A voice-interactive graphene AI smart coffee table control method
By incorporating a built-in microphone array and training noise scene dataset into the smart coffee table, combined with a dialect recognition model and cloud-based text parsing, the accuracy of voice interaction in complex environments has been solved, enabling convenient voice control.
Patent Information
- Application Number
- CN202512030378.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2045-12-30
AI Technical Summary
The existing voice interaction methods of smart coffee tables have low recognition accuracy in complex environments and cannot adapt to different dialects and accents, resulting in inconvenient interaction.
Preliminary noise reduction is performed using a built-in microphone array. The model's adaptability is enhanced by training on a dataset of noisy scenes. Data sets of dialects and accents from different regions are collected to train the recognition model. Text parsing and intent analysis are performed on a cloud server to ensure accurate recognition and execution of commands.
It improves the quality of voice acquisition, reduces the false trigger rate, expands the applicable population, ensures the reliability of recognition and the accuracy of interaction in complex environments, and provides a convenient natural and intelligent interactive experience.
Smart Images

Figure CN121963725B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of home voice interaction technology, specifically to a voice-interactive graphene AI smart coffee table control method. Background Technology
[0002] With the rapid development of smart home technology, various smart terminal devices are gradually being integrated into people's daily lives. Traditional coffee tables only have the single function of carrying items, while modern graphene coffee tables are a new type of home furnishing product that integrates the characteristics of graphene materials with intelligent functions. They not only retain the function of traditional coffee tables as the core furniture of the living room, but also expand more practical value through the unique advantages of graphene materials and the integration of intelligent technology. They have gradually integrated touch operation, information display and environmental monitoring functions, but there are still many limitations in human-computer interaction. Currently, most mainstream smart coffee tables rely on physical buttons or touchscreens for interaction, requiring users to be in close proximity to the device to operate it, which is extremely inconvenient when their hands are occupied or relaxed. Voice interaction, as a natural and convenient human-computer interaction method, has been widely used in smartphones and smart speaker devices, but its application in specific scenarios such as smart coffee tables still faces many technical challenges. Environmental noise is a significant factor affecting the accuracy of voice recognition. Sounds from televisions and conversations around us can interfere with voice commands, leading to inaccurate control commands received by the smart coffee table and hindering effective voice interaction. Furthermore, the user's pronunciation, dialect, or accent can also prevent the voice recognition system from accurately recognizing all commands. The recognition rate varies considerably among people from different regions, resulting in recognition errors and hindering effective interaction. Therefore, there is an urgent need to propose a voice-interactive AI smart coffee table control method. Summary of the Invention
[0003] To address the aforementioned technical issues, a voice-interactive graphene AI smart coffee table control method is provided, which solves the problems described above.
[0004] To achieve the above objectives, the technical solution adopted by this invention is: a voice-interactive graphene AI smart coffee table control method, the control steps of which are as follows: S1. Based on the built-in microphone of the coffee table to listen to the commands of surrounding users, after the coffee table receives the voice command, it performs preliminary noise reduction processing at the same time to retain the voice segment of the command. S2. Train a dataset containing noisy scenes, allowing the built-in model to learn features that distinguish speech from noise and filter non-speech signals in real time. S3. Collect large-scale datasets for different regional dialects and accents, train dialect recognition models, input the acquired speech signals into the dialect recognition model to identify and judge the current dialect and accent, and perform text conversion. S4. Parse the converted text, analyze the intent and instructions. If the instructions are ambiguous, actively ask the user to clarify their needs. If there are multiple intents, break them down into sub-instructions and execute them sequentially. Pass the instructions to the corresponding functional modules to perform interactive operations.
[0005] Preferably, in step S1, the smart coffee table has a built-in microphone array distributed along its edges and bottom. The noise reduction process uses the hardware's built-in low-pass filter to filter out high-frequency electromagnetic interference above 16kHz and low-frequency vibration noise below 20Hz, retaining signals in the 200Hz-8kHz frequency band.
