of the generative model are adjusted, and based on the reward value, it is determined whether to select the false wake-up text and false wake-up speech as sample corpus for training the wake-up model. The process then returns to the steps of constructing model instructions based on the product requirements of the wake-up model, until the termination condition is met. This is because the false wake-up corpus is mainly generated through... Generative models actively generate data, which, compared to passive methods such as manual collection and rule synthesis, improves the efficiency of acquiring false wake-up corpora. Furthermore, by combining the wake-up model's response to false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, a reward value is obtained for the generative model's current generation. This reward value is then used to adjust the network parameters of the generative model. This creates a real-time closed-loop feedback between generation and sample evaluation through reward values, dynamically adjusting the generation strategy based on the weaknesses of the wake-up model. This helps force the generative model to generate content that yields higher rewards. Moreover, the continuous iteration process helps to continuously uncover the decision boundary defects of the wake-up model, leading to the generation of more challenging false wake-up corpora. This avoids sample redundancy caused by blind generation, making the false wake-up corpora more targeted and effective, thus improving the data quality. Therefore, this improves the efficiency and quality of acquiring false wake-up corpora. In addition, it can also minimize the false wake-up rate.

[0093] In some embodiments, the functions or modules of the apparatus provided in this disclosure can be used to perform the methods described in the above method embodiments. The specific implementation can be referred to the description of the above method embodiments, and for the sake of brevity, it will not be repeated here.

[0094] The description of the various embodiments above tends to emphasize the differences between the various embodiments. The similarities or similarities between them can be referred to, and for the sake of brevity, they will not be repeated here.

[0095] In the several embodiments provided in this application, it should be understood that the disclosed methods and apparatus can be implemented in other ways. For example, the apparatus implementations described above are merely illustrative. For instance, the division of modules or 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 devices or units may be electrical, mechanical, or other forms.

[0096] 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, depending on actual needs.

[0097] 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.

[0098] 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.) or processor to execute all or part of the steps of the methods of 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.

[0099] If the technical solution of this application involves personal information, the product using this technical solution has clearly informed the user of the personal information processing rules and obtained the user's voluntary consent before processing the personal information. If the technical solution of this application involves sensitive personal information, the product using this technical solution has obtained the user's separate consent before processing the sensitive personal information, and also meets the requirement of "express consent". For example, at personal information collection devices such as cameras, clear and prominent signs are set up to inform users that they have entered the scope of personal information collection and that personal information will be collected. If an individual voluntarily enters the collection scope, it is deemed that they have agreed to the collection of their personal information; or on the personal information processing device, with clear signs / information informing users of the personal information processing rules, authorization is obtained from the individual through pop-up information or by asking the individual to upload their personal information; wherein, the personal information processing rules may include information such as the personal information processor, the purpose of personal information processing, the processing method, and the types of personal information processed.

Claims

1. A method for acquiring false wake-up corpus, characterized in that, include: Based on the product requirements of the wake-up model, model instructions are constructed; wherein, the product requirements include the target wake-up word of the wake-up model, and the model instructions are used to instruct the generative model to generate false wake-up text; Based on the false wake-up text output by the generative model in response to the model instruction, speech synthesis is performed to obtain false wake-up speech; Based on the wake-up model's response to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text, the reward value generated by the generative model this time is obtained; wherein, the reward value is positively correlated with the probability that the false wake-up speech successfully falsely wakes up the wake-up model; Based on the reward value, the network parameters of the generative model are adjusted, and based on the reward value, it is determined whether to select the false wake-up text and the false wake-up speech as sample corpus for training the wake-up model, and the steps of returning the product requirements based on the wake-up model and constructing model instructions are completed until the termination condition is met.

2. The method according to claim 1, characterized in that, The reward value generated by the generative model for this instance is obtained based on the wake-up model's response to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text. This reward value includes: Detect whether the response result indicates a successful false wake-up and whether the identified text contains the target wake-up word; In response to the response result being characterized as a successful false wake-up and the recognized text not containing the target wake-up word, the upper limit of the reward is selected as the reward value; In response to the response result being characterized as a false wake-up failure and the recognized text not containing the target wake-up word, the reward value is obtained based on the phoneme similarity between the phoneme sequence and the target wake-up word and the wake-up score value of the wake-up model triggered by the false wake-up speech; In response to the recognition text containing the target wake word, a lower limit of the reward is selected as the reward value; wherein, the upper limit of the reward is a positive value and the lower limit of the reward is a negative value.

3. The method according to claim 2, characterized in that, The method further includes: In response to the phoneme sequence containing a complete initial consonant sequence and / or a complete final vowel sequence of the target wake word, the lower limit of the reward is selected as the reward value; In response to the similarity condition between the phoneme sequence and the pronunciation sequence of the target wake word in the elision state, a negative value between the lower limit of the reward and the upper limit of the reward is selected as the reward value.

