Voice interaction method of intelligent glasses

Abstract

The invention discloses a voice interaction method for intelligent glasses, and the method specifically comprises the steps: carrying out the wake-up word extraction of environment voice data through employing a convolutional neural network model based on an attention mechanism, and obtaining a wake-up word recognition result; verifying the wake-up word recognition result through a recurrent neural network model to obtain a wake-up word verification result, and correcting the wake-up word recognition result according to the wake-up word verification result to obtain a target wake-up word recognition result; and extracting voiceprint features when a user registers by adopting a voiceprint recognition technology, constructing a voiceprint recognition model according to the voiceprint features, and performing voice wake-up on the intelligent glasses through the voiceprint recognition model and the target wake-up word recognition result. According to the method, the accuracy, robustness and user experience of intelligent glasses wake-up word detection are remarkably improved, meanwhile, system resource consumption is reduced, and reliable technical support is provided for voice interaction of the intelligent glasses.

Description

Technical Field

[0001] The present invention relates to the field of artificial intelligence technology, and in particular to a voice interaction method for smart glasses. Background Art

[0002] As an emerging type of wearable device, smart glasses provide users with a more convenient operating experience through integrated voice interaction functions. However, in daily use, the voice interaction function of smart glasses faces severe challenges, especially in complex and changeable noise environments. Although traditional voice wake-up technologies, such as dynamic time warping (DTW) and hidden Markov models (HMM), have achieved certain results in certain specific scenarios, their performance depends to a large extent on the quality of training data. In actual application scenarios with severe noise interference and complex and changeable environments, these traditional algorithms often show the problem of insufficient robustness, which can easily lead to false triggering or failure to work properly.

[0003] In recent years, the rapid development of deep learning technology has provided new solutions for voice wake-up. Deep learning models have shown excellent performance in silent environments and can accurately identify and wake up smart devices. However, when deep learning models are applied to scenarios with high real-time requirements such as smart glasses, they face problems such as large computing volume and high power consumption. Smart glasses are limited by their size and battery capacity and have strict requirements on power consumption. An overly complex wake-up model will not only significantly increase the amount of computing, causing the device to heat up and shorten battery life, but also affect the user experience.

[0004] In addition, the voice wake-up technology of smart glasses needs to be improved in environmental adaptability. When moving or engaging in other activities, the environmental noise of the user is often large and varied. Traditional wake-up word detection algorithms have difficulty accurately separating the user's wake-up word voice signal in such a complex environment, resulting in an increase in the false trigger rate. Therefore, how to improve the wake-up accuracy of smart glasses in complex noisy environments while ensuring the real-time and power consumption performance of the system has become a technical problem that needs to be solved urgently. Summary of the invention

[0005] The object of the present invention is to provide a voice interaction method for smart glasses, which significantly improves the accuracy, robustness and user experience of wake-up word detection of smart glasses, while reducing system resource consumption, and provides reliable technical support for voice interaction of smart glasses to solve at least one of the above-mentioned prior art problems.

[0006] In a first aspect, the present invention provides a voice interaction method for smart glasses, the method specifically comprising:

[0007] A convolutional neural network model based on the attention mechanism is used to extract wake-up words from ambient voice data to obtain wake-up word recognition results;

[0008] Verifying the wake-up word recognition result through a recurrent neural network model to obtain a wake-up word verification result, and correcting the wake-up word recognition result according to the wake-up word verification result to obtain a target wake-up word recognition result;

[0009] Adopting adaptive noise suppression technology to estimate and model environmental noise in real time, and obtaining a noise detection model, wherein the noise detection model is used to automatically adjust the parameters of the convolutional neural network model so that the convolutional neural network model dynamically adapts to different noise environments;

[0010] Voiceprint recognition technology is used to extract the voiceprint features of the user when registering, a voiceprint recognition model is constructed according to the voiceprint features, and the smart glasses are voice-woken up through the voiceprint recognition model and the target wake-up word recognition result.

[0011] In a second aspect, the present invention provides a voice interaction system for smart glasses, the system specifically comprising:

[0012] The first voice interaction module is used to extract wake-up words from environmental voice data using a convolutional neural network model based on an attention mechanism to obtain a wake-up word recognition result;

[0013] A second voice interaction module is used to verify the wake-up word recognition result through a recurrent neural network model to obtain a wake-up word verification result, and to correct the wake-up word recognition result according to the wake-up word verification result to obtain a target wake-up word recognition result;

[0014] A third voice interaction module is used to use adaptive noise suppression technology to estimate and model environmental noise in real time to obtain a noise detection model, and the noise detection model is used to automatically adjust the parameters of the convolutional neural network model so that the convolutional neural network model dynamically adapts to different noise environments;

[0015] The fourth voice interaction module is used to use voiceprint recognition technology to extract the voiceprint features of the user when registering, build a voiceprint recognition model according to the voiceprint features, and perform voice wake-up on the smart glasses through the voiceprint recognition model and the target wake-up word recognition result.

[0016] In a third aspect, the present invention provides a computer device, comprising: a memory and a processor and a computer program stored in the memory, and when the computer program is executed on the processor, a voice interaction method for smart glasses as described in any one of the above methods is implemented.

[0017] In a fourth aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, and when the computer program is executed by a processor, the voice interaction method for smart glasses as described in any one of the above methods is implemented.

[0018] Compared with the prior art, the present invention has at least one of the following technical effects:

[0019] 1. The present invention significantly improves the accuracy, robustness and user experience of wake-up word detection in smart glasses, while reducing system resource consumption, providing reliable technical support for voice interaction of smart glasses.

[0020] 2. By introducing the attention mechanism, the present invention can more accurately focus on the key wake-up word features in the voice signal, improve the accuracy and robustness of wake-up word recognition, and effectively extract wake-up words even in complex noisy environments, thereby improving the accuracy and efficiency of voice wake-up.

[0021] 3. By training the recurrent neural network model, the present invention can learn the temporal dependencies in the speech signal, further improve the accuracy of wake-up word recognition, and reduce the word error rate, providing a reliable basis for subsequent verification and correction.

[0022] 4. The present invention adopts adaptive noise suppression technology to estimate and model the environmental noise in real time, and automatically adjusts the parameters of the convolutional neural network model so that the model can dynamically adapt to different noise environments and improve the anti-interference ability and robustness of voice wake-up.

[0023] 5. The present invention can accurately identify the user's voiceprint characteristics through voiceprint recognition technology and compare it with the current environment voice data to ensure that only registered users can successfully wake up the smart glasses, thereby improving the security and accuracy of voice interaction.