[0006] Preferably, the specific steps in step S2 are as follows: During the training phase, a noisy speech dataset is first constructed. Clean speech and various noisy sample data are collected and mixed with signal-to-noise ratios ranging from -5dB to 20dB to generate samples. The training, validation and test sets are divided into training, validation and test sets in a 7:2:1 ratio, and the speech region timestamps are labeled. The audio is converted into frequency domain features, resulting in a two-dimensional feature matrix that intuitively reflects the spectral differences between speech and noise. Choosing CNN as the model architecture, the model was trained to minimize the loss and achieve the highest accuracy in predicting speech frames as 1 and noise frames as 0. The Adam optimizer was used to avoid overfitting through early stopping strategy and data augmentation. The speech recognition rate and noise filtering rate were used as evaluation metrics to evaluate the model output. The trained model is deployed to the coffee table system for real-time filtering. The audio stream is collected and cut into segments with a fixed frame length. Features in the two-dimensional feature matrix are extracted and input into the model to obtain frame-by-frame speech probabilities. After smoothing through a sliding window, a threshold is used to determine the speech and noise regions, and the denoised speech is output.
[0007] Preferably, the duration of speech in the sample data is 1-10 seconds. The mixed samples are generated by superimposing speech and noise in the time domain using an audio processing library. The superposition formula is as follows: Mixed signal = Clean speech signal * Superposition factor 1 + Noise signal * Superposition factor 2 The superposition coefficient is obtained by setting the target signal-to-noise ratio according to the scene, calculating the sum of the energy of clean speech and noise, fixing the speech coefficient, and deriving the noise coefficient; the clean speech and noise are quantized by converting the continuous analog signal of the collected sample data into a discrete digital signal to obtain the quantized value, which is the clean speech signal and the noise signal.
[0008] Preferably, the steps for obtaining the two-dimensional feature matrix are as follows: The continuous audio in the mixed samples is segmented into short frames of 20-30ms, with a 10ms overlap between frames; A Hanning window is applied to each frame of signal, and a fast Fourier transform is performed on each windowed frame to convert the time-domain signal into the frequency domain, thereby obtaining the amplitude of each frequency component and forming the spectrum. By using a Mel filter bank, the nonlinear perception of frequency by the human ear is simulated, and the spectrum is mapped to the Mel frequency scale, preserving key frequency band information of speech. Taking the logarithm of the Mel spectrum compresses the dynamic range and enhances the discriminability of low-energy features, resulting in a two-dimensional feature matrix. In the matrix, speech exhibits regular spectral stripes, while noise shows a disordered distribution.
[0009] Preferably, the specific steps in step S3 are as follows: The included dialect types were clearly defined, and voice data was collected based on recording equipment. The collected voice data was standardized into WAV and MP3 formats, and the sampling rate was set to 16kHz. Data with labeling errors and poor audio quality are removed, and speech features are extracted. The speech features are obtained after processing by framing, windowing, Fourier transform, Mel filtering and discrete cosine transform. The processed data is divided proportionally for model training, validation, and evaluation. The model is then optimized, and the optimization steps are the same as those in step S2, where the built-in model learns to distinguish speech from noise features. The optimized model is deployed on a cloud server. After acquiring the speech signal, it is preprocessed and then input into the dialect recognition model. The model outputs the recognition result, determines the dialect and accent of the current speech, and converts the speech into the corresponding text content.
[0010] Preferably, the cloud server is an 8-core, 16GB memory cloud server. The optimized model is stored in ONNX format, and the model initialization time is less than 3 seconds. Preprocessing includes format verification and conversion. After receiving the audio, the format is first verified. If it is MP3 format, it is converted to WAV format in real time using FFmpeg. CRC verification is enabled during the conversion process to ensure data integrity. After inputting into the dialect recognition model, the long-term dependencies of the feature sequence are captured by a 6-layer self-attention mechanism. A 15-dimensional logits are output through a fully connected layer. The Softmax function is applied to convert the logits into a probability distribution, satisfying that the sum of all probabilities is 1. The top 3 candidate categories with the highest probabilities and their probability values are extracted. If the highest probability is ≥0.6, the corresponding category is directly used as the recognition result. If the highest probability is <0.6, a second inference is triggered. The 500ms segment with the highest energy in the audio is extracted again to re-extract features. The input is fed into the model to calculate the probability again. The highest average probability value of the two inferences is taken as the final category.
