Voice wake-up model training method, voice wake-up method, device and medium

Abstract

The invention discloses a voice wake-up model training method, a voice wake-up method, equipment and a medium, and relates to the technical field of artificial intelligence, and the method comprises the steps: obtaining a to-be-trained voice wake-up model and a preset target wake-up word; inputting the target wake-up word into a speech synthesis model, outputting to obtain positive sample speech data, and generating negative sample speech data opposite to the positive sample speech data; and constructing a training data set based on the positive sample voice data and the negative sample voice data, training a voice wake-up model based on the training data set, and obtaining a trained voice wake-up model until a preset training ending condition is met. According to the invention, effective training of the voice wake-up model is realized under the condition of small samples or zero samples.

Description

Technical Field

[0001] This application relates to the field of artificial intelligence technology, and in particular to a voice wake-up model training method, voice wake-up method, device and medium. Background Technology

[0002] Voice wake-up technology is a key entry point for smart devices to achieve voice interaction. Its working principle involves the device continuously listening to ambient sounds in a low-power standby state. When it detects that the user has spoken a specific keyword, it automatically activates and performs subsequent full voice command recognition. This technology acts like an "auditory trigger switch" for the device and is widely used in IoT devices such as smartphones, smart speakers, and in-vehicle systems, providing a foundation for natural and convenient human-computer interaction.

[0003] In recent years, deep learning models such as deep neural networks, convolutional neural networks, recurrent neural networks, and their variants have been widely applied in the keyword wake-up field due to their powerful feature extraction and sequence modeling capabilities. However, the performance of these deep learning methods is highly dependent on large-scale labeled data. In practical applications, collecting sufficient data for each new wake-up word or specific user group presents numerous challenges. Data collection and manual annotation are time-consuming, labor-intensive, and costly. To ensure the robustness of the model under different environments, accents, and channel conditions, the collected data must cover various acoustic scenarios, further increasing the complexity and cost of data collection. In addition, in cold-start scenarios or scenarios targeting niche keywords, there may be problems of "few samples" or even "zero samples," making traditional methods based on large amounts of real data difficult to implement effectively.

[0004] Therefore, how to effectively train a voice wake-up model under conditions of few or no samples has become a technical challenge that urgently needs to be solved in this field. Summary of the Invention

[0005] The main purpose of this application is to provide a voice wake-up model training method, voice wake-up method, device and medium, which aims to solve the technical problem of how to effectively train a voice wake-up model under small sample or zero sample conditions.

[0006] To achieve the above objectives, this application provides a voice wake-up model training method, which includes: Obtain the voice wake-up model to be trained and the preset target wake-up word; The target wake word is input into the speech synthesis model, and positive sample speech data is output. Negative sample speech data is generated relative to the positive sample speech data. A training dataset is constructed based on the positive sample speech data and the negative sample speech data. The voice wake-up model is trained based on the training dataset until the preset training termination condition is met, and then the trained voice wake-up model is obtained.

[0007] Furthermore, to achieve the above objectives, this application provides a voice wake-up method, which includes the following steps: Acquire voice interaction data, and perform feature extraction on the voice interaction data to obtain feature data; The feature data is input into a pre-trained voice wake-up model, and the voice wake-up recognition result is output. The voice wake-up model is a model trained using the voice wake-up model training method described above. If the voice wake-up recognition result indicates that the target wake-up word has been recognized, then the corresponding voice interaction device will be woken up.

[0008] In addition, to achieve the above objectives, this application also provides a voice interaction device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, the computer program being configured to implement the steps of the method described above.

[0009] In addition, to achieve the above objectives, this application also provides a readable storage medium, which is a computer-readable storage medium storing a computer program that is executed by a processor to implement the steps of the above method.

[0010] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the method described above.

[0011] One or more technical solutions proposed in this application have at least the following technical effects: This application effectively solves the technical challenge of training voice wake-up models under conditions of few or no samples by automatically generating training data using speech synthesis technology. Specifically, by inputting the target wake-up word into the speech synthesis model to generate positive sample speech data and constructing corresponding negative sample speech data, a sufficient training dataset is created. This dataset construction method bypasses the dependence on large-scale real-world labeled data, as the generation process of synthetic data does not require time-consuming and laborious data collection and manual annotation, significantly reducing costs. More importantly, when facing new wake-up words or niche scenarios where real-world speech data is lacking (i.e., cold start), the required training samples can be quickly generated from scratch, transforming the "zero-sample" or "few-sample" problem into a supervised learning problem supported by sufficient synthetic data. This allows the training dataset to replace or supplement real-world data, thus providing a feasible training foundation for the voice wake-up model under conditions of few or even zero samples, achieving effective training of the voice wake-up model. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0013] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 A flowchart illustrating the first embodiment of the voice wake-up model training method of this application; Figure 2 A schematic diagram of the DTC block structure involved in an embodiment of the voice wake-up model training method of this application; Figure 3 A schematic diagram of the voice wake-up model structure involved in one embodiment of the voice wake-up model training method of this application; Figure 4 A schematic diagram of the overall architecture of voice wake-up according to an embodiment of this application; Figure 5 This is a schematic diagram of the hardware operating environment of the voice wake-up model training method device in the embodiments of this application.

[0015] The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0016] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0017] With the rapid development of artificial intelligence and Internet of Things technologies, voice interaction has become an important way of human-computer interaction. As the entry point for voice interaction, Keyword Spotting (KWS) technology plays a crucial role, enabling devices to be activated from standby mode and enter a complete voice recognition state through predetermined keywords.

[0018] Traditional voice wake-up systems heavily rely on large amounts of precisely labeled speech data for model training. Early methods were mainly based on Hidden Markov Models (HMMs) and Gaussian Mixture Models (GMMs), but their performance limitations are significant. In recent years, deep learning models such as Deep Neural Networks (DNNs), Convolutional Neural Networks (CNNs), Recurrent Neural Networks (RNNs), and their variants (such as GRUs and Long Short-Term Memory) have been widely applied in this field. Their powerful feature extraction and sequence modeling capabilities have significantly improved the accuracy and robustness of wake-up models.