[0024] 6. The present invention can intelligently judge the user's interaction status, avoid erroneous triggering of voice wake-up operation in a discontinuous interaction state, and improve the intelligence level of voice interaction and user experience.

[0025] 7. The present invention can continuously learn and improve the performance of the voice wake-up algorithm through the use of user feedback data and model optimization processing, thereby improving the accuracy of the wake-up results and user satisfaction, while reducing the false trigger rate and power consumption. BRIEF DESCRIPTION OF THE DRAWINGS

[0026] In order to more clearly illustrate the technical solutions in the embodiments of the present application, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present application. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.

[0027] Figure 1 It is a flowchart of a voice interaction method of smart glasses provided by an embodiment of the present invention;

[0028] Figure 2 is a structural schematic diagram of a voice interaction system for smart glasses provided by an embodiment of the present invention;

[0029] Figure 3 It is a structural schematic diagram of a computer device provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0030] In the following description, specific details such as specific system structures, technologies, etc. are provided for the purpose of illustration rather than limitation, so as to provide a thorough understanding of the embodiments of the present application. However, it should be clear to those skilled in the art that the present application may also be implemented in other embodiments without these specific details. In other cases, detailed descriptions of well-known systems, devices, circuits, and methods are omitted to prevent unnecessary details from obstructing the description of the present application.

[0031] It should be understood that when used in the present specification and the appended claims, the term "comprising" indicates the presence of described features, wholes, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, wholes, steps, operations, elements, components and / or combinations thereof.

[0032] It should also be understood that the term “and / or” used in the specification and appended claims refers to any and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0033] As used in the specification and appended claims of this application, the term "if" can be interpreted as "when" or "uponce" or "in response to determining" or "in response to detecting", depending on the context. Similarly, the phrase "if it is determined" or "if [described condition or event] is detected" can be interpreted as meaning "uponce it is determined" or "in response to determining" or "uponce [described condition or event] is detected" or "in response to detecting [described condition or event]", depending on the context.

[0034] In addition, in the description of the present application specification and the appended claims, the terms "first", "second", "third", etc. are only used to distinguish the descriptions and cannot be understood as indicating or implying relative importance.

[0035] References to "one embodiment" or "some embodiments" etc. described in the specification of this application mean that one or more embodiments of the present application include specific features, structures or characteristics described in conjunction with the embodiment. Therefore, the statements "in one embodiment", "in some embodiments", "in some other embodiments", "in some other embodiments", etc. that appear in different places in this specification do not necessarily refer to the same embodiment, but mean "one or more but not all embodiments", unless otherwise specifically emphasized in other ways. The terms "including", "comprising", "having" and their variations all mean "including but not limited to", unless otherwise specifically emphasized in other ways.

[0036] In the embodiment of the present application, the execution subject of the process includes a terminal device, which includes but is not limited to: a server, a computer, a smart phone, a tablet computer, and other devices capable of executing the method disclosed in the present application. Figure 1 A flow chart of a voice interaction method for smart glasses disclosed in an embodiment of the present invention is shown, and the details are as follows:

[0037] S101, using a convolutional neural network model based on the attention mechanism to extract wake-up words from environmental voice data to obtain wake-up word recognition results.

[0038] In this embodiment, first, a large amount of environmental voice data is collected, which includes user voices in various noise backgrounds, such as street noise, office conversations, wind noise, etc. At the same time, we also annotate these voice data to clarify the location and content of the wake-up word.

[0039] Next, a convolutional neural network model based on the attention mechanism was constructed, which includes an input layer, a convolution layer, an attention mechanism layer, a fully connected layer, and an output layer. The input layer is responsible for receiving environmental voice data and converting it into a format suitable for model processing; the convolution layer is responsible for extracting multi-scale features in the voice signal; the attention mechanism layer calculates the attention weights of each wake-up word feature based on these features and determines the key wake-up word features; the fully connected layer classifies the key wake-up word features and obtains the wake-up word recognition results; the output layer is responsible for outputting the recognition results.

[0040] Then, the model is trained using the prepared environmental voice data. During the training process, we continuously adjust the model's parameters to minimize the recognition error rate. At the same time, we also use methods such as cross-validation to evaluate the performance of the model to ensure that it performs well in various environments. Once the model is trained, it can be used to extract wake-up words from new environmental voice data. In actual applications, users only need to say the wake-up word, and the model can quickly recognize and respond.

[0041] In this embodiment, due to the use of a convolutional neural network model based on the attention mechanism, this embodiment can more accurately capture the key wake-up word features in the voice signal, thereby improving the accuracy of wake-up word recognition. Even in a complex noise environment, the model can effectively filter out noise interference and accurately identify the user's wake-up word. It can adapt to changes in various environmental noises and maintain stable performance under different noise backgrounds. This means that users can use the voice interaction system of smart glasses in various scenarios without worrying about recognition errors caused by noise interference.

[0042] S102, verifying the wake-up word recognition result through a recurrent neural network model to obtain a wake-up word verification result, and correcting the wake-up word recognition result according to the wake-up word verification result to obtain a target wake-up word recognition result.

[0043] In this embodiment, an RNN verification model is constructed. This model can be a long short-term memory network (LSTM) or a gated recurrent unit (GRU), which are special forms of RNN and can better process long sequence data and capture time relationships. The input of the RNN verification model is the preliminary wake-up word recognition result and its corresponding speech features, and the output is a verified wake-up word recognition result. During the RNN verification process, the model captures the long-distance dependencies in the speech signal through recurrent connections and hidden layer states based on the input speech features and the preliminary wake-up word recognition results. This mechanism enables the RNN to more accurately judge the correctness of the recognition results and give the verified results. Finally, the preliminary wake-up word recognition results are corrected according to the results output by the RNN verification model. If the verification result is inconsistent with the original recognition result, we will use the verification result as the final wake-up word recognition result; if consistent, the original recognition result remains unchanged. In this way, we obtain the target wake-up word recognition result.

[0044] In this embodiment, by introducing the RNN verification model, we can more accurately judge the correctness of the wake-up word recognition results. The recurrent connection and hidden layer state of RNN enable the model to capture long-distance dependencies in the voice signal, thereby improving the accuracy of recognition. The sensitivity of the RNN verification model to voice features and its ability to process long sequence data enable the system to maintain stable performance in various environments. Even in the case of high noise or poor voice quality, the system can accurately recognize the user's wake-up word.

[0045] S103, using adaptive noise suppression technology to estimate and model the environmental noise in real time to obtain a noise detection model, wherein the noise detection model is used to automatically adjust the parameters of the convolutional neural network model so that the convolutional neural network model dynamically adapts to different noise environments.