[0011] Preferably, the step of analyzing the statement intent and instruction information in step S4 is as follows: Clean up redundant symbols, retain core semantic components, unify encoding, and map dialect words to general semantics; Perform word segmentation, split the text using a dialect adaptation tool, and handle adjoining words; label each word segmentation result with part-of-speech tags, with a focus on marking the part-of-speech tags related to instructions. Construct an intent classification system and pre-define common intent categories to cover home control, information query and social interaction scenarios. Basic intents include: turn on the lights, turn off the lights, adjust the temperature, check the weather and play music. Input a segmented text sequence and output the intent probability distribution, then use the context to resolve ambiguity. Extract key information from the identified instruction intent, including the target of the operation, the operation action, attribute parameters, and the time. For missing elements, the missing requirements are filled in by actively querying; for complex instructions, the interactive operations are executed step by step after being broken down, and finally the results are output in a structured manner for intelligent control.
[0012] Preferably, redundant symbols include modal particles and repetition symbols. The word segmentation process is as follows: For the standardized dialect text, a word segmentation tool is used for processing. The word segmentation tool combines the collected dialect samples and uses the bidirectional maximum matching method to split the text. Starting from the left and right sides of the text, it prioritizes matching the longest valid words in the vocabulary, and then compares the splitting results in the two directions. The sequence with fewer words is selected as the initial word segmentation result. For adjoining words not included in the lexicon, identify the basic morphemes contained in the adjoining words and split the adjoining words into two or more words that can independently express part of the semantics; Each word segmented by word is labeled with its corresponding part of speech. The labeling is based on the word’s conventional grammatical function in the dialect and references the dialect word class library in the big data. Special marking is used for core words related to instructions to highlight their functions.
[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention improves voice acquisition quality and reduces command omissions through initial noise reduction and microphone array layout. Secondly, by training on multi-noise scene datasets, it enhances the model's adaptability to complex environments, reduces false triggers, and ensures reliable recognition. Furthermore, it optimizes for dialects and accents, expanding the applicable population and improving product universality. Finally, the text parsing and intent analysis mechanism can handle ambiguous and multi-intent commands, ensuring accurate execution and precise transmission to the corresponding modules, thus improving the interactive experience. Overall, it is user-centric, achieving natural and intelligent interaction, making operation more convenient, fitting home scenarios, and bringing an efficient and user-friendly experience. Attached Figure Description
[0014] Figure 1 This is a flowchart of the control steps of the present invention. Detailed Implementation
[0015] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.
[0016] Reference Figure 1 As shown, a voice-interactive graphene AI smart coffee table control method includes the following control steps: S1. Based on the built-in microphone of the coffee table to listen to the commands of surrounding users, after the coffee table receives the voice command, it performs preliminary noise reduction processing at the same time to retain the voice segment of the command. S2. Train a dataset containing noisy scenes, allowing the built-in model to learn features that distinguish speech from noise and filter non-speech signals in real time. S3. Collect large-scale datasets for different regional dialects and accents, train dialect recognition models, input the acquired speech signals into the dialect recognition model to identify and judge the current dialect and accent, and perform text conversion. S4. Parse the converted text, analyze the intent and instructions. If the instructions are ambiguous, actively ask the user to clarify their needs. If there are multiple intents, break them down into sub-instructions and execute them sequentially. Pass the instructions to the corresponding functional modules to perform interactive operations.
[0017] First, by implementing initial noise reduction and microphone array layout, the quality of voice acquisition is improved, and command omissions are reduced. Second, by training on datasets of multiple noisy scenarios, the model's adaptability to complex environments is enhanced, false triggers are reduced, and recognition reliability is ensured. Third, optimizations for dialects and accents expand the applicable population and improve product universality. Finally, the text parsing and intent analysis mechanism can handle ambiguous and multi-intent commands, ensuring accurate execution and precise transmission to the corresponding modules, thus improving the interactive experience. Overall, the system is user-centric, achieving natural and intelligent interaction, making operation more convenient, fitting home scenarios, and bringing an efficient and user-friendly experience.
