A method for Internet voice wake-up in complex acoustic environments
By employing a multi-interference joint constraint parallel processing and an edge-cloud adaptive hierarchical wake-up mechanism, the problems of missed wake-up, false wake-up, and high latency in Internet voice wake-up under complex acoustic environments are solved, achieving Internet voice wake-up with low computing power, high precision, and low power consumption.
Patent Information
- Application Number
- CN202611520630.0
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-09-29
- Publication Date
- 2026-11-17
AI Technical Summary
Existing Internet voice wake-up methods suffer from problems such as missed wake-up, false wake-up, high latency, and poor adaptability in complex acoustic environments, and cannot simultaneously achieve low computing power, high precision, and low power consumption.
A multi-interference joint constraint parallel processing method is adopted, combined with the signal-to-noise ratio linked dynamic VAD algorithm and two-level lightweight cascaded judgment, and speech feature extraction and verification are performed through the end-to-cloud adaptive hierarchical wake-up mechanism to achieve dynamic adaptive optimization.
With extremely low computing power, it effectively filters out complex acoustic interference, reduces the probability of missed detection and false detection, balances high precision and low power consumption, adapts to complex acoustic environments, and achieves accurate wake-up.
Smart Images

Figure CN123119901A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of intelligent voice wake-up technology, and in particular relates to an Internet voice wake-up method in complex acoustic environments. Background Technology
[0002] With the popularization of smart homes, in-vehicle terminals, and wearable smart devices, voice wake-up has become widely used as a human-computer interaction interface. Internet voice wake-up relies on a terminal-cloud collaborative architecture of lightweight local inference and high-precision cloud verification to achieve all-weather voice monitoring and long-distance voice-triggered interaction. Currently, mainstream Internet voice wake-up technologies fall into two categories: one is a purely local offline wake-up solution, which relies on a lightweight model built into the terminal to complete wake-up word detection; the other is a terminal-cloud collaborative wake-up solution, which uploads audio to the cloud for secondary recognition after local coarse screening. Both solutions can work normally in quiet indoor environments, but they both have inherent technical defects that are difficult to overcome when facing complex acoustic environments with superimposed noise, spatial reverberation, replayed audio, and interference from multiple people's voices.
[0003] Existing Internet voice wake-up methods generally suffer from poor adaptability to complex acoustic environments, including: traditional voice front-end processing uses a serial and independent processing method, with each algorithm module iteratively optimized independently, lacking a joint constraint mechanism; existing wake-up algorithms have inherent contradictions in computing power, accuracy, and latency; traditional single-stage lightweight matching methods have weak noise resistance and generalization capabilities; and deep learning high-precision algorithms have high inference overhead and cannot be adapted to local real-time operation on low-computing-power Internet terminals.
[0004] Meanwhile, traditional internet voice wake-up methods are divided into two single modes: pure local wake-up and full audio cloud wake-up, which are rigid in their logic. Summary of the Invention
[0005] The purpose of this invention is to provide an Internet voice wake-up method in complex acoustic environments. Without relying on multi-device collaboration or dedicated hardware architecture, it solves the technical problems of missed wake-up, false wake-up, high latency, and poor adaptability in complex acoustic environments, while taking into account low computing power, high precision, low power consumption, and privacy and security.
[0006] To achieve the above objectives, the present invention provides an Internet voice wake-up method in complex acoustic environments, comprising the following steps: S1. Enter the original relevant data, preset the wake word template, and perform initial configuration; S2. After the initial configuration is completed, the acoustic parameters in the environment are collected in real time and joint preprocessing is performed. S3. Based on the preprocessed acoustic parameters, perform dynamic adaptive speech endpoint detection, filter out noise in the acoustic parameters, and obtain effective speech features; S4. Perform a two-level hierarchical judgment on the effective speech features and the preset wake word template to obtain high-confidence speech features; S5. Perform multi-dimensional verification of high-confidence speech features to prevent false wake-up and obtain suspected valid wake-up commands; S6. Real-time detection of the terminal's internet network status, execution of end-to-cloud adaptive hierarchical wake-up determination, obtaining the final wake-up determination result, and completing internet voice wake-up; S7. Dynamically iterate and optimize scene parameters by combining historical parameters from Internet voice wake-up.