[0046] In this embodiment, first, the key algorithms in the adaptive noise suppression technology, such as spectral subtraction, Wiener filtering, wavelet noise reduction or adaptive filtering, are used to estimate the environmental noise in real time. These algorithms can dynamically adjust the filter parameters based on the statistical characteristics of the input signal, thereby minimizing the noise components in the output signal. Next, the estimated environmental noise is modeled to generate a noise detection model. This model can accurately describe the characteristics of the current noise environment and provide a basis for the subsequent adjustment of the CNN model parameters. The noise detection model is combined with the CNN model so that the CNN model can automatically adjust its parameters according to the output of the noise detection model. Specifically, when the noise environment changes, the noise detection model will capture this change in time and generate corresponding noise features. The CNN model dynamically adjusts its filter weights, convolution kernel size and other parameters according to these noise features to adapt to the new noise environment. This adjustment is real-time and can ensure that the CNN model can maintain high performance in different noise environments. Before actual deployment, the entire system is strictly verified and tested. By simulating various noise environments, we evaluated the recognition performance of the CNN model after adaptively adjusting the parameters. Based on the test results, we optimized the CNN model and adaptive noise suppression technology to improve the robustness and accuracy of the system.

[0047] In this embodiment, by adaptively adjusting the parameters of the CNN model, we can significantly reduce the interference of environmental noise on speech recognition, thereby improving recognition accuracy. This enables the intelligent in-vehicle voice assistant system to accurately recognize the user's voice commands in a high-speed car environment. Adaptive noise suppression technology can estimate and model environmental noise in real time, allowing the CNN model to dynamically adapt to different noise environments. This mechanism enhances the robustness of the system, enabling it to maintain stable performance in various complex environments.

[0048] S104, using voiceprint recognition technology to extract the voiceprint features of the user when registering, building a voiceprint recognition model according to the voiceprint features, and performing voice wake-up on the smart glasses through the voiceprint recognition model and the target wake-up word recognition result.

[0049] In this embodiment, when the user uses the smart glasses for the first time, the system will prompt the user to register. During the registration process, the user needs to repeat the preset registration voice (for example, the user's name or a specific phrase) in a quiet environment multiple times according to the system's prompts. The system will use voiceprint recognition technology to extract voiceprint features from these registration voices. These features include the spectral features, fundamental frequency features, formant features, etc. of the voice, which can uniquely identify the user's voice identity. The extracted voiceprint features will be used to build a voiceprint recognition model. This model can be implemented using algorithms such as Gaussian mixture model (GMM), support vector machine (SVM), neural network (such as deep neural network DNN), etc. During the model construction process, the system will use a large amount of voiceprint data to train the model to ensure that it can accurately identify the user's voiceprint features. When the user wants to wake up the smart glasses, they only need to say the preset target wake-up word (for example, "Hi, glasses"). The system will first use the wake-up word recognition model to detect whether the user's voice contains the target wake-up word. If the target wake-up word is detected, the system will further use the voiceprint recognition model to verify whether the user's voiceprint features match the voiceprint features extracted during registration. Only when the target wake-up word is correctly recognized and the user's voiceprint features match the voiceprint features extracted during registration, the smart glasses will be awakened and ready to receive and execute the user's subsequent voice commands.

[0050] In this embodiment, by combining voiceprint recognition technology and target wake-up word recognition results, smart glasses can ensure that only specific users can wake up and control it through voice. This greatly improves the security of the system and prevents unauthorized access and control. Voiceprint recognition technology can accurately identify the user's voiceprint characteristics and maintain a high recognition rate even in different environmental noises. This ensures that smart glasses can accurately respond to users' voice commands in various environments.

[0051] In some embodiments, in the above step S101, the convolutional neural network model based on the attention mechanism includes an input layer, a convolutional layer, an attention mechanism layer, a fully connected layer and an output layer; the convolutional neural network model based on the attention mechanism is used to extract the wake-up word from the ambient voice data to obtain the wake-up word recognition result, specifically including:

[0052] Converting the environmental speech data into time-frequency domain feature data, and inputting the time-frequency domain feature data into the input layer;

[0053] Extracting multi-scale features of the time-frequency domain feature data according to the convolution layer, capturing speech patterns of different time spans, and obtaining wake-up word features;

[0054] Calculate the attention weight of each wake-up word feature according to the attention mechanism layer to determine the key wake-up word feature;

[0055] Classify the critical wake word features according to the fully connected layer to obtain a wake word recognition result;

[0056] Output the wake word recognition result according to the output layer.

[0057] In this embodiment, the convolutional neural network model based on the attention mechanism is an advanced speech recognition technology, which is particularly suitable for wake word recognition. The core of this model lies in combining the feature extraction ability of the convolutional neural network and the key information focusing ability of the attention mechanism.

[0058] First, convert the ambient speech data into time-frequency domain feature data, which is usually done by methods such as short-time Fourier transform (STFT) or Mel-frequency cepstral coefficients (MFCC). For example, for a 2-second speech, it can be converted into a spectrogram composed of 128 Mel filters, forming a two-dimensional feature matrix of 128×200.

[0059] The convolutional layer is responsible for extracting multi-scale features. Convolutional kernels of different sizes can capture speech patterns with different time spans. For example, using convolutional kernels of 3×3, 5×5, and 7×7 to extract local, medium, and large-scale features respectively. This can simultaneously focus on speech features at the short phoneme, syllable, and word levels, which is beneficial to improving the robustness of the model.

[0060] The attention mechanism layer calculates the importance weights of each feature. Suppose 100 feature maps are extracted, and the attention mechanism will assign a weight between 0 and 1 to each feature map. For example, for the wake word "Xiaoi Bot", the model may assign higher weights to the features representing the two sounds "Xiao" and "Bot", because these two sounds play a key role in recognition.

[0061] The fully connected layer classifies the weighted features. For example, a two-layer fully connected network can be used. The first layer has 1024 neurons, and the second layer has the number of neurons equal to the number of wake word categories. If 5 different wake words need to be recognized, then the last layer has 5 neurons, and the output of each neuron represents the probability of the corresponding wake word.

[0062] The output layer is responsible for presenting the final recognition result. Usually, the softmax function is used to convert the neuron output into a probability distribution, and the category with the highest probability is selected as the recognition result. For example, if the neuron output probability corresponding to "Xiaoi Bot" is 0.85, which is much higher than that of other wake words, then the model will determine that the current input contains the wake word "Xiaoi Bot".