[0018] In step S1, the smart coffee table has a built-in microphone array distributed along its edges and bottom. The noise reduction process uses the hardware's built-in low-pass filter to filter out high-frequency electromagnetic interference above 16kHz and low-frequency vibration noise below 20Hz, retaining signals in the 200Hz-8kHz frequency band.
[0019] In step S1, the smart coffee table listens to user commands through a built-in microphone array. These microphones are distributed along the edges and bottom of the coffee table. This layout can capture voice information from different directions more comprehensively. The initial noise reduction processing, combined with hardware low-pass filtering, accurately filters high-frequency electromagnetic interference above 16kHz and low-frequency vibration sounds below 20Hz, retaining only the 200Hz-8kHz frequency band where human speech is mainly concentrated, providing a cleaner voice foundation for subsequent processing.
[0020] The specific steps in step S2 are as follows: During the training phase, a noisy speech dataset is first constructed. Clean speech and various noisy sample data are collected and mixed with signal-to-noise ratios ranging from -5dB to 20dB to generate samples. The training, validation and test sets are divided into training, validation and test sets in a 7:2:1 ratio, and the speech region timestamps are labeled. The audio is converted into frequency domain features, resulting in a two-dimensional feature matrix that intuitively reflects the spectral differences between speech and noise. Choosing CNN as the model architecture, the model was trained to minimize the loss and achieve the highest accuracy in predicting speech frames as 1 and noise frames as 0. The Adam optimizer was used to avoid overfitting through early stopping strategy and data augmentation. The speech recognition rate and noise filtering rate were used as evaluation metrics to evaluate the model output. The trained model is deployed to the coffee table system for real-time filtering. The audio stream is collected and cut into segments with a fixed frame length. Features in the two-dimensional feature matrix are extracted and input into the model to obtain frame-by-frame speech probabilities. After smoothing through a sliding window, a threshold is used to determine the speech and noise regions, and the denoised speech is output.
[0021] In the model training of step S2, the construction of a noisy speech dataset is crucial. The speech samples are 1-10 seconds long, covering the common length of daily instructions. Samples are generated by mixing them with a signal-to-noise ratio of -5dB to 20dB, covering a variety of environments from noisy to quiet. In the formula for calculating the mixed signal, the superposition coefficient is determined based on the target signal-to-noise ratio of the scene. By calculating the speech and noise energy and fixing the speech coefficient, the noise coefficient is derived, which can accurately control the signal-to-noise ratio after mixing. Quantization processing converts the continuous analog audio signal into a discrete digital signal that can be processed by a computer, which is convenient for subsequent feature extraction and model training. The process of obtaining the two-dimensional feature matrix involves dividing the continuous audio into short frames of 20-30ms and overlapping them by 10ms. This framing method ensures the independence of each frame while preserving the continuity between frames through overlap. Applying the Hanning window reduces spectral leakage, making the frequency domain information after Fourier transform more accurate. The Mel filter bank simulates the nonlinear perception characteristics of the human ear for frequency, highlighting key speech information. Then, taking the logarithm of the Mel spectrum further compresses the dynamic range, making low-energy speech features easier to identify. At this point, the regular spectral stripes of speech in the two-dimensional feature matrix contrast sharply with the chaotic distribution of noise, which is beneficial for model differentiation.
[0022] The speech duration in the sample data ranges from 1 to 10 seconds. The mixed samples are generated by superimposing speech and noise in the time domain using an audio processing library. The superposition formula is as follows: Mixed signal = Clean speech signal * Superposition factor 1 + Noise signal * Superposition factor 2 The superposition coefficient is obtained by setting the target signal-to-noise ratio according to the scene, calculating the sum of the energy of clean speech and noise, fixing the speech coefficient, and deriving the noise coefficient; the clean speech and noise are quantized by converting the continuous analog signal of the collected sample data into a discrete digital signal to obtain the quantized value, which is the clean speech signal and the noise signal.