[0007] Preferably, the process of S1 is as follows: S11. Initialize the basic parameters of the local terminal's Internet connection; S12. Preset and save the wake-up word template; S13. Collect the unique voiceprint features of the target user to obtain the user's baseline voiceprint template. The process is as follows: Extract the voiceprint feature vectors of the target user's wake-up voice at different volumes and speeds, and perform mean fusion on the voiceprint feature vectors to generate the user's baseline voiceprint template. S14. Initialize the network and cloud.
[0008] Preferably, the process of S2 is as follows: S21. The original voice signal is continuously collected through the deployed terminal microphones at a preset sampling rate; S22. Detect four types of acoustic parameters in the original speech signal: noise energy, signal-to-noise ratio, reverberation duration, and echo intensity. S23, Total Loss To constrain the target, signal optimization is performed on the original speech signal to obtain clean speech frames, as follows: ; ; in Indicates echo residual loss, Indicates the combined noise loss. Indicates speech resonance loss, , and All are preset weighted balance coefficients; S24. Propose FBank acoustic features of clean speech frames, desensitize the original speech information, and unify all FBank acoustic features into a standard feature matrix of fixed length to obtain the preprocessed acoustic parameters.
[0009] Preferably, the process of S3 is as follows: S31, Combining the signal-to-noise ratio in S2 Calculate the signal-to-noise ratio linked dynamic threshold The process is as follows: ; in This represents the VAD basic energy threshold in the terminal's basic internet parameters. This represents the scene adjustment coefficient in the terminal's basic internet parameters; S32. Calculate the frame average energy of the FBank acoustic features for each frame, and use it as a quantitative indicator of the speech activity level of the current frame to obtain the single-frame feature energy. S33. Compare the single-frame feature energy with the signal-to-noise ratio linked dynamic threshold, as follows: If single-frame feature energy If the frame is valid, it is marked as a valid speech frame, and the FBank acoustic features of the current frame are retained; otherwise, it is marked as invalid noise and the features of the frame are discarded directly. S34. Set a continuous frame buffer window. If the proportion of valid audio frames in the window exceeds the preset judgment threshold, the entire segment is judged as a valid audio segment and buffered completely; otherwise, it is judged as continuous background noise and discarded entirely. S35. Perform smoothing filtering on all effective speech segments and concatenate them into a complete speech segment feature sequence to obtain effective speech features.
[0010] Preferably, the determination process for the signal-to-noise ratio linkage dynamic threshold in S31 is as follows: If the current If the value is low, it indicates a noisy environment, in which case the exponential term will be larger. Adaptive elevation; If the current A higher value indicates a quiet indoor environment or a low-noise far-field scene, where the exponential result will be smaller. Adaptive reduction.
[0011] Preferably, the process of S4 is as follows: S41. First-level template coarse judgment is performed by combining effective speech features with preset wake word templates. The process is as follows: S411. Normalize the effective speech features and align them with the memory timing of the preset wake word template. S412. Calculate the wake word normalized matching score. The process is as follows: ; in The shortest regular distance representing effective speech features. Indicates the maximum frame length of the preset wake word template; S413, if It was determined to be a suspected wake-up fragment, its feature sequence was preserved, and it proceeded to the second-level model for further refinement. This indicates the DTW matching threshold; otherwise, it is judged as irrelevant speech, and it is considered that no voice wake-up has been performed. The feature segment is discarded directly, and the process returns to S2 to collect the next frame of speech. S42. Perform secondary fine-tuning on suspected wake-up segments, as follows: S421. Output the suspected wake-up fragment to the local CRNN network; The S422 local CRNN network sequentially performs three layers of one-dimensional convolution to extract frequency domain noise-resistant features from suspected wake-up segments, two layers of unidirectional GRU to model speech temporal dependencies, and two layers of fully connected layers to map features. Finally, it outputs a binary classification result through Softmax activation to obtain the model wake-up confidence. And high-confidence speech features.