[0063] The advantage of this model structure is that it can fully utilize the advantages of convolutional neural networks in speech feature extraction, and can also highlight key information through the attention mechanism to improve recognition accuracy. Especially in complex environmental noise, the attention mechanism can help the model better focus on the key phonemes of the wake-up word, thereby improving anti-interference capabilities. At the same time, the extraction of multi-scale features also enables the model to adapt to the speech speed and pronunciation characteristics of different speakers, enhancing the generalization ability of the model.

[0064] In some embodiments, in the above step S102, the training step of the recurrent neural network model includes:

[0065] Collect historical voice data, and annotate the wake-up word position and wake-up word context of the historical voice data to form a training voice data set;

[0066] Extracting Mel-frequency cepstral coefficient features of the training speech data set to obtain acoustic feature sequence data;

[0067] Decoding the training speech data set using a speech recognition tool to obtain a decoded text, and performing word segmentation processing on the decoded text using a word segmentation tool to obtain word sequence data;

[0068] The acoustic feature sequence data and the word sequence data are used as input to train the recurrent neural network model to obtain a trained recurrent neural network model, and the trained recurrent neural network model is used to determine whether the speech contains a wake-up word and the word error rate of the wake-up word.

[0069] In this embodiment, according to the preset historical voice data collection rules, voice data within a certain period of time is obtained to construct an original voice data set. For the original voice data set, the position of the wake-up word in each voice data and the context information of the wake-up word are determined by manual annotation to obtain a training voice data set with annotations. The Mel frequency cepstral coefficient (MFCC) feature extraction algorithm is used to extract acoustic features from the training voice data set to obtain acoustic feature sequence data. The training voice data set is input into the speech recognition model, and the corresponding text data, i.e., the decoded text, is obtained by decoding the acoustic model and the language model. The decoded text is segmented using a pre-trained Chinese word segmentation tool to convert the text into word sequence data in units of words. The acoustic feature sequence data and the word sequence data are aligned to construct an acoustic-text correspondence to form a complete training data set. The acoustic feature sequence data and the standard word sequence data are input into the recurrent neural network model, and the trained recurrent neural network model is obtained through multiple rounds of iterative training. The target voice data to be judged is obtained, and pre-processed and acoustic feature extracted to obtain the acoustic feature sequence of the target voice. The acoustic feature sequence of the target speech is input into the trained recurrent neural network model for judgment. If the result output by the model indicates that the speech data contains the wake-up word, the speech data is judged as the wake-up speech; otherwise, the speech data is judged as non-wake-up speech. For the target speech data judged as the wake-up speech, its acoustic feature sequence is compared with the standard word sequence of the wake-up word, and the difference between the two is calculated to obtain the word error rate of the wake-up speech.

[0070] Exemplary, the first step is data collection and annotation. For example, voice samples containing the wake-up word "Xiao Ai Tongxue" can be collected, and the precise position of the wake-up word and its context can be marked. For example, in a 10-second recording, "Xiao Ai Tongxue" appears from the 3rd to the 5th second, followed by environmental noise and user instructions. This precise annotation helps the model learn the acoustic characteristics of the wake-up word and its performance in different contexts. Next is feature extraction, usually using Mel frequency cepstral coefficients (MFCC). For a 2-second voice clip, 13-dimensional MFCC features can be extracted, one frame every 10 milliseconds, and a 13×200 feature matrix is ​​obtained. This representation method not only retains the key spectral information of the speech, but also greatly reduces the amount of data, which is conducive to subsequent processing. Speech recognition decoding is the process of converting speech into text. Speech recognition tools can be used to convert speech data into a preliminary text form. Subsequently, a word segmentation tool (such as stuttering word segmentation) is used to process the text and cut the continuous string into a meaningful sequence of words. The purpose of this step is to provide the model with semantic information to help it understand the context of the wake-up word in the sentence. Recurrent neural network (RNN) model training is the core link. Taking the long short-term memory network (LSTM) as an example, a bidirectional LSTM structure can be designed. The input layer receives acoustic features and word sequences, the hidden layer contains 256 LSTM units, and the output layer uses the softmax function to predict whether each frame belongs to the wake-up word. This structure can simultaneously consider past and future contextual information to improve recognition accuracy. During the training process, the model learns to determine whether the speech contains the wake-up word and evaluates the word error rate of the wake-up word. For example, in 1,000 hours of training data, the model may achieve 98% wake-up word detection accuracy and a word error rate as low as 2%. This means that the model can not only accurately identify the existence of the wake-up word, but also correctly identify the specific content of the wake-up word, effectively reducing the probability of false wake-up and missed wake-up. The advantage of this training method is that it comprehensively utilizes acoustic and semantic information. Acoustic features capture the pronunciation pattern of the wake-up word, while word sequence data provides contextual information, enabling the model to better understand the various variants and contexts of the wake-up word in actual use. This multimodal learning method significantly improves the robustness of the model, enabling it to accurately identify the wake-up word in various complex environments.

[0071] In some embodiments, in the above step S103, the use of adaptive noise suppression technology to estimate and model the environmental noise in real time to obtain a noise detection model specifically includes:

[0072] Acquire environmental noise data, and obtain multiple frames of noise speech signals by performing frame processing on the environmental noise data;

[0073] Performing Hamming windowing on the noisy speech signal of each frame, and performing short-time Fourier transform to obtain the frequency domain characteristics of the noisy speech signal;

[0074] According to the frequency domain characteristics of the noise speech signal, the environmental noise data is estimated and statistically modeled in real time using spectral subtraction to obtain a noise detection model.

[0075] In this embodiment, noise data in the environment is obtained through a microphone array, and the noise data is converted into a digital signal for subsequent processing. The digitized noise signal is framed by a frame processing technology, each frame length is 20-30 milliseconds, and the frame shift is 10-15 milliseconds, so as to obtain multiple frames of noise speech signals. For each frame of noise speech signal, a Hamming window function is used for windowing processing to reduce spectrum leakage and improve spectrum resolution. Short-time Fourier transform is performed on each frame of noise speech signal after windowing, and the time domain signal is converted into a frequency domain signal to obtain the frequency domain characteristics of the frame of noise speech signal. According to the frequency domain feature vector of the noise speech, the spectral subtraction method is used to estimate the environmental noise in real time to obtain an estimated environmental noise spectrum. The estimated environmental noise spectrum is used as a priori knowledge to perform statistical modeling on the environmental noise data to obtain a statistical model of environmental noise. The statistical model of environmental noise is used as a basis for discrimination, and noise detection is performed on the noise speech signal. If it is detected as noise, it is determined as environmental noise, otherwise it is determined as a speech signal. According to the noise detection results, the noisy speech signal is classified to obtain the classification label of the noisy speech signal, and the classification label is used as the training data of the noise detection model. The noise detection model is trained using machine learning algorithms, such as support vector machines and Gaussian mixture models, to obtain the trained noise detection model for subsequent noise detection tasks.