[0019] Despite the tremendous success of deep learning methods, their performance is highly dependent on large-scale, high-quality labeled training data. In practical applications, collecting sufficient data for each new wake word or specific user group presents significant challenges: First, data collection and manual annotation are time-consuming, labor-intensive, and extremely costly; second, to ensure the robustness of the model under different environments, accents, and channel conditions, the collected data needs to cover as many acoustic scenarios as possible (such as noisy environments, reverberant environments, etc.), which further increases the complexity and cost of data collection; finally, in cold-start scenarios or scenarios targeting niche keywords, there may be "few samples" or even "zero samples," making traditional methods based on large amounts of real data training difficult to implement.

[0020] To alleviate the problem of data scarcity, some data augmentation methods have emerged, such as adding background noise, changing speech rate, and pitch to expand existing data. However, these methods are essentially transformations of existing datasets and cannot create entirely new speech content. They have limited effectiveness in expanding the pronunciation diversity of the wake word itself, especially when there is very little or no real speech data for the target wake word, in which case these methods are difficult to implement.

[0021] Based on this, the main solution of this application is: to obtain a voice wake-up model to be trained and a preset target wake-up word; to input the target wake-up word into a speech synthesis model, output positive sample speech data, and generate negative sample speech data corresponding to the positive sample speech data; to construct a training dataset based on the positive sample speech data and the negative sample speech data, and to train the voice wake-up model based on the training dataset until the preset training termination condition is met to obtain the trained voice wake-up model.

[0022] This application creates a sufficient training dataset by inputting the target wake word into a speech synthesis model to generate positive sample speech data and constructing corresponding negative sample speech data. This dataset construction method bypasses the dependence on large-scale real-world labeled data, as the generation process of synthetic data does not require time-consuming and laborious data collection and manual annotation, significantly reducing costs. More importantly, when facing new wake words or niche scenarios lacking real-world speech data (i.e., cold start), the required training samples can be quickly generated from scratch, transforming the "zero-sample" or "few-sample" problem into a supervised learning problem supported by sufficient synthetic data. This allows the training dataset to replace or supplement real-world data, providing a feasible training foundation for the speech wake-up model under conditions of few or even zero samples, thus achieving effective training of the speech wake-up model.

[0023] It should be noted that the execution subject of each embodiment of the voice wake-up model training method of this application can be a computing service device with data processing, network communication and program running functions, such as a server, tablet computer, personal computer, mobile phone, etc., or a voice interaction device that can realize the above functions, such as headphones, AR (Augmented Reality) headset, VR (Virtual Reality) headset, etc. The embodiments of the voice wake-up model training method of this application do not impose specific limitations on this.

[0024] Based on this, this application proposes a voice wake-up model training method according to the first embodiment, referring to... Figure 1 As shown, the voice wake-up model training method includes the following steps S10~S30: Step S10: Obtain the voice wake-up model to be trained and the preset target wake-up word; The "voice wake-up model to be trained" can be an untrained or only pre-trained deep learning model. Regarding model architecture selection, besides common convolutional neural networks (CNNs), recurrent neural networks (RNNs), or their variants (such as LSTM and GRU), temporal convolutional networks (TCNs) and their variants exhibit significant advantages in sequence modeling tasks like speech processing due to their ability to process long sequences in parallel, their larger receptive field to capture long-range dependencies, and their structural stability, making them less prone to gradient vanishing problems. Therefore, they are a preferred model architecture. The task of this voice wake-up model is to determine whether the input audio contains the target wake-up word.

[0025] The preset "target wake word" is a specific word or phrase that the user wants the device to recognize and respond to, which can usually be entered by the user in text form or selected from a preset word library.

[0026] Step S20: Input the target wake-up word into the speech synthesis model, output positive sample speech data, and generate negative sample speech data corresponding to the positive sample speech data; The pre-set target wake word is input into the speech synthesis model. The speech synthesis model can be a text-to-speech (TTS) model that can convert the text wake word into speech data.

[0027] To construct a complete training dataset, negative sample speech data is also generated, corresponding to the positive sample speech data. This negative sample speech data is used to train the model to distinguish the target wake word from other non-wake words. This can be achieved in various ways, such as: Non-target word synthesis: Select other words that are similar to or unrelated to the target wake word and use the same speech synthesis model to generate their speech data as negative samples.

[0028] Silent or background noise segments: Collect or generate silent segments or various environmental background noises (such as street noise, office conversations) without any clear speech as negative samples.

[0029] Positive sample manipulation: Local cropping, frequency perturbation, or timing distortion of generated positive sample speech data to destroy its integrity and render it no longer a valid wake word. Such manipulated data can also be used as negative samples.

[0030] Step S30: Construct a training dataset based on the positive sample speech data and the negative sample speech data, and train the voice wake-up model based on the training dataset until the preset training termination condition is met to obtain the trained voice wake-up model.

[0031] The generated positive and negative sample speech data are aggregated, and each data point is labeled accordingly (e.g., 1 for positive samples and 0 for negative samples), thus constructing a complete and labeled training dataset. Subsequently, this training dataset is used to perform supervised learning on the voice wake-up model to be trained.

[0032] The training termination condition can be a pre-set condition, such as reaching a predetermined number of iterations, the loss function value falling below a predetermined threshold, the loss function converging, computational resources being exhausted, reaching a time limit, or a predetermined accuracy threshold. This embodiment does not impose specific restrictions on this.

[0033] It should be noted that the deployment method of the voice wake-up model can be flexibly selected according to the actual system architecture: if the model is trained on a cloud device, its lightweight version or the parameters required for inference can be deployed to the edge device, so that the edge device has the ability to independently perform routing decisions; if the model itself is trained on the edge, the edge device can directly use the locally trained model to make real-time voice wake-up decisions.