[0012] Preferably, the process of S5 is as follows: S51. Extract the voiceprint feature vector of the current high-confidence speech features, and perform cosine similarity calculation with the locally stored user baseline voiceprint to obtain the target user voiceprint similarity. ; S52. Compare the current speech spectrum features with the interference sample spectrum library to calculate the interference voice spectrum matching calculation. ; S53. Calculate the comprehensive wake-up credibility score of high-confidence speech features. The process is as follows: ; in , and All are preset weighting coefficients; S54, when both conditions are met , , If the high-confidence speech feature at this time is determined to be a suspected valid wake-up command, otherwise it is considered that no speech wake-up has been performed, and the process returns to S2 to repeat the loop acquisition. in This indicates the minimum threshold for matching user voiceprints. This indicates the upper limit threshold of the spectrum of interfering human voices. This represents the global comprehensive wake-up trust score threshold.
[0013] Preferably, the process of S6 is as follows: S61. Detect the terminal's internet link. If it is in any of the following situations: network disconnection, latency greater than 1000ms, or packet loss, perform local wake-up (S62); otherwise, perform cloud-based secondary high-precision verification wake-up (S63). S62. Dynamically lower the global comprehensive wake-up confidence score threshold to obtain the offline relaxed threshold. If the comprehensive wake-up confidence score is not less than the offline relaxed threshold, the wake-up is deemed to be effective and jumps to S7. Otherwise, it is deemed to be interference speech and it is considered that no speech wake-up has been performed. The feature segment is discarded directly and S2 is returned to collect the next frame of speech in a loop. S63. Perform differential lightweight compression on the desensitized FBank feature matrix, acoustic environment parameters, voiceprint matching index, and comprehensive wake-up credibility score, and upload it to the cloud; The cloud decrypts the uploaded data, restores the standardized feature matrix, inputs the features into a large-scale noise-resistant Transformer verification model in the cloud, and outputs the global wake-up confidence score in the cloud. ; S64. Retrieve the interference audio library from the cloud, perform a secondary verification to determine whether the current segment belongs to the broadcast or video playback type of recorded wake-up word interference, and generate a cloud-based interference penalty score. ; S65. Calculate the joint confidence score between the cloud and the endpoint, as follows: ; in , and All represent weighting coefficients; S66, The cloud matches scene-specific thresholds based on the environment tags reported by the device. And perform a wake-up determination: like Once the wake-up is successful, the cloud sends a wake-up confirmation command and the corresponding scenario's optimized algorithm parameter package. like If the system determines that the wake-up is false, it will only upload the current acoustic environment log, without sending any parameters. The terminal will discard the feature and return to S2.
[0014] Preferably, the process of S7 is as follows: S71. Update the local acoustic scene sample library; Verify the validity of the acoustic features of the wake-up speech, remove impulse noise and truncated fragments, and store the SNR, reverberation duration, average interference energy and environmental label corresponding to this wake-up into the local circular cache. Cluster the cache data to distinguish three basic acoustic scenarios: quiet bedroom, noisy living room, and high reverberation vehicle, and update the current device's resident scenario label; S72. Based on the current clustering scenario and cloud-deployed parameters, adjust the core thresholds and weighting coefficients of the entire process in layers; S73. When the terminal is in a normal network state, compress the local scene statistics log and upload the log using a low-priority asynchronous background upload method. After receiving the log, the cloud updates the global scene parameter library and iteratively optimizes the standard thresholds of various scenes for subsequent distribution to similar acoustic environment devices. S74. After completing parameter updates and log storage, clear the current speech feature cache and temporary scoring variables. The terminal returns to S2 to continuously collect raw speech frames and start the next round of wake-up detection loop.
[0015] Therefore, the present invention employs the above-described method for Internet voice wake-up in a complex acoustic environment, which has the following advantages: 1. An innovative parallel processing method with joint constraints for multiple interferences is introduced to replace the traditional serial step-by-step filtering. By using a multi-dimensional joint loss function to simultaneously optimize noise reduction, echo removal, reverberation removal, and speech fidelity, the method addresses industry pain points such as feature distortion, music noise, and far-field wake-up failure under complex acoustic interference.