[0076] Exemplarily, first obtain environmental noise data. This process can be achieved by recording environmental sounds in different scenarios. For example, noise samples are collected in typical environments such as offices, streets or homes. These samples should contain various common noises, such as air conditioning sounds, traffic noise, human voices, etc. The purpose of framing the environmental noise data is to divide the continuous audio signal into short time segments. Typically, the length of each frame is 20 to 30 milliseconds, and there is a 50% overlap between adjacent frames. This processing method is based on the assumption that speech signals have relatively stable characteristics in a short time. For example, for a 10-second environmental noise recording, with a frame length of 25 milliseconds and a frame shift of 12.5 milliseconds, about 800 frames of noisy speech signals can be obtained. Hamming windowing is to reduce spectral leakage. Applying the Hamming window function to each frame of the signal can smooth the signal edge and reduce the edge effect in the frequency domain analysis. This step is crucial to improving the accuracy of subsequent spectrum analysis. Short-time Fourier transform (STFT) converts the time domain signal to the frequency domain. For each frame of the windowed signal, its spectrum representation can be obtained by STFT. This representation method can reveal the energy distribution of the noise signal at different frequencies. For example, for a 25 millisecond speech frame, a 512-point FFT can obtain a complex spectrum of 257 frequency points. Spectral subtraction is a commonly used noise suppression technique based on the assumption that noise and speech signals are additive in the frequency domain. By estimating the power spectrum of the noise and subtracting it from the power spectrum of the noisy speech signal, a relatively pure speech signal can be obtained. In practical applications, the noise spectrum can be estimated using the speech inactive period (such as the silent segment between speech segments). The establishment of a noise detection model involves statistical modeling of noise characteristics. Common methods include Gaussian mixture model (GMM) or hidden Markov model (HMM). These models can capture the changing characteristics of noise in time and frequency. For example, a GMM with 16 Gaussian components can better describe the spectral feature distribution of office environmental noise. Through this series of processing, a model that can estimate and identify environmental noise in real time can be obtained. This model plays an important role in applications such as speech enhancement and speech recognition, and can significantly improve the performance of the system in a noisy environment.

[0077] In some embodiments, in the above step S104, the extracting the voiceprint features of the user during registration by using voiceprint recognition technology and constructing a voiceprint recognition model according to the voiceprint features specifically includes:

[0078] Acquire user voice data recorded during the registration phase when the user uses the smart glasses, and extract voiceprint feature parameters of the user voice data;

[0079] The voiceprint feature parameters are taken as input and trained using a Gaussian mixture model to obtain a voiceprint recognition model. The voiceprint recognition model is used to calculate the likelihood ratio of the current environmental voice data and the voiceprint feature parameters, and compare the likelihood ratio with a preset likelihood ratio threshold to determine whether it is the registered user himself.

[0080] In this embodiment, voice data recorded by the user through the smart glasses during the registration stage is obtained, and voiceprint feature parameters of the voice data are extracted; the extracted voiceprint feature parameters are used as input, and the voiceprint feature parameters are trained using a Gaussian mixture model to obtain a model for voiceprint recognition; when using the smart glasses, voice data in the current environment is obtained, and the voiceprint feature parameters of the voice data are extracted; the voiceprint feature parameters of the current voice data are input into the voiceprint recognition model, and the likelihood ratio of the current voice data to the voiceprint feature parameters of the registered user is calculated; it is determined whether the calculated likelihood ratio is greater than a preset likelihood ratio threshold, and if so, it is determined that the current voice data is the voice of the registered user himself; if the likelihood ratio is less than or equal to the preset threshold, it is determined that the current voice data is not the voice of the registered user himself, and is further determined to be the voice of another user; according to the judgment result, the functional modules of the smart glasses are controlled to achieve the distinction between registered users and non-registered users, thereby improving the security and personalized experience of the smart glasses.

[0081] Exemplarily, it is first necessary to obtain the voice data recorded by the user during the registration phase, which usually includes the user reading a preset text or speaking freely, and the duration may vary from 10 to 30 seconds. For example, the user may be asked to read a sentence such as "Smart glasses are my personal assistant" many times. Extracting voiceprint feature parameters from the acquired voice data is a key step. Common features include Mel frequency cepstral coefficients (MFCC), linear prediction coefficients (LPC), etc. Taking MFCC as an example, the system may extract 13-dimensional MFCC features, plus its first-order and second-order differences, to form a 39-dimensional feature vector. These features can effectively capture the speaker's vocal tract characteristics and pronunciation habits. Gaussian mixture model (GMM) is a widely used statistical model in voiceprint recognition. It can be regarded as a weighted sum of multiple Gaussian distributions and can describe complex probability distributions. During the training process, the system may use the expectation maximization (EM) algorithm to estimate the parameters of GMM. Typically, a voiceprint model may contain 16 to 256 Gaussian components, and the specific number depends on the amount of training data and the complexity requirements of the system. After training, the voiceprint recognition model is used to calculate the likelihood ratio of the current environment voice data and the registered user's voiceprint features. This process is actually comparing whether the current voice data is more likely to come from the registered user or other speakers. The calculation of the likelihood ratio involves evaluating the probability of the current voice data under the user model and the general background model. The preset likelihood ratio threshold is the key to the system's judgment of the speaker's identity. The setting of this threshold requires a balance between the security and usability of the system. A higher threshold will improve the security of the system, but may increase the false rejection rate; a lower threshold will improve the usability of the system, but may increase the false recognition rate. In actual applications, the threshold may be dynamically adjusted according to the usage scenario. For example, when performing sensitive operations, the system may increase the threshold to ensure higher security. The application of this voiceprint recognition technology in smart glasses can greatly improve the security and personalization of the device. For example, when a user says "open my mailbox", the system will first perform voiceprint verification, and only when it is confirmed that it is the registered user will the corresponding operation be performed. This not only prevents unauthorized access, but also accurately identifies the instructions of legitimate users in a multi-person environment, improving the user experience.