[0034] In recent years, text-to-speech (TTS) technology, especially end-to-end speech synthesis models based on neural networks, has made groundbreaking progress, with the naturalness and fidelity of synthesized speech approaching that of real human recordings. This provides a technological possibility for artificially generating high-quality, diverse speech training data. However, how to effectively utilize synthesized data to train high-performance wake-up models and address the distributional discrepancies between synthesized and real data remains an unsolved technical challenge in the field.

[0035] Based on this and the first embodiment of the voice wake-up model training method of this application, in the second embodiment of the voice wake-up model training method of this application, the content that is the same as or similar to the first embodiment can be referred to the above description, and will not be repeated hereafter. On this basis, the step of inputting the target wake-up word into the speech synthesis model and outputting positive sample speech data includes: Step A10: Obtain the preset configuration parameters, wherein the configuration parameters include gender, tone, speech rate, prompt audio, and prompt words; A series of configuration parameters can be obtained from a preset parameter library or user-defined settings. These parameters will guide the speech synthesis model to generate diverse positive sample speech data. The configuration parameters constitute a parameter space, specifically including: Gender: Used to control the basic timbre category of synthesized speech, typically male or female, affecting the fundamental frequency range and formant structure of synthesized speech.

[0036] Pitch: Used for fine adjustment of the volume of a voice. It can take values ​​within a preset continuous or discrete range (e.g., very low, low, medium, high, very high) to produce pitch variations within the same gender category.

[0037] Speech rate: Used to control the playback speed of speech per unit time. It can be set to different levels such as very slow, slow, medium, fast, and very fast to simulate different speaking habits and contexts.

[0038] Cue audio and cue words: These are used to achieve speaker adaptation with zero or few samples. The cue audio is a sample speech segment; the cue word is the corresponding text. By receiving this pair of information, the speech synthesis model can extract the speaker's voice features (voiceprint) from the sample speech segment, and then synthesize the target wake word using these voice features as a reference. This allows for the generation of personalized speech samples without recording a large amount of data for the speaker. Configuring this parameter enables the generation of speech samples for the target speaker without recording any speech data for the target wake word, effectively solving the cold start problem for new users. The cue audio and cue words can be easily obtained from existing speech databases or public datasets, such as the aishell open-source dataset, greatly reducing the data preparation threshold.

[0039] Step A20: Input the target wake word into the speech synthesis model, and synthesize and output multiple positive sample speech data according to the configuration parameters.

[0040] The acquired text sequence of the target wake word, along with the acquired configuration parameter combination, is input into a pre-trained speech synthesis model that supports conditional control, such as VITS (Variational Inference with Adversarial Learning for End-to-End Text-to-Speech) and FastSpeech (Fast and High-Quality End-to-End Text to Speech). This model parses and generates speech waveforms that conform to specified acoustic characteristics based on the input parameters. To maximize data diversity, the configuration parameter space can be automatically traversed or randomly sampled to generate multiple positive sample speech data. For example, it can generate speech of the same wake word in various combinations such as "male, high pitch, fast," "female, mid pitch, slow," and "timbre corresponding to cue audio X, normal speed." This systematic parameterized generation method can efficiently and cost-effectively create a rich set of positive samples covering a variety of timbres, intonations, and rhythms, greatly enhancing the robustness of the future trained wake-up model to speaker variations and changes in pronunciation style from the data source.

[0041] This embodiment constructs a configuration parameter space encompassing multiple dimensions such as gender, tone, speech rate, and prompt audio and prompt words. By utilizing a speech synthesis model for systematic parameterized generation, it can automatically create a highly diverse positive sample speech dataset in terms of timbre, intonation, and rhythm. This method is not only efficient and low-cost, but more importantly, by actively and systematically simulating various variations present in real human speech (such as different speakers' timbre and pronunciation habits), it significantly enhances the diversity and coverage of the synthetic dataset from the source of data generation. This directly improves the robustness of the wake-up model trained using synthetic data, enabling it to better adapt to speaker variations and changes in pronunciation style. Furthermore, through this intentional approximation and extensive sampling of the real speech distribution, it effectively reduces the distributional difference between synthetic and real data, providing an effective technical path to solve the core challenge of distributional discrepancies.

[0042] Crucially, this embodiment selects a combination of gender, pitch, speech rate, and prompt audio and prompt words, rather than randomly or generically choosing other acoustic parameters. This selection is not arbitrary but based on a deep understanding of the nature of speech signals and key sources of variation in real-world application scenarios, thus achieving a good balance between generation efficiency, model performance, and practical problem-solving capabilities. Specifically, gender, pitch, and speech rate are chosen because they collectively constitute the core dimensions of pronunciation style and prosodic features in human speech that are most significant and influential on perception and model judgment. Gender and pitch together determine the basic timbre and frequency range of the voice, while speech rate directly affects the temporal structure and unit duration of speech. The combination of these three can efficiently simulate the most significant pronunciation differences between different speakers in the real world. If too few parameters are selected (e.g., only controlling speech rate), the generated data will lack diversity, failing to cover basic timbre variations and hindering effective training of the model's robustness. Conversely, selecting too many or additional minor or difficult-to-control parameters (e.g., deliberately simulating specific channel distortions or complex background noise) unnecessarily increases the complexity and uncontrollability of the synthesis process, potentially introducing interference unrelated to the core speech content, thus distracting the model and hindering its focused learning of the acoustic patterns of the wake word itself. Furthermore, introducing the parameter pair of cue audio and cue word transcends the method of simulating timbre through limited, discrete categories (e.g., male / female), achieving accurate replication of a specific speaker's timbre with continuous, high-fidelity reproduction. This choice directly addresses the core application challenge of "cold start." Thus, the combination of these parameters constitutes a complete parameter space that covers broad common variations while accurately capturing individual characteristics, systematically reducing the distributional differences between synthesized and real data. This ensures that the diversity and effectiveness of the generated data directly contribute to improving the model's performance in real-world scenarios.