4. The method according to claim 2, characterized in that, The reward value is obtained based on the phoneme similarity between the phoneme sequence and the target wake-up word and the wake-up score of the wake-up model triggered by the false wake-up speech, including: The reward value is obtained by fusing the phoneme similarity and the wake-up score; wherein the reward value is positively correlated with the phoneme similarity and the wake-up score, and the wake-up model triggers wake-up when the wake-up score meets the triggering condition of the wake-up score threshold.

5. The method according to claim 1, characterized in that, The method further includes: Clustering is performed based on the corpus features of each of the accumulated sample corpora to obtain several cluster sets; Based on the corpus features of the sample corpus obtained this time, it is determined whether the sample corpus obtained this time can be assigned to any of the cluster sets mentioned above. The termination condition includes: the number of consecutive target iterations fails to generate sample corpora that cannot be assigned to any of the cluster sets.

6. The method according to claim 5, characterized in that, The method further includes: In response to the fact that the sample corpus obtained this time cannot be assigned to any of the cluster sets, a new cluster set is created based on the sample corpus obtained this time, and interference words that replace the target wake-up word to trigger false wake-ups are extracted from the sample corpus obtained this time as target interference words. When returning to the step of constructing model instructions based on the product requirements of the wake-up model, a new model instruction is constructed based on the product requirements and the target interference words.

7. The method according to claim 5, characterized in that, The method further includes: Based on the corpus features of the sample corpus obtained this time and the central features of each cluster set, the feature distances between the sample corpus obtained this time and each cluster set are obtained respectively. The reward value is updated based on the minimum feature distance to obtain a new reward value; wherein, the generative model adjusts its parameters based on the new reward value, and the magnitude of the minimum feature distance is positively correlated with the update magnitude of the reward value.

8. The method according to any one of claims 1 to 7, characterized in that, Before the step of returning the product requirements based on the wake-up model and constructing model instructions, the wake-up model is first trained based on the sample corpus, or the wake-up model is trained based on the sample corpus when the sample corpus has accumulated to a target quantity value; And / or, the product requirements may also include at least one of the following: several interfering words with similar pronunciation to the target wake word, or several application scenarios of the wake-up model; And / or, the model instruction is specifically used to instruct the generative model to generate the false wake-up text based on the application scenario that the wake-up model needs to be deployed in, wherein the false wake-up text does not contain the target wake-up word but contains interference words that are similar in pronunciation to the target wake-up word; And / or, the generative model adjusts its parameters based on the reward value through proximal policy optimization.

9. A training method for a wake-up model, characterized in that, include: Obtain sample corpus; wherein the sample corpus is obtained by the false wake-up corpus acquisition method according to any one of claims 1 to 8; The wake-up model is trained based on the sample corpus.

10. A voice wake-up method, characterized in that, include: Acquire the voice recordings from the voice device; Based on the response of the wake-up model to the collected speech, it is determined whether to wake up the speech device; wherein the wake-up model is obtained based on the training method of the wake-up model as described in claim 9.

11. A device for acquiring false wake-up corpus, characterized in that, include: The instruction construction module is used to construct model instructions based on the product requirements of the wake-up model; wherein, the product requirements include the target wake-up word of the wake-up model, and the model instructions are used to instruct the generative model to generate false wake-up text; The speech synthesis module is used to synthesize speech based on the false wake-up text output by the generative model in response to the model instruction, so as to obtain the false wake-up speech; The reward determination module is used to obtain the reward value generated by the generative model this time based on the response result of the wake-up model to the false wake-up speech, the recognized text of the false wake-up speech, and the phoneme sequence of the false wake-up text; wherein, the reward value is positively correlated with the probability that the false wake-up speech successfully falsely wakes up the wake-up model; The adjustment loop module is used to adjust the network parameters of the generative model based on the reward value, and to determine whether to select the false wake-up text and the false wake-up speech as sample corpus for training the wake-up model based on the reward value, and to return the product requirements based on the wake-up model and construct model instructions, until the termination condition is met.

12. A training device for a wake-up model, characterized in that, include: A corpus acquisition module is used to acquire sample corpus; wherein the sample corpus is obtained through the false wake-up corpus acquisition device as described in claim 11; The model training module is used to train the wake-up model based on the sample corpus.

13. A voice wake-up device, characterized in that, include: The voice acquisition module is used to acquire the voice data collected by the voice device. A wake-up determination module is used to determine whether to wake up the voice device based on the response result of the wake-up model to the collected voice; wherein the wake-up model is obtained based on the training device of the wake-up model as described in claim 12.

14. An electronic device, characterized in that, It includes at least a memory and a processor coupled to each other, wherein the memory stores at least program instructions, and the processor is used to execute the program instructions to implement the false wake-up corpus acquisition method according to any one of claims 1 to 8, or to implement the wake-up model training method according to claim 9, or to implement the voice wake-up method according to claim 10.

15. A computer-readable storage medium, characterized in that, The system stores program instructions that can be executed by a processor, the program instructions being used to implement the false wake-up corpus acquisition method according to any one of claims 1 to 8, or to implement the wake-up model training method according to claim 9, or to implement the voice wake-up method according to claim 10.

Images