[0016] 2. The signal-to-noise ratio linked dynamic VAD algorithm is adopted to overcome the rigidity of fixed thresholds. It can adapt to quiet, far-field, strong noise, and transient interference scenarios in real time, reducing the probability of missed detection and false detection from the source.
[0017] 3. A two-level lightweight cascaded judgment + multi-dimensional quantization scoring and verification system was constructed, which balances noise resistance accuracy with extremely low computing power overhead, effectively filters similar human voices and broadcast audio interference, and significantly reduces the false wake-up rate in complex scenarios.
[0018] 4. It pioneered an adaptive dual-mode wake-up mechanism that enables high-precision cloud verification when connected to the internet and local availability when disconnected from the internet. At the same time, it transmits only de-identified feature data, which greatly reduces traffic and power consumption and prevents the leakage of original voice privacy.
[0019] 5. The method has the ability to smoothly iterate parameters and can automatically update algorithm parameters according to the real-time acoustic environment. It can adapt to various complex dynamic scenarios without manual debugging, and its generalization ability and stability are significantly better than traditional fixed parameter algorithms.
[0020] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description
[0021] Figure 1 This is an overall flowchart of an Internet voice wake-up method for complex acoustic environments according to the present invention. Figure 2 This is a flowchart illustrating the process of obtaining high-confidence speech features in an internet voice wake-up method under complex acoustic environments according to the present invention. Figure 3 This is a diagram illustrating the process of completing Internet voice wake-up in a complex acoustic environment according to the present invention. Detailed Implementation
[0022] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Specific model specifications need to be selected and determined according to the actual specifications of the device, etc. The specific selection calculation method adopts existing technology in the art, and therefore will not be described in detail.
[0023] Example like Figures 1-3 As shown, this invention provides an internet voice wake-up method in complex acoustic environments, comprising the following steps: S1. Enter the original relevant data, preset the wake word template, and perform initial configuration; S11. Initialize the basic parameters of the local terminal Internet, including the factory-fixed basic hyperparameters: VAD basic energy threshold, scene adjustment coefficient, DTW matching threshold, weighted coefficients of each loss, iterative smoothing coefficient, transmission threshold 1000ms latency limit, single packet feature capacity limit, interference human voice judgment threshold, and voiceprint matching minimum threshold. Configure the underlying parameters for voice acquisition and load the lightweight, noise-resistant CRNN inference weight file.
[0024] S12. Preset and save the wake-up word template; S13. Collect the unique voiceprint features of the target user to obtain the user's baseline voiceprint template. The process is as follows: Extract the voiceprint feature vectors of the target user's wake-up voice at different volumes and speeds, and perform mean fusion on the voiceprint feature vectors to generate the user's baseline voiceprint template. S14. Initialize the network and cloud.
[0025] Deploy high-precision noise-resistant Transformer verification models, global multi-scene acoustic parameter libraries, and massive general-purpose speaker sample libraries in the cloud. Pre-store the optimal algorithm parameters (noise reduction coefficient, confidence threshold, VAD adaptation parameters) for various scenarios. Establish a service chain for receiving encrypted features, decrypting, secondary inference, and distributing lightweight parameters, with preset differential encryption / decryption algorithms.
[0026] S2. After the initial configuration is completed, the acoustic parameters in the environment are collected in real time and joint preprocessing is performed. S21. The original voice signal is continuously collected through the deployed terminal microphones at a preset sampling rate; S22. Detect four types of acoustic parameters in the original speech signal: noise energy, signal-to-noise ratio, reverberation duration, and echo intensity. Noise energy: The average spectral energy of the silent region, representing the strength of background noise; Signal-to-noise ratio (SNR): The ratio of effective speech frequency band energy to noise energy; Reverberation duration: The time required for room sound to decay by 60dB, characterizing spatial reflection interference; Echo intensity: The residual coupling energy of the audio played by the device itself, representing the echo interference during playback.