[0043] In one possible implementation, the step of generating negative sample speech data relative to the positive sample speech data includes: Step B10: Obtain the pre-set non-wake-up obfuscation terms; The non-wake-up confusion word is another word or phrase that is highly similar to the target wake-up word in pronunciation, phoneme sequence, or prosody. For example, if the target wake-up word is "Goltek", then the non-wake-up confusion word can be set to "Koltek", "Golmic", etc. These confusion words are high-risk samples that the model is most likely to misjudge. Targeted generation of such negative samples can directly and effectively improve the model's discrimination and anti-interference ability.

[0044] Step B20: Input the non-wake-up obfuscation word into the speech synthesis model and output synthesized obfuscated speech data; To ensure consistency in data generation conditions, the same speech synthesis model and configuration parameter space (such as gender, pitch, and speech rate) can be used as those used to generate positive samples to synthesize multiple speech data for each non-wake-up confusion word, i.e., confused speech data. This ensures that the confused speech and the positive sample speech are comparable in acoustic characteristics such as timbre and intonation, forcing the model to focus on the differences in the speech content of the words themselves, rather than irrelevant acoustic features, thereby enabling more refined learning.

[0045] Step B30: Select general speech data from the preset general speech corpus and select pure noise speech data from the preset noise database; The pre-set general speech corpus contains a large number of speech segments such as daily conversations, news broadcasts, and commands that are unrelated to the target wake word. Randomly selected general speech data from this corpus is used to simulate non-target speech interference that the model may encounter in real-world environments. The pre-set noise database collects pure noise speech data from various environments, such as white noise, street noise, and background noise from office conversations. This noise data, which contains no clear semantics, is used to train the model's robustness in noisy environments and to teach it to remain silent when there is no wake word.

[0046] Step B40: Combine the confused speech data, the general speech data, and the pure noise speech data to obtain negative sample speech data that is opposite to the positive sample speech data.

[0047] The generated obfuscated speech data, selected general speech data, and pure noise speech data can be combined or directly aggregated in a certain proportion to form a comprehensive, multi-level negative sample speech data set. This combination ensures that the negative sample set includes not only high-risk easily obfuscated samples, but also general irrelevant speech interference and pure background noise, thereby systematically training the wake-up model to cope with various complex negative scenarios and significantly reducing the false wake-up rate.

[0048] This embodiment selects non-wake-up confusion words that are highly similar to the target wake-up word in phoneme sequence and pronunciation, and uses the same speech synthesis model to generate confused speech data. This creates highly targeted and difficult-to-distinguish "high-risk" negative samples, allowing the model to focus on learning the subtle differences in the wake-up word's speech features rather than relying on simple acoustic characteristics for judgment, thus improving the model's discriminative power. Secondly, by introducing irrelevant speech segments from a general speech corpus and pure environmental noise from a noise database, the coverage of negative samples is greatly expanded, simulating various non-target speech interferences and complex acoustic backgrounds that may be encountered in real-world scenarios. This trains the model's ability to distinguish meaningful speech from irrelevant interference and background noise, enhancing its robustness in noisy environments and significantly reducing the false wake-up rate. Finally, by combining easily confused samples, general interference samples, and pure noise samples, a comprehensive negative sample set is constructed. This construction method ensures that the model can undergo balanced training for various negative scenarios, fundamentally solving the problem of weak model generalization ability and high false alarm rate caused by a single or insufficient negative sample, and achieving systematic optimization of model performance.

[0049] Based on the first and / or second embodiments of the voice wake-up model training method of this application, in the third embodiment of the voice wake-up model training method of this application, the content that is the same as or similar to the first and second embodiments described above can be referred to the above description and will not be repeated hereafter. Based on this, the step of training the voice wake-up model based on the training dataset includes: Step C10: Preprocess the sample speech data in the training dataset to obtain the preprocessed sample speech data, wherein the preprocessing includes pre-emphasis, frame segmentation and windowing; Pre-emphasis processes the speech signal using a high-pass filter to enhance the amplitude of high-frequency signals. This preprocessing reduces the attenuation of high-frequency signals during speech acquisition, improving the clarity of the speech signal.

[0050] Frame segmentation divides the pre-emphasized long-time speech signal into a sequence of short-time frames. Typically, each frame is 20-30 milliseconds long and the frame shift is 10-15 milliseconds.

[0051] Windowing reduces the signal truncation effect caused by framing by multiplying each frame of signal by a window function (such as a Hamming window), thereby reducing spectral leakage.

[0052] Step C20: Extract the log-Mel spectrogram of the preprocessed sample speech data to obtain sample features; Log-Mel spectrograms are a speech feature representation method that effectively captures the spectral information of speech signals. The specific extraction steps involve performing a Fourier transform on each windowed frame of the speech signal to obtain its amplitude spectrum. Then, the amplitude spectrum is passed through a set of Mel-scale filters (whose frequency spacing is designed based on human auditory characteristics, with high resolution at low frequencies and low resolution at high frequencies) to map the linear frequency scale to a Mel scale that better matches human hearing, resulting in the Mel spectrum. Finally, the logarithm of this Mel spectrum is taken to obtain the log-Mel spectrogram. This operation not only compresses the dynamic range, making the feature distribution closer to a normal distribution, which is beneficial for model optimization, but also simulates the approximate logarithmic perception of sound loudness by the human ear. This log-Mel spectrogram serves as the sample features for training the voice wake-up model.

[0053] Step C30: Train the voice wake-up model using the sample features as input.

[0054] The training process begins by using the extracted sample features as input to the voice wake-up model. After model initialization, the sample features are input into the model for forward propagation calculation, yielding the model's output, typically a probability value representing the probability that the input speech data is the target wake-up word. Based on the model's output and the true labels, a loss function is calculated; commonly used loss functions include cross-entropy loss and mean squared error loss. The gradient of the loss function with respect to the model parameters is calculated using the backpropagation algorithm, and optimization algorithms (such as stochastic gradient descent, Adam, etc.) are used to update the model parameters to minimize the loss function. The forward propagation, loss calculation, and backpropagation steps are repeated until the preset training termination condition is met, resulting in a trained voice wake-up model. This model can effectively recognize the target wake-up word under conditions of small samples or zero samples, making it suitable for various practical application scenarios.