[0023] The steps for obtaining the two-dimensional feature matrix are as follows: The continuous audio in the mixed samples is segmented into short frames of 20-30ms, with a 10ms overlap between frames; A Hanning window is applied to each frame of signal, and a fast Fourier transform is performed on each windowed frame to convert the time-domain signal into the frequency domain, thereby obtaining the amplitude of each frequency component and forming the spectrum. By using a Mel filter bank, the nonlinear perception of frequency by the human ear is simulated, and the spectrum is mapped to the Mel frequency scale, preserving key frequency band information of speech. Taking the logarithm of the Mel spectrum compresses the dynamic range and enhances the discriminability of low-energy features, resulting in a two-dimensional feature matrix. In the matrix, speech exhibits regular spectral stripes, while noise shows a disordered distribution.
[0024] The specific training and real-time deployment inference process of the model is as follows: Clean speech and noise are mixed together to generate diverse samples according to different SNRs. Each frame is labeled with a tag: speech frame = 1, noise frame = 0. The audio is divided into fixed frame lengths, such as 25ms / frame, and a two-dimensional frequency domain feature matrix is generated through temporal framing → windowing → FFT → Mel filtering → logarithmic transformation. Design a lightweight CNN network to adapt to real-time requirements; receive a two-dimensional frequency domain feature matrix, such as a Mel spectrogram, with dimensions [time frame, Mel band, 1], and adapt to a single-channel CNN; Multi-layer convolution (such as 3×3 convolution kernel) is used to extract local spectral features and capture the differences in speech formants and disordered noise distribution; Dimensionality reduction is achieved through max pooling / average pooling, preserving key features while reducing computational cost. The convolutional features are mapped to a binary classification, speech / noise, and the output is the frame-by-frame predicted probability (0~1). The difference between the predicted probability and the true label (1 / 0) is quantified using binary cross-entropy loss, with the goal of minimizing the loss. The Adam optimizer is used to dynamically adjust the learning rate, and during training, the audio speed (±10%), volume (±20%), and random noise are randomly adjusted to increase data diversity. The validation set loss is monitored, and if it does not decrease for 5 consecutive rounds, training is stopped and the optimal model is retained; the proportion of speech frames correctly predicted as 1; the proportion of noise frames correctly predicted as 0; The audio stream is acquired through a microphone / audio interface and cut into fixed frame lengths. The feature transformation process during training (temporal framing → windowing → FFT → Mel filtering → logarithmic transformation) is repeated for each frame of audio to generate a two-dimensional feature matrix, which is used as the model input. The feature matrix is input into the CNN model, and the output is the speech probability of each frame. The probability sequence is filtered by a sliding window of 3-5 frames to avoid misjudgment of a single frame. A probability threshold is set. If the probability is greater than the threshold, it is determined to be a speech region. Otherwise, it is a noise region. For areas identified as noise, noise is suppressed through spectral subtraction / energy attenuation; The processed audio frames are spliced into a continuous stream, and the denoised speech is output. During training, the shape of the feature matrix needs to be clearly defined to ensure that the CNN input dimension matches; during real-time inference, the dimensions of single-frame features need to be expanded to adapt to the model input. The window length needs to be adjusted according to the scenario; the smoothing method can be flexibly selected to balance response speed and false alarm rate. Fixed thresholds may not be suitable for all scenarios; adaptive thresholds can be designed to improve robustness.
[0025] Inference speed needs to be optimized through model quantization and hardware acceleration to ensure that the end-to-end latency (acquisition → processing → output) is less than 100ms, in order to meet the requirements of real-time interaction.
[0026] The specific steps in step S3 are as follows: The included dialect types were clearly defined, and voice data was collected based on recording equipment. The collected voice data was standardized into WAV and MP3 formats, and the sampling rate was set to 16kHz. Data with labeling errors and poor audio quality are removed, and speech features are extracted. The speech features are obtained after processing by framing, windowing, Fourier transform, Mel filtering and discrete cosine transform. The processed data is divided proportionally for model training, validation, and evaluation. The model is then optimized, and the optimization steps are the same as those in step S2, where the built-in model learns to distinguish between speech and noise features; these will not be elaborated upon here. The optimized model is deployed on a cloud server. After acquiring the speech signal, it is preprocessed and then input into the dialect recognition model. The model outputs the recognition result, determines the dialect and accent of the current speech, and converts the speech into the corresponding text content.