[0082] In some embodiments, in the above steps S101 to S104, the method further includes:

[0083] If it is detected that the current ambient voice data contains a wake-up word, the current time point is obtained, and the current time point is compared with the time point of the last user interaction to determine the time difference;

[0084] Compare the time difference with a preset time difference threshold, and if the time difference is greater than the preset time difference threshold, determine it as a discontinuous interaction state, otherwise it is a continuous interaction state;

[0085] If the non-continuous interaction state is currently in place, context information related to the wake-up word is obtained, and it is determined whether to trigger a voice wake-up operation on the smart glasses based on the context information.

[0086] In this embodiment, the ambient voice data is analyzed in real time by a speech recognition model to determine whether it contains a preset wake-up word; if the wake-up word is detected, the current time point is obtained, and the time point of the last user interaction is obtained from the interaction record; the time difference between the current time point and the last interaction time point is calculated, and the time difference is compared with a preset time difference threshold; if the time difference is greater than the preset threshold, the interaction state is marked as a discontinuous interaction state, otherwise it is marked as a continuous interaction state; in the discontinuous interaction state, relevant context information is obtained from the knowledge graph according to the wake-up word to construct a context semantic vector; the context semantic vector is weighted using an attention mechanism to obtain a comprehensive context semantic representation; the context semantic representation is input into a deep neural network to determine whether the conditions for voice awakening the smart glasses are met, and if so, the voice awakening operation is triggered, otherwise no awakening is performed.

[0087] For example, a user may say "Hi, smart glasses" as a wake-up word. After the system detects the wake-up word, it records the current time point and compares it with the time point of the last interaction to calculate the time difference. The calculation of this time difference is crucial to determining the interaction state. Assume that the preset time difference threshold set by the system is 5 minutes. If the user uses the wake-up word again within 3 minutes, the system will regard it as a continuous interaction state. In this case, the smart glasses may maintain the previous conversation context to make the interaction smoother. On the contrary, if the user uses the wake-up word again after 30 minutes, the system will regard it as a discontinuous interaction state and need to re-establish the conversation context. In the discontinuous interaction state, the system will obtain contextual information related to the wake-up word, which may include the user's location, current time, recent schedule, etc. For example, if the user uses the wake-up word in the office, the system may prioritize work-related functions; if it is at home at night, it may tend to prefer entertainment or leisure-related functions. The analysis of contextual information is crucial to determining whether to trigger the voice wake-up operation. Assuming that the user uses the wake-up word late at night, the system may consider that the user may have unconsciously said the wake-up word in his sleep, so it may increase the trigger threshold or require additional confirmation. On the contrary, if the user has an upcoming meeting in the schedule and uses the wake-up word near the meeting room, the system may be more likely to trigger the wake-up operation and proactively provide meeting-related information. This context-based intelligent wake-up mechanism can significantly improve the user experience. For example, when the user uses the wake-up word while driving, the system may automatically start the navigation function; when the user uses the wake-up word in the gym, the system may prepare the fitness data recording function. This intelligent response not only improves the ease of use of the device, but also reduces unnecessary interaction steps, allowing users to focus more on the current task. In addition, this mechanism can also improve the safety of the system. By analyzing contextual information, the system can better identify potential false triggering situations and avoid performing sensitive operations at inappropriate times. For example, if the system detects that the user is talking to others, even if the wake-up word is heard, it may choose not to respond immediately to avoid interrupting important conversations. In general, this voice wake-up mechanism based on time difference and contextual information enables smart glasses to respond to user needs more intelligently, safely, and humanely, thereby providing a more personalized and efficient user experience.

[0088] In some embodiments, in the above steps S101 to S104, the method further includes:

[0089] Obtaining each awakening result generated by the user in the process of using the smart glasses, as well as feedback data of the user on each awakening result;

[0090] According to the feedback data, each awakening result is marked to form an awakening result data set;

[0091] The wake-up result data set is used as input, and a logistic regression model is used for training to obtain a wake-up result feedback model;

[0092] Based on the wake-up result feedback model, the convolutional neural network model of the attention mechanism and the recurrent neural network model are optimized.

[0093] In this embodiment, the original data generated by the user during use is obtained through smart glasses, including multimodal data such as wake-up voice, image, video, etc. The obtained original data is preprocessed to remove invalid information such as noise and redundancy, and extract key features. The preprocessed data is input into the wake-up model for reasoning to determine whether it is an effective wake-up and obtain the wake-up result. The user's feedback data on the wake-up result is obtained, including feedback information such as whether the user is satisfied and whether it is accurate. The wake-up result is labeled according to the user feedback, and the wake-up result and the corresponding feedback label are combined into a data pair. The labeled wake-up result data pair is divided into a training set, a validation set, and a test set according to a certain ratio to construct a wake-up result data set. The wake-up result data set is preprocessed to extract feature vectors. The extracted feature vectors are input into the logistic regression model, and the wake-up result feedback model is obtained through training. According to the wake-up result feedback model, the parameters of the convolutional neural network model of the attention mechanism are adjusted to optimize the network structure. The output of the optimized convolutional neural network model is used as the input of the recurrent neural network model to construct a cascade model. The cascade model is trained using the back propagation algorithm, and the model parameters are fine-tuned through error feedback. During the training process, the weight distribution of the attention mechanism is dynamically adjusted to enhance the model's ability to capture key features. Continuously iterate and optimize until the model converges, and obtain an excellent wake-up result discrimination model for subsequent wake-up result analysis and application.

[0094] Exemplarily, each time a user uses a wake-up word, the system records the wake-up result, including whether the recognition is successful, the response time, etc. At the same time, the system collects user feedback data, such as whether the user needs to repeat the wake-up word, whether the wake-up operation is manually canceled, etc. These data are used to annotate each wake-up result to form a comprehensive wake-up result data set. For example, suppose a user uses the wake-up word "Hi, smart glasses" in a quiet office environment, and the system responds quickly and accurately executes the user's instructions. This wake-up will be marked as a positive result. On the contrary, if the user uses the same wake-up word on a noisy street, the system does not respond or misunderstands the instruction, and this wake-up will be marked as a negative result. By inputting these annotated data into the logistic regression model for training, a model that can predict the quality of the wake-up result can be obtained. This model can take into account multiple factors, such as the ambient noise level, the clarity of the user's speech, the current time, etc. Through this model, the system can better understand which factors affect the wake-up effect, thereby optimizing the wake-up strategy. For example, the model may find that the user's wake-up success rate is low between 6 and 8 in the morning, perhaps because the user has just gotten up and the voice is not clear enough. Based on this finding, the system can lower the threshold for wake-up word recognition during this time period and improve recognition sensitivity. Using the trained wake-up result feedback model, the system can further optimize the convolutional neural network model and the recurrent neural network model of the attention mechanism. Guided by the feedback model, the parameters of these neural networks can be adjusted to better adapt to the wake-up needs in different scenarios. For example, the feedback model may indicate that in an outdoor environment, the system is too sensitive to sounds in certain frequency ranges, resulting in an increased false wake-up rate. Based on this information, the convolution kernel parameters in the convolutional neural network can be adjusted to reduce sensitivity to these frequencies. At the same time, the recurrent neural network can be optimized to better capture the temporal characteristics of wake-up words in different environments and improve verification accuracy. This continuous optimization process enables the voice wake-up system of smart glasses to continuously adapt to user habits and various complex environments. Over time, the system may find that some users tend to use a specific tone or volume to wake up the device in specific scenarios. Through personalized adjustments, the system can provide a more accurate wake-up experience for each user. In general, this optimization mechanism based on user feedback can significantly improve the user experience of smart glasses. It not only improves the accuracy and response speed of voice wake-up, but also reduces false wake-ups and missed wake-ups, making the device more intelligent and user-friendly. The advantage of this method is that it can continuously learn and adapt, so that smart glasses become more and more "understanding" of users during long-term use, thereby providing more personalized and efficient services.