[0055] In one possible implementation, the step of training the voice wake-up model using the sample features as input includes: Step D10: Input the sample features into the voice wake-up model to obtain a wake-up probability sequence, wherein the wake-up probability sequence includes at least one wake-up probability, and each wake-up probability represents the probability that the sample voice data at a corresponding time step is the target wake-up word; The extracted sample features (i.e., the time series of log-Melograms) are input into the speech wake-up model. The model processes the input feature sequence frame by frame and outputs a wake-up probability for each time step (i.e., each time frame). This probability is a value between 0 and 1, representing the probability that the model considers the speech data at the corresponding time step to belong to the target wake-up word. All these probability values ​​are arranged in chronological order to form a wake-up probability sequence.

[0056] Step D20: Select the maximum value in the wake-up probability sequence as the confidence score, and calculate the prediction loss based on the confidence score and the wake-up labels corresponding to the sample speech data in the training dataset; The maximum value in the wake-up probability sequence is selected as the confidence score for the entire sample. This confidence score is then compared to the wake-up label corresponding to the sample's speech data, and the prediction loss is calculated using a predefined loss function (e.g., binary cross-entropy loss). This loss function quantifies the difference between the model's predicted confidence score and the true label; a larger loss value indicates a more severe prediction error.

[0057] Step D30: Iteratively optimize the voice wake-up model based on the predicted loss.

[0058] The voice wake-up model can be iteratively optimized using the backpropagation algorithm based on the calculated prediction loss. The backpropagation algorithm calculates the gradient of the loss function with respect to the model parameters and uses optimization algorithms (such as stochastic gradient descent, Adam, etc.) to update the model parameters to minimize the prediction loss.

[0059] This embodiment uses the "maximum value of the wake-up probability sequence" as the confidence score for the entire speech segment and optimizes the loss accordingly. This directly corresponds to the practical requirement of voice wake-up: "it should wake up as soon as a keyword is successfully detected once in the audio stream." This design ensures that the model training objective is consistent with the final application objective, allowing the model to learn key rules: for a positive sample speech containing a wake-up word, the model does not need to guarantee high output probabilities at all time steps, but only needs to provide high confidence at one or several key moments when the keyword appears to significantly reduce the overall loss. This effectively avoids the detection sluggishness caused by the model over-focusing on non-key segments or averaging output, significantly reducing the false positive rate. At the same time, for negative samples, this loss function requires that all values ​​in the entire probability sequence not be too high, thereby ensuring the model's ability to distinguish non-target speech and suppressing false wake-ups. Through iterative optimization using gradient feedback based on sequence maximum values, the model is guided to accurately locate and identify the local features most representative of the target wake word in the speech stream, rather than learning irrelevant global patterns. This enables sensitive detection of keyword events without the need for fine-grained frame-level annotation, and the ultimately trained model performs excellently in balancing detection rate and false alarm rate.

[0060] In one possible implementation, the voice wake-up model is a multi-scale dilated temporal convolutional network (DTC network). The DTC network includes at least one DTC stack, a fusion layer, and a fully connected layer connected sequentially. The dilated temporal convolution stack includes multiple dilated temporal convolutional blocks connected sequentially. The step of inputting the sample features into the voice wake-up model to obtain the wake-up probability sequence includes: Step E10: Extract features from the input data through each DTC stack to obtain a feature vector. When the DTC stack is the first layer stack, the input data is a vector obtained by fusing the historical input data of the DTC stack with the sample features. When the DTC stack is not the first layer stack, the input data is a vector obtained by fusing the historical input data of the empty DTC stack with the output of the previous layer DTC stack. Multi-scale Dilated Temporal Convolutional Network (MDTC) is an advanced deep neural network architecture designed specifically for modeling time-series signals. Its core idea is to capture the dependencies at different time scales in the input sequence simultaneously through parallel or stacked dilated convolutional layers with different dilation rates.

[0061] In this embodiment, the multi-scale dilated temporal convolutional network includes at least one DTC stack, and each DTC stack consists of multiple sequentially connected DTC (Dilated Temporal Convolution Block) blocks. For example, refer to... Figure 3 As shown, four DTC blocks are stacked to form a DTC stack, and the inflation rates of the four DTC blocks are set to 1, 2, 4 and 8, respectively.

[0062] The DTC block is used to capture contextual information at different time scales in the input audio sequence. For example, refer to... Figure 2 As shown, each DTC block consists of a series of interconnected layers: a Dilated-Depth TCN, a Batch Normalization (BN) layer, a Point-Conv layer, a BN+ReLU (Rectified LinearUnit) activation function layer, another Point-Conv layer, a BN layer, a Scale layer, and ReLU. Furthermore, a Squeeze-and-Excitation (SE) module is added after the final PointConv layer to learn attention information between different channels. To prevent gradient vanishing and gradient exploding, an Add (summation) residual connection is used between the input and the last ReLU activation function.

[0063] For the first-layer DTC stack, its input data is a vector formed by fusing the historical input data of that block with sample features. For non-first-layer DTC stacks, the input data is a vector formed by fusing the historical input data of that stack with the output of the previous layer's DTC stack. Through this cascaded structure, each layer performs deeper feature extraction based on the fusion of historical information and abstract features from the previous layer, thus providing the model with dynamic and continuously evolving temporal context information. Specifically, the historical input data, as the "memory" of the DTC stack of previous time steps, is fused with the current new input (sample features or the output of the previous layer), ensuring that the model's processing at each moment is not only based on the current instantaneous information, but also on a continuous temporal context containing recent historical states. This mechanism effectively simulates the inherent temporal continuity of speech signals, ensuring that the model can perceive the dependencies between consecutive frames in the speech stream, thereby enabling more accurate identification of context-dependent key patterns such as phoneme transitions and syllable boundaries. This design significantly enhances the model's ability to capture temporal dynamic features, enabling it to make more coherent and accurate sequence judgments and effectively avoid misjudgments caused by focusing only on local segments.