[0027] This application clearly defines the dialect types and collects data in a targeted manner to avoid interference from irrelevant data and ensure the targeted nature of model training; it unifies the audio format (WAV / MP3) and sampling rate (16kHz) to reduce the impact of format differences on model learning and ensure data consistency; it removes erroneous and low-quality data, which is equivalent to "purifying" the training data, reducing the misleading influence of noisy data on the model, and allowing the model to focus more on effective features. By extracting speech features through professional techniques such as frame segmentation and Fourier transform, the model can accurately capture key information about the unique tones and rhythms of dialects, making it easier for the model to distinguish the differences between different dialects / accents. The training, validation, and evaluation sets are divided proportionally, which conforms to the scientific process of machine learning. This makes it easy to adjust parameters through the validation set and check the effect through the evaluation set, avoiding overfitting or underfitting of the model and ensuring stable performance in real-world scenarios.
[0028] The cloud server is an 8-core, 16GB memory cloud server. The optimized model is stored in ONNX format, and the model initialization time is less than 3 seconds. Preprocessing includes format verification and conversion. After receiving the audio, the format is first verified. If it is in MP3 format, it is converted to WAV format in real time using FFmpeg. CRC check is enabled during the conversion process to ensure data integrity. After inputting into the dialect recognition model, a 6-layer self-attention mechanism is used to capture the long-term dependencies of the feature sequence. A 15-dimensional logits is output through a fully connected layer. The Softmax function is applied to convert the logits into a probability distribution, satisfying that the sum of all probabilities is 1. The top 3 candidate categories with the highest probabilities and their probability values are extracted. If the highest probability is ≥0.6, the corresponding category is directly used as the recognition result. If the highest probability is <0.6, a second inference is triggered. The 500ms segment with the highest energy in the audio is extracted again to re-extract features. The input is fed into the model to calculate the probability again. The highest average probability value of the two inferences is taken as the final category.
[0029] In the S3 step of dialect recognition, the collected speech data is uniformly in WAV and MP3 formats with a sampling rate of 16kHz to ensure data format consistency. Format verification and conversion in preprocessing are crucial. CRC verification is enabled when converting MP3 to WAV to effectively ensure the integrity of the transcoded data. An 8-core, 16GB memory cloud server provides sufficient computing power for model operation. The model initialization time for ONNX format storage is less than 3 seconds, ensuring response speed. The 6-layer self-attention mechanism can effectively capture long-term dependencies in speech feature sequences, improving the accuracy of dialect recognition. When the highest probability is <0.6, the 500ms segment with the highest energy in the audio is extracted for secondary inference. This segment usually contains the clearest speech information. The final category is determined by the average probability of the two inferences, further improving the reliability of recognition.
[0030] The steps in step S4 for analyzing statement intent and instruction information are as follows: Clean up redundant symbols, retain core semantic components, unify encoding, and map dialect words to general semantics; Perform word segmentation, split the text using a dialect adaptation tool, and handle adjoining words; label each word segmentation result with part-of-speech tags, with a focus on marking the part-of-speech tags related to instructions. Construct an intent classification system and pre-define common intent categories to cover home control, information query and social interaction scenarios. Basic intents include: turn on the lights, turn off the lights, adjust the temperature, check the weather and play music. Input a segmented text sequence and output the intent probability distribution, then use the context to resolve ambiguity. Extract key information from the identified instruction intent, including the target of the operation, the operation action, attribute parameters, and the time. For missing elements, the missing requirements are filled in by actively querying; for complex instructions, the interactive operations are executed step by step after being broken down, and finally the results are output in a structured manner for intelligent control.
[0031] Redundant symbols include modal particles and repetition symbols. The word segmentation process is as follows: For the standardized dialect text, a word segmentation tool is used for processing. The word segmentation tool combines the collected dialect samples and uses the bidirectional maximum matching method to split the text. Starting from the left and right sides of the text, it prioritizes matching the longest valid words in the vocabulary, and then compares the splitting results in the two directions. The sequence with fewer words is selected as the initial word segmentation result. For adjoining words not included in the lexicon, identify the basic morphemes contained in the adjoining words and split the adjoining words into two or more words that can independently express part of the semantics; Each word segmented by word is labeled with its corresponding part of speech. The labeling is based on the word’s conventional grammatical function in the dialect and references the dialect word class library in the big data. Special marking is used for core words related to instructions to highlight their functions.