[0095] Reference Figure 2 An embodiment of the present invention provides a voice interaction system 2 for smart glasses, and the system 2 specifically includes:

[0096] The first voice interaction module 201 is used to extract the wake-up word from the ambient voice data using a convolutional neural network model based on an attention mechanism to obtain a wake-up word recognition result;

[0097] The second voice interaction module 202 is used to verify the wake-up word recognition result through a recurrent neural network model to obtain a wake-up word verification result, and correct the wake-up word recognition result according to the wake-up word verification result to obtain a target wake-up word recognition result;

[0098] The third voice interaction module 203 is used to use adaptive noise suppression technology to estimate and model the environmental noise in real time to obtain a noise detection model, and the noise detection model is used to automatically adjust the parameters of the convolutional neural network model so that the convolutional neural network model dynamically adapts to different noise environments;

[0099] The fourth voice interaction module 204 is used to use voiceprint recognition technology to extract the voiceprint features of the user when registering, build a voiceprint recognition model according to the voiceprint features, and perform voice wake-up on the smart glasses through the voiceprint recognition model and the target wake-up word recognition result.

[0100] It is understandable that if Figure 1 The contents of the voice interaction method embodiment of the smart glasses shown in the figure are applicable to the voice interaction system embodiment of the smart glasses. The functions specifically implemented by the voice interaction system embodiment of the smart glasses are similar to those in the embodiment of the voice interaction method of the smart glasses shown in the figure. Figure 1 The voice interaction method embodiment of the smart glasses shown in the figure is the same as that of the embodiment of the smart glasses shown in the figure, and the beneficial effects achieved are the same as those of the embodiment of the smart glasses shown in the figure. Figure 1 The beneficial effects achieved by the embodiment of the voice interaction method for the smart glasses shown are also the same.

[0101] It should be noted that the information interaction, execution process and other contents between the above-mentioned systems are based on the same concept as the embodiment of the method of the present invention. Their specific functions and technical effects can be found in the method embodiment part and will not be repeated here.

[0102] Those skilled in the art can clearly understand that, for the convenience and simplicity of description, only the division of the above-mentioned functional units and modules is used as an example for illustration. In practical applications, the above-mentioned function allocation can be completed by different functional units and modules as needed, that is, the internal structure of the system can be divided into different functional units or modules to complete all or part of the functions described above. The functional units and modules in the embodiment can be integrated in a processing unit, or each unit can exist physically separately, or two or more units can be integrated in one unit. The above-mentioned integrated unit can be implemented in the form of hardware or in the form of software functional units. In addition, the specific names of the functional units and modules are only for the convenience of distinguishing each other, and are not used to limit the scope of protection of this application. The specific working process of the units and modules in the above-mentioned system can refer to the corresponding process in the aforementioned method embodiment, which will not be repeated here.

[0103] Reference Figure 3 The embodiment of the present invention further provides a computer device 3, comprising: a memory 302 and a processor 301 and a computer program 303 stored in the memory 302. When the computer program 303 is executed on the processor 301, the voice interaction method of the smart glasses described in any one of the above methods is implemented.

[0104] The computer device 3 may be a computing device such as a desktop computer, a notebook, a PDA, or a cloud server. The computer device 3 may include, but is not limited to, a processor 301 and a memory 302. Those skilled in the art will appreciate that Figure 3 It is only an example of computer device 3 and does not constitute a limitation on computer device 3. It may include more or fewer components than shown in the figure, or a combination of certain components, or different components, for example, it may also include input and output devices, network access devices, etc.

[0105] The processor 301 may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSP), application-specific integrated circuits (ASIC), field-programmable gate arrays (FPGA) or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. A general-purpose processor may be a microprocessor or any conventional processor, etc.

[0106] In some embodiments, the memory 302 may be an internal storage unit of the computer device 3, such as a hard disk or memory of the computer device 3. In other embodiments, the memory 302 may also be an external storage device of the computer device 3, such as a plug-in hard disk, a smart memory card (Smart Media Card, SMC), a secure digital (Secure Digital, SD) card, a flash card (Flash Card), etc. equipped on the computer device 3. Further, the memory 302 may also include both an internal storage unit of the computer device 3 and an external storage device. The memory 302 is used to store an operating system, an application program, a boot loader (Boot Loader), data, and other programs, such as the program code of the computer program, etc. The memory 302 may also be used to temporarily store data that has been output or is to be output.

[0107] An embodiment of the present invention further provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, the voice interaction method for smart glasses as described in any one of the above methods is implemented.

[0108] In this embodiment, if the integrated unit is implemented in the form of 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 present application implements all or part of the processes in the above-mentioned embodiment method, which can be completed by instructing the relevant hardware through a computer program, and the computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor, the steps of the above-mentioned various method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in source code form, object code form, executable file or some intermediate form. The computer-readable medium can at least include: any entity or device that can carry the computer program code to the camera / terminal device, recording medium, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electric carrier signal, telecommunication signal and software distribution medium. For example, a USB flash drive, a mobile hard disk, a disk or an optical disk. In some jurisdictions, according to legislation and patent practice, computer-readable media cannot be electric carrier signals and telecommunication signals.

[0109] In the above embodiments, the description of each embodiment has its own emphasis. For parts that are not described or recorded in detail in a certain embodiment, reference can be made to the relevant descriptions of other embodiments.