[0064] The historical input data refers to the pre-cached input data of the DTC stack at the previous inference time step. The fusion of the two vectors can employ a sliding window concatenation strategy to achieve smooth transition of temporal context. Specifically, if the historical input data is a 3×64-dimensional vector, and the input data at the current time step (i.e., sample features or the output of the previous DTC stack) is also a 3×64-dimensional vector, then the last 2×64-dimensional vector is extracted from the historical input data and concatenated with the first 1×64-dimensional vector of the current input data, thereby constructing a new fused vector that maintains the 3×64 dimension. This fusion mechanism not only ensures the consistency of the input dimensions but, more importantly, achieves efficient integration of historical context and current information, generating an enhanced feature vector containing continuous temporal semantics, providing a rich contextual foundation for subsequent deep feature extraction by the DTC stack.

[0065] Step E20: The feature vectors obtained from each DTC stack are fused through the fusion layer to obtain a fused feature vector; The fusion layer receives feature vectors from the output of a dilated temporal convolution stack. Fusion can be achieved through operations such as summation. By fusing these feature vectors from different depths, the fusion layer effectively combines low-level detailed acoustic features (such as phoneme information) with high-level abstract semantic features (such as word-level information), forming a fused feature vector containing rich multi-scale information. This provides a more comprehensive and robust feature representation for the final classification.

[0066] Step E30: The fused feature vector is subjected to probability mapping processing through the fully connected layer to obtain a wake-up probability sequence.

[0067] The resulting fused feature vector is input into a fully connected layer. This fully connected layer maps the high-dimensional fused feature vector to a scalar output. Since the model processes sequential data, this operation is performed independently at each time step, ultimately outputting a wake-up probability between 0 and 1 for each time step in the input sample feature sequence. These probabilities, arranged chronologically, constitute the final wake-up probability sequence. This sequence intuitively reflects the model's perception of the probability that the speech at each time step belongs to the target wake-up word.

[0068] For example, refer to Figure 3 As shown, the sample feature Input x is input into a multi-scale dilated temporal convolutional network, and then multi-scale feature extraction is performed through multiple layers of DTC-Stack. The fusion layer fuses the outputs of each DTC-Stack layer by adding. The output of the fusion layer passes through a fully connected (FC) layer, and then is activated by the Sigmoid function to obtain the wake-up probability sequence Output Y.

[0069] Furthermore, this application also proposes a voice wake-up method. The execution subject of each embodiment of the voice wake-up method can be a computing service device with data processing, network communication and program running functions, such as a server, tablet computer, personal computer, mobile phone, etc., or a voice interaction device capable of realizing the above functions, such as headphones, AR headsets, VR headsets and other wearable devices. The embodiments of the voice wake-up method in this application do not impose specific limitations on this.

[0070] Based on this, the voice wake-up method in the first embodiment includes the following steps S100~S300: Step S100: Obtain voice interaction data, and perform feature extraction on the voice interaction data to obtain feature data; Raw voice interaction data, i.e., audio streams, can be acquired via microphone arrays or audio input interfaces. Subsequently, a feature extraction process identical to that used in the model training phase is performed on this audio stream to ensure consistency in the feature space. For example, this feature extraction process specifically includes: first, preprocessing, including pre-emphasis to enhance high-frequency components; then, framing and windowing (such as Hamming windowing) to convert continuous audio into short-time stationary frame sequences; next, calculating the log-Melogram features for each frame of audio, transforming the entire audio stream into a temporal feature sequence, which serves as the feature data input to the subsequent model.

[0071] Step S200: Input the feature data into the pre-trained voice wake-up model and output the voice wake-up recognition result. The voice wake-up model is a model trained using the voice wake-up model training method described in any of the above embodiments. The obtained feature data is input into a pre-trained voice wake-up model, and the wake-up probability, i.e., the voice wake-up recognition result, is output.

[0072] Step S300: If the voice wake-up recognition result indicates that the target wake-up word has been recognized, then the corresponding voice interaction device is woken up.

[0073] The voice wake-up recognition result can be compared with a set classification threshold. If the voice wake-up recognition result is greater than the classification threshold, it is determined that the voice wake-up recognition result indicates that the target wake-up word has been recognized; otherwise, if the voice wake-up recognition result is less than or equal to the classification threshold, it is determined that the voice wake-up recognition result indicates that the target wake-up word has not been recognized.

[0074] When the target wake-up word is successfully recognized, a wake-up signal is triggered to wake up the corresponding voice interaction device. The device then starts its full speech recognition engine to receive and process subsequent user voice commands, completing the entire interaction process. If the target wake-up word is not recognized, the device remains in its current state and continues listening.

[0075] Based on the first embodiment of the voice wake-up method, in the second embodiment of the voice wake-up method, the content that is the same as or similar to that in the first embodiment can be referred to the above description, and will not be repeated hereafter. Based on this, before the step of extracting features from the voice interaction data to obtain feature data, the method further includes: Step A100: Preprocess the voice interaction data to obtain the framed voice interaction data, wherein the preprocessing includes at least frame division; The acquired voice interaction data undergoes preprocessing. Voice interaction data is typically continuous speech signals captured from a microphone or other audio input device. The purpose of preprocessing is to transform these continuous speech signals into a form suitable for subsequent feature extraction and model processing. Preprocessing includes at least framing, which divides the continuous speech signal into a series of short frames, each containing a certain length of speech signal. Typically, each frame is 20-30 milliseconds long, with a frame shift of 10-15 milliseconds to ensure some overlap between adjacent frames, helping to capture the continuity of the speech signal.

[0076] Step A200: For each frame of the voice interaction data, perform the step of extracting features from the voice interaction data to obtain feature data, and stream process each frame of the voice interaction data to obtain the voice wake-up recognition result corresponding to each frame of the voice interaction data.

[0077] For each frame of segmented speech data, feature extraction is performed, such as calculating the log-Melogram of that frame. The extracted features of the current frame are then input into the voice wake-up model for inference. The model outputs a wake-up probability for the current moment, which is the voice wake-up recognition result corresponding to that frame. This streaming processing mechanism enables near real-time generation of detection results while audio is flowing in, significantly reducing the latency of wake-up detection and meeting the stringent real-time requirements of voice interaction applications.