[0032] In the S4 step's analysis of sentence intent and instruction information, cleaning redundant symbols simplifies the text and highlights the core semantics. The word segmentation process uses a bidirectional maximum matching method, combined with a dialect sample lexicon, to simultaneously match the longest effective words from both sides, and then select the segmentation result with fewer words, which can more accurately segment dialect text. For unrecorded adjoining words, they are segmented by identifying basic morphemes, ensuring the integrity of word segmentation. Based on the common categories preset by the classification system, it comprehensively covers home control, information query, and social interaction scenarios. When extracting key information, if there are missing elements, the system will actively ask for completion. For compound instructions, they are split into sub-instructions and executed sequentially to ensure the accuracy and orderliness of instruction execution, ultimately achieving precise and intelligent control of the smart coffee table.
[0033] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.
Claims
1. A method for controlling a graphene AI smart coffee table with voice interaction, characterized in that, The control steps are as follows: S1. Based on the built-in microphone of the coffee table to listen to the commands of surrounding users, after the coffee table receives the voice command, it performs preliminary noise reduction processing at the same time to retain the voice segment of the command. S2. Train the dataset containing noisy scenes, allowing the built-in model to learn features that distinguish speech from noise and filter non-speech signals in real time; the specific steps in step S2 are as follows: During the training phase, a noisy speech dataset is first constructed. This is achieved by collecting clean speech and various noise sample data and generating samples by mixing them with a signal-to-noise ratio ranging from -5dB to 20dB. The audio is then converted into frequency domain features to obtain a two-dimensional feature matrix. A CNN is selected as the model architecture to train the model, ensuring that the model has the highest accuracy in predicting 1 for speech frames and 0 for noise frames. The trained model is deployed to the coffee table system for real-time filtering. After audio acquisition, segmentation, and smoothing, the denoised speech is output. S3. Collect large-scale datasets for different regional dialects and accents, train a dialect recognition model, input the acquired speech signals into the dialect recognition model to identify and judge the current dialect and accent, and perform text conversion; the specific steps in step S3 are as follows: The included dialect types are clearly defined. Speech data is collected based on recording equipment. The optimized model is deployed on a cloud server. After acquiring the speech signal, it is preprocessed and then input into the dialect recognition model. After being input into the dialect recognition model, the long-term dependencies of the feature sequence are captured through a 6-layer self-attention mechanism. A 15-dimensional logits is output through a fully connected layer. The Softmax function is applied to convert the logits into a probability distribution, satisfying that the sum of all probabilities is 1. The top 3 candidate categories with the highest probabilities and their probability values are extracted. If the highest probability is ≥0.6, the corresponding category is directly used as the recognition result. If the highest probability is <0.6, a second inference is triggered. The 500ms segment with the highest energy in the audio is extracted again, and the probabilities are recalculated by inputting it into the model. The highest average probability value of the two inferences is taken as the final category. The model outputs the recognition result, determines the dialect and accent of the current speech, and converts the speech into the corresponding text content. S4. Parse the converted text, analyze the intent and instructions. If the instructions are ambiguous, actively ask the user to clarify their needs. If there are multiple intents, break them down into sub-instructions and execute them sequentially. Pass the instructions to the corresponding functional modules to perform interactive operations.
2. The method for controlling a graphene AI smart coffee table with voice interaction according to claim 1, characterized in that: In step S1, the smart coffee table has a built-in microphone array distributed along its edges and bottom. The noise reduction process uses the hardware's built-in low-pass filter to filter out high-frequency electromagnetic interference above 16kHz and low-frequency vibration noise below 20Hz, retaining signals in the 200Hz-8kHz frequency band.
3. The method for controlling a graphene AI smart coffee table with voice interaction according to claim 1, characterized in that: The speech duration in the sample data ranges from 1 to 10 seconds. The mixed samples are generated by superimposing speech and noise in the time domain using an audio processing library. The superposition formula is as follows: Mixed signal = Clean speech signal * Superposition factor 1 + Noise signal * Superposition factor 2; The superposition coefficient is obtained by setting the target signal-to-noise ratio according to the scene, calculating the sum of the energy of clean speech and noise, fixing the speech coefficient, and deriving the noise coefficient; the clean speech and noise are quantized by converting the continuous analog signal of the collected sample data into a discrete digital signal to obtain the quantized value, which is the clean speech signal and the noise signal.