[0110] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0111] In the embodiments disclosed in the present application, it should be understood that the disclosed devices / terminal equipment and methods can be implemented in other ways. For example, the device / terminal equipment embodiments described above are only schematic. For example, the division of the modules or units is only a logical function division. There may be other division methods in actual implementation, such as multiple units or components can be combined or integrated into another system, or some features can be ignored or not executed. Another point is that the mutual coupling or direct coupling or communication connection shown or discussed can be through some interfaces, indirect coupling or communication connection of devices or units, which can be electrical, mechanical or other forms.

[0112] The units described as separate components may or may not be physically separated, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed on multiple network units. Some or all of the units may be selected according to actual needs to achieve the purpose of the solution of this embodiment.

Claims

1. A voice interaction method for smart glasses, characterized in that: The method specifically comprises: A convolutional neural network model based on the attention mechanism is used to extract wake-up words from ambient voice data to obtain wake-up word recognition results; Verifying the wake-up word recognition result through a recurrent neural network model to obtain a wake-up word verification result, and correcting the wake-up word recognition result according to the wake-up word verification result to obtain a target wake-up word recognition result; Adopting adaptive noise suppression technology to estimate and model environmental noise in real time, and obtaining a noise detection model, wherein the noise detection model is used to automatically adjust the parameters of the convolutional neural network model so that the convolutional neural network model dynamically adapts to different noise environments; Voiceprint recognition technology is used to extract the voiceprint features of the user when registering, a voiceprint recognition model is constructed according to the voiceprint features, and the smart glasses are voice-woken up through the voiceprint recognition model and the target wake-up word recognition result.

2. The method according to claim 1, characterized in that The convolutional neural network model based on the attention mechanism includes an input layer, a convolutional layer, an attention mechanism layer, a fully connected layer and an output layer; the convolutional neural network model based on the attention mechanism is used to extract the wake-up word from the ambient voice data to obtain the wake-up word recognition result, which specifically includes: Converting the environmental speech data into time-frequency domain feature data, and inputting the time-frequency domain feature data into the input layer; Extracting multi-scale features of the time-frequency domain feature data according to the convolution layer, capturing speech patterns of different time spans, and obtaining wake-up word features; Calculate the attention weight of each wake-up word feature according to the attention mechanism layer to determine the key wake-up word feature; Classifying the key wake-up word features according to the fully connected layer to obtain a wake-up word recognition result; The wake-up word recognition result is output according to the output layer.

3. The method according to claim 1, characterized in that The training steps of the recurrent neural network model include: Collect historical voice data, and annotate the wake-up word position and wake-up word context of the historical voice data to form a training voice data set; Extracting Mel-frequency cepstral coefficient features of the training speech data set to obtain acoustic feature sequence data; Decoding the training speech data set using a speech recognition tool to obtain a decoded text, and performing word segmentation processing on the decoded text using a word segmentation tool to obtain word sequence data; The acoustic feature sequence data and the word sequence data are used as input to train the recurrent neural network model to obtain a trained recurrent neural network model, and the trained recurrent neural network model is used to determine whether the speech contains a wake-up word and the word error rate of the wake-up word.

4. The method according to claim 1, characterized in that: The method of using the adaptive noise suppression technology to estimate and model the environmental noise in real time to obtain a noise detection model specifically includes: Acquire environmental noise data, and obtain multiple frames of noise speech signals by performing frame processing on the environmental noise data; Performing Hamming windowing on the noisy speech signal of each frame, and performing short-time Fourier transform to obtain the frequency domain characteristics of the noisy speech signal; According to the frequency domain characteristics of the noise speech signal, the environmental noise data is estimated and statistically modeled in real time using spectral subtraction to obtain a noise detection model.

5. The method according to claim 1, characterized in that The method of extracting the voiceprint features of the user during registration by using the voiceprint recognition technology and constructing a voiceprint recognition model according to the voiceprint features specifically includes: Acquire user voice data recorded during the registration phase when the user uses the smart glasses, and extract voiceprint feature parameters of the user voice data; The voiceprint feature parameters are taken as input and trained using a Gaussian mixture model to obtain a voiceprint recognition model. The voiceprint recognition model is used to calculate the likelihood ratio of the current environmental voice data and the voiceprint feature parameters, and compare the likelihood ratio with a preset likelihood ratio threshold to determine whether it is the registered user himself.

6. The method according to any one of claims 1 to 5, characterized in that The method further comprises: If it is detected that the current ambient voice data contains a wake-up word, the current time point is obtained, and the current time point is compared with the time point of the last user interaction to determine the time difference; Compare the time difference with a preset time difference threshold, and if the time difference is greater than the preset time difference threshold, determine it as a discontinuous interaction state, otherwise it is a continuous interaction state; If the non-continuous interaction state is currently in place, context information related to the wake-up word is obtained, and it is determined whether to trigger a voice wake-up operation on the smart glasses based on the context information.

7. The method according to any one of claims 1 to 5, characterized in that The method further comprises: Obtaining each awakening result generated by the user in the process of using the smart glasses, as well as feedback data of the user on each awakening result; According to the feedback data, each awakening result is marked to form an awakening result data set; The wake-up result data set is used as input, and a logistic regression model is used for training to obtain a wake-up result feedback model; Based on the wake-up result feedback model, the convolutional neural network model of the attention mechanism and the recurrent neural network model are optimized.

8. A voice interaction system for smart glasses, characterized in that: The system specifically comprises: The first voice interaction module is used to extract wake-up words from environmental voice data using a convolutional neural network model based on an attention mechanism to obtain a wake-up word recognition result; A second voice interaction module is used to verify the wake-up word recognition result through a recurrent neural network model to obtain a wake-up word verification result, and to correct the wake-up word recognition result according to the wake-up word verification result to obtain a target wake-up word recognition result; A third voice interaction module is used to use adaptive noise suppression technology to estimate and model environmental noise in real time to obtain a noise detection model, and the noise detection model is used to automatically adjust the parameters of the convolutional neural network model so that the convolutional neural network model dynamically adapts to different noise environments; The fourth voice interaction module is used to use voiceprint recognition technology to extract the voiceprint features of the user when registering, build a voiceprint recognition model according to the voiceprint features, and perform voice wake-up on the smart glasses through the voiceprint recognition model and the target wake-up word recognition result.

9. A computer device, characterized in that: include: A memory and a processor and a computer program stored in the memory, when the computer program is executed on the processor, implements the voice interaction method of the smart glasses as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: A computer program is stored thereon, and when the computer program is executed by the processor, the voice interaction method of the smart glasses as described in any one of claims 1 to 7 is implemented.

Images