[0078] For example, to aid in understanding the technical concept or principle of the voice wake-up model training method and voice wake-up method combined with the above embodiments, a specific embodiment is now provided. In this specific embodiment, refer to... Figure 4 As shown, the overall process includes: Phase 1: Synthetic Data Generation Phase. The goal of this phase is to build a training dataset for training the wake-up model, which contains positive and negative sample speech data.

[0079] 1. Wake-up word audio synthesis (generation of positive sample speech data): The target wake word is used as the input text to the pre-trained speech synthesis model. A large number of positive sample speech data are generated in batches according to the configuration parameters shown in Table 1 below.

[0080] Table 1 Configuration Parameter Table

[0081] By configuring the above parameters, the following two points can be guaranteed: 1) Speaker diversity: Multiple speaker models of different genders, ages, and timbres are synthesized to simulate the pronunciation characteristics of different users.

[0082] 2) Prosodic diversity: Adjust the speech rate, pitch, stress and emotion of synthesized speech to generate samples with rich prosodic variation.

[0083] Ultimately, tens of thousands of high-quality, clean synthesized speech words with target wake words are generated, forming positive sample speech data.

[0084] 2. Non-wake word audio synthesis and collection (negative sample speech data generation): Randomly select non-wake words or unrelated phrases of similar length that are easily confused with the target wake word in terms of phonemes, and use the above TTS method to generate a large number of speech samples as easily confused negative samples.

[0085] A large number of speech segments unrelated to the wake word are randomly selected from publicly available general speech corpora (such as LibriSpeech) as general negative samples.

[0086] Various types of environmental noise are collected from publicly available noise databases (such as ESC-50) as pure noise negative samples.

[0087] 3. Training dataset construction: The positive and negative samples generated and collected above are combined, and each data point is labeled accordingly, such as 0 for positive sample speech data and -1 for negative sample speech data, to form the final training dataset.

[0088] Phase Two: Model Training Phase This stage utilizes the training dataset constructed in the first stage to train an MDTC voice wake-up model.

[0089] 1. Preprocessing and Feature Extraction Each audio waveform in the synthetic dataset is preprocessed, including pre-emphasis, framing, and windowing. Then, the log-Mel spectrogram of the audio is extracted as the input feature of the model. This feature can well preserve the spectral information of speech and conforms to the auditory characteristics of the human ear.

[0090] 2. MDTC Model Structure The voice wake-up model employs a multi-scale dilated temporal convolutional network (MDTC) as its core classifier. This network consists of multiple dilated temporal convolution blocks. Each convolutional block uses a different dilation rate, enabling the network to simultaneously capture contextual information at different time scales within the input audio sequence. Branches with low dilation rates focus on fine-grained local features (such as phonemes), while branches with high dilation rates focus on global macro-level features (such as words and intonation). The high-level features output from each parallel branch are fused through summation.

[0091] 3. Training based on Max-Pooling Loss: Max-Pooling Loss is used as the training objective function for the model. The fused feature sequence is input into a fully connected layer, which outputs a score for each time step, representing the probability that the time step belongs to the wake word, forming a score sequence S=[s1, s2, ..., sT]. Global Max-Pooling is performed on this score sequence S, and the maximum value s_max=max(S) in the entire sequence is selected as the confidence score of the speech sample.

[0092] When calculating the loss, s_max and the true labels of the samples are substituted into the loss function (such as binary cross-entropy) for calculation. The Adam optimizer is used to train the MDTC model end-to-end on the training dataset with Max-Pooling Loss as the optimization objective, iteratively optimizing the network parameters until the model converges.

[0093] Phase 3: Model Deployment and Streaming Inference The trained MDTC model is deployed in actual voice interaction devices. Its inference process adopts an efficient streaming processing mode to achieve low-latency real-time wake-up. An audio frame (e.g., 25ms long) is taken at certain intervals (e.g., every 10ms) and preprocessing (pre-emphasis, windowing) and feature extraction (calculating log-Mel spectrogram features) are applied.

[0094] Thanks to MDTC's temporal convolutional structure, the model can process frames one by one without waiting for the entire audio segment to be acquired, and efficiently maintains the historical input data of each layer of the DTC stack using dilated convolution and caching mechanisms. For each input frame, the model outputs a confidence score s_t corresponding to that moment. This score represents the probability that the model judges the occurrence of the wake word ending at the current moment (time t). When a confidence score s_t exceeds a preset activation threshold, it is determined that the target wake word has been recognized, and the device wake-up interruption is triggered, entering the subsequent voice interaction process.

[0095] It should be noted that the above examples are only used to help understand this embodiment and do not constitute a limitation on the voice wake-up model training process and voice wake-up process of this embodiment. Any simple modifications based on this technical concept are within the protection scope of this application.

[0096] Furthermore, this application also proposes a voice interaction device, which includes: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program is configured to implement the steps of the method described above.

[0097] refer to Figure 5The diagram illustrates a structural schematic of a voice interaction device suitable for implementing the embodiments of this application. The voice interaction device in the embodiments of this application may also include, but is not limited to, mobile terminals such as headphones, AR headsets, VR headsets, mobile phones, servers, laptops, PDAs (Personal Digital Assistants), PADs (Portable Application Description), PMPs (Portable Media Players), etc., as well as fixed terminals such as digital TVs, desktop computers, etc. Figure 5 The voice interaction device shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0098] like Figure 5 As shown, the voice interaction device may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for the operation of the voice interaction device. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following systems can be connected to the I / O interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the voice interaction device to communicate wirelessly or wiredly with other devices to exchange data. Although the figures show voice interaction devices with various systems, it should be understood that it is not required to implement or possess all of the systems shown. More or fewer systems may be implemented alternatively.