4. The voice-interactive graphene AI smart coffee table control method according to claim 1, characterized in that, The steps in step S4 for analyzing statement intent and instruction information are as follows: Clean up redundant symbols, retain core semantic components, unify encoding, and map dialect words to general semantics; Perform word segmentation, split the text using a dialect adaptation tool, and handle adjoining words; label each word segmentation result with part-of-speech tags, with a focus on marking the part-of-speech tags related to instructions. Construct an intent classification system and pre-define common intent categories to cover home control, information query and social interaction scenarios. Basic intents include: turn on the lights, turn off the lights, adjust the temperature, check the weather and play music. Input a segmented text sequence and output the intent probability distribution, then use the context to resolve ambiguity. Extract key information from the identified instruction intent, including the target of the operation, the operation action, attribute parameters, and the time. For missing elements, the missing requirements are filled in by actively querying; for complex instructions, the interactive operations are executed step by step after being broken down, and finally the results are output in a structured manner for intelligent control.
5. The voice-interactive graphene AI smart coffee table control method according to claim 4, characterized in that: Redundant symbols include modal particles and repetition symbols. The word segmentation process is as follows: For the standardized dialect text, a word segmentation tool is used for processing. The word segmentation tool combines the collected dialect samples and uses the bidirectional maximum matching method to split the text. Starting from the left and right sides of the text, it prioritizes matching the longest valid words in the vocabulary, and then compares the splitting results in the two directions. The sequence with fewer words is selected as the initial word segmentation result. For adjoining words not included in the lexicon, identify the basic morphemes contained in the adjoining words and split the adjoining words into two or more words that can independently express part of the semantics; Each word segmented by word is labeled with its corresponding part of speech. The labeling is based on the word’s conventional grammatical function in the dialect and references the dialect word class library in the big data. Special marking is used for core words related to instructions to highlight their functions.
6. The method for controlling a graphene AI smart coffee table with voice interaction according to claim 1, characterized in that: In step S2, The steps for obtaining the two-dimensional feature matrix are as follows: the continuous audio in the mixed samples is divided into short frames of 20-30ms with an inter-frame overlap of 10ms; A Hanning window is applied to each frame of the signal, and a Fast Fourier Transform is performed on each windowed frame to convert the time-domain signal into the frequency domain, obtaining the amplitude of each frequency component and forming a spectrum. A Mel filter bank is used to simulate the nonlinear perception of frequency by the human ear, mapping the spectrum to the Mel frequency scale and preserving key frequency band information of the speech. The logarithm of the Mel spectrum is taken to compress the dynamic range and enhance the recognition of low-energy features, resulting in a two-dimensional feature matrix. In the matrix, speech presents regular spectral stripes, while noise appears to be randomly distributed.
7. The voice-interactive graphene AI smart coffee table control method according to claim 1, characterized in that: In step S3, the collected voice data is uniformly formatted as WAV and MP3, and the sampling rate is set to 16kHz. Data with labeling errors and poor audio quality are removed, and speech features are extracted. The speech features are obtained after processing by framing, windowing, Fourier transform, Mel filtering and discrete cosine transform. The processed data is divided proportionally for model training, validation, and evaluation. The model is then optimized, and the optimization steps are the same as those in step S2, where the built-in model learns to distinguish speech from noise features. The cloud server is an 8-core, 16GB memory cloud server. The optimized model is stored in ONNX format. The model initialization time is less than 3 seconds. Preprocessing includes format verification and conversion. After receiving the audio, the format is first verified. If it is in MP3 format, it is converted to WAV format in real time using FFmpeg. CRC check is enabled during the conversion process to ensure data integrity.
Citation Information
Patent Citations
Intelligent tea table, voice instruction recognition method and storage medium
CN112056763A
Intelligent electric appliance voice recognition control method and system
CN120472900A