[0099] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0100] The voice interaction device provided in this application adopts the voice wake-up model training method in the above embodiments, which can solve the problem of how to effectively train the voice wake-up model under conditions of small samples or zero samples. Compared with the prior art, the beneficial effects of the voice interaction device provided in this application are the same as the beneficial effects of the voice wake-up model training method provided in the above embodiments, and other technical features in this voice interaction device are the same as the features disclosed in the method of the previous embodiment, and will not be repeated here.

[0101] It should be understood that the various parts disclosed in this application can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics can be combined in any suitable manner in one or more embodiments or examples.

[0102] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0103] In addition, to achieve the above objectives, this application also provides a readable storage medium having computer-readable program instructions (i.e., a computer program) stored thereon, which are used to execute the voice wake-up model training method in the above embodiments.

[0104] The computer-readable storage medium provided in this application embodiment may be, for example, a USB flash drive, but is not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, portable compact disk read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this embodiment, the computer-readable storage medium may be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, system, or device. The program code contained on the computer-readable storage medium may be transmitted using any suitable medium, including but not limited to: wires, optical cables, RF (Radio Frequency), etc., or any suitable combination thereof.

[0105] The aforementioned computer-readable storage medium may be included in the voice interaction device; or it may exist independently and not be assembled into the voice interaction device.

[0106] The aforementioned computer-readable storage medium carries one or more programs, which, when executed by a voice interaction device, cause the voice interaction device to implement the process steps of any of the above embodiments.

[0107] The readable storage medium provided in this application is a computer-readable storage medium that stores computer-readable program instructions (i.e., a computer program) for executing the above-described voice wake-up model training method. This program is capable of effectively training the voice wake-up model under conditions of few or zero samples. Compared with the prior art, the beneficial effects of the computer-readable storage medium provided in this application are the same as those of the voice wake-up model training method provided in the above embodiments, and will not be elaborated upon here.

[0108] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for training a voice wake-up model, characterized in that, The voice wake-up model training method includes the following steps: Obtain the voice wake-up model to be trained and the preset target wake-up word; The target wake word is input into the speech synthesis model, and positive sample speech data is output. Negative sample speech data is generated relative to the positive sample speech data. A training dataset is constructed based on the positive sample speech data and the negative sample speech data. The voice wake-up model is trained based on the training dataset until the preset training termination condition is met, and then the trained voice wake-up model is obtained.

2. The voice wake-up model training method as described in claim 1, characterized in that, The step of inputting the target wake-up word into the speech synthesis model and outputting positive sample speech data includes: Obtain preset configuration parameters, wherein the configuration parameters include gender, tone, speech rate, prompt audio, and prompt words; The target wake word is input into the speech synthesis model, which then synthesizes and outputs multiple positive sample speech data according to the configuration parameters.

3. The voice wake-up model training method as described in claim 1, characterized in that, The step of generating negative sample speech data relative to the positive sample speech data includes: Retrieve pre-set non-wake-up obfuscation terms; The non-wake-up obfuscation word is input into the speech synthesis model, and synthesized obfuscated speech data is output. Select general speech data from a preset general speech corpus and select pure noise speech data from a preset noise database; By combining the obfuscated speech data, the general speech data, and the pure noise speech data, negative sample speech data corresponding to the positive sample speech data is obtained.

4. The voice wake-up model training method as described in claim 1, characterized in that, The step of training the voice wake-up model based on the training dataset includes: The sample speech data in the training dataset is preprocessed to obtain the preprocessed sample speech data, wherein the preprocessing includes pre-emphasis, frame segmentation and windowing; The log-Mel spectrogram of the preprocessed sample speech data is extracted to obtain the sample features; The voice wake-up model is trained using the sample features as input.

5. The voice wake-up model training method as described in claim 4, characterized in that, The step of training the voice wake-up model using the sample features as input includes: The sample features are input into the voice wake-up model to obtain a wake-up probability sequence, wherein the wake-up probability sequence includes at least one wake-up probability, and each wake-up probability represents the probability that the sample voice data at a time step is the target wake-up word; The maximum value in the wake-up probability sequence is selected as the confidence score, and the prediction loss is calculated based on the confidence score and the wake-up label corresponding to the sample speech data in the training dataset. The voice wake-up model is iteratively optimized based on the predicted loss.

6. The voice wake-up model training method as described in claim 5, characterized in that, The voice wake-up model is a multi-scale dilated temporal convolutional network, which includes at least one DTC stack, a fusion layer, and a fully connected layer connected in sequence. The step of inputting the sample features into the voice wake-up model to obtain the wake-up probability sequence includes: Feature vectors are obtained by extracting features from the input data through each DTC stack. When the DTC stack is the first layer stack, the input data is a vector obtained by fusing the historical input data of the DTC stack with the sample features. When the DTC stack is not the first layer stack, the input data is a vector obtained by fusing the historical input data of the DTC stack with the output of the previous layer DTC stack. The feature vectors obtained from each DTC stack are fused through the fusion layer to obtain a fused feature vector. The fully connected layer performs probability mapping on the fused feature vector to obtain a wake-up probability sequence.

7. A voice wake-up method, characterized in that, The voice wake-up method includes the following steps: Acquire voice interaction data, and perform feature extraction on the voice interaction data to obtain feature data; The feature data is input into a pre-trained voice wake-up model, and the voice wake-up recognition result is output. The voice wake-up model is a model trained using the voice wake-up model training method as described in any one of claims 1 to 6. If the voice wake-up recognition result indicates that the target wake-up word has been recognized, then the corresponding voice interaction device will be woken up.

8. The voice wake-up method as described in claim 7, characterized in that, Before the step of extracting features from the voice interaction data to obtain feature data, the method further includes: The voice interaction data is preprocessed to obtain framed voice interaction data, wherein the preprocessing includes at least frame division; For each frame of the voice interaction data, the step of extracting features from the voice interaction data to obtain feature data is performed, and each frame of the voice interaction data is processed in a streaming manner to obtain the voice wake-up recognition result corresponding to each frame of the voice interaction data.

9. A voice interaction device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the computer program, when executed by the processor, implements the method as described in any one of claims 1 to 8.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method as described in any one of claims 1 to 8.

Images