[0027] The four sets of parameters are synchronously stored in the frame-level environment parameter array and bound to each voice frame.
[0028] S23, Total Loss To constrain the target, signal optimization is performed on the original speech signal to obtain clean speech frames, as follows: ; ; in Indicates echo residual loss, Indicates the combined noise loss. Indicates speech resonance loss, , and All are preset weighted balance coefficients; S24. Propose FBank acoustic features of clean speech frames, desensitize the original speech information, and unify all FBank acoustic features into a standard feature matrix of fixed length to obtain the preprocessed acoustic parameters.
[0029] S3. Based on the preprocessed acoustic parameters, perform dynamic adaptive speech endpoint detection, filter out noise in the acoustic parameters, and obtain effective speech features; S31, Combining the signal-to-noise ratio in S2 Calculate the signal-to-noise ratio linked dynamic threshold The process is as follows: ; in This represents the VAD basic energy threshold in the terminal's basic internet parameters. This represents the scene adjustment coefficient in the terminal's basic internet parameters; If the current If the value is low, it indicates a noisy environment, in which case the exponential term will be larger. Adaptive elevation; If the current A higher value indicates a quiet indoor environment or a low-noise far-field scene, where the exponential result will be smaller. Adaptive reduction.
[0030] S32. Calculate the frame average energy of the FBank acoustic features for each frame, and use it as a quantitative indicator of the speech activity level of the current frame to obtain the single-frame feature energy. S33. Compare the single-frame feature energy with the signal-to-noise ratio linked dynamic threshold, as follows: If single-frame feature energy If the frame is valid, it is marked as a valid speech frame, and the FBank acoustic features of the current frame are retained; otherwise, it is marked as invalid noise and the features of the frame are discarded directly. S34. Set a continuous frame buffer window. If the proportion of valid audio frames in the window exceeds the preset judgment threshold, the entire segment is judged as a valid audio segment and buffered completely; otherwise, it is judged as continuous background noise and discarded entirely. S35. Perform smoothing filtering on all effective speech segments and concatenate them into a complete speech segment feature sequence to obtain effective speech features.
[0031] S4. Perform a two-level hierarchical judgment on the effective speech features and the preset wake word template to obtain high-confidence speech features; S41. First-level template coarse judgment is performed by combining effective speech features with preset wake word templates. The process is as follows: S411. Normalize the effective speech features and align them with the memory timing of the preset wake word template. S412. Calculate the wake word normalized matching score. The process is as follows: ; in The shortest regular distance representing effective speech features. Indicates the maximum frame length of the preset wake word template; S413, if It was determined to be a suspected wake-up fragment, its feature sequence was preserved, and it proceeded to the second-level model for further refinement. This indicates the DTW matching threshold; otherwise, it is judged as irrelevant speech, and it is considered that no voice wake-up has been performed. The feature segment is discarded directly, and the process returns to S2 to collect the next frame of speech. S42. Perform secondary fine-tuning on suspected wake-up segments, as follows: S421. Output the suspected wake-up fragment to the local CRNN network; The S422 local CRNN network sequentially performs three layers of one-dimensional convolution to extract frequency domain noise-resistant features from suspected wake-up segments, two layers of unidirectional GRU to model speech temporal dependencies, and two layers of fully connected layers to map features. Finally, it outputs a binary classification result through Softmax activation to obtain the model wake-up confidence. And high-confidence speech features.
[0032] S5. Perform multi-dimensional verification of high-confidence speech features to prevent false wake-up and obtain suspected valid wake-up commands; S51. Extract the voiceprint feature vector of the current high-confidence speech features, and perform cosine similarity calculation with the locally stored user baseline voiceprint to obtain the target user voiceprint similarity. ; S52. Compare the current speech spectrum features with the interference sample spectrum library to calculate the interference voice spectrum matching calculation. ; S53. Calculate the comprehensive wake-up credibility score of high-confidence speech features. The process is as follows: ; in , and All are preset weighting coefficients; S54, when both conditions are met , , If the high-confidence speech feature at this time is determined to be a suspected valid wake-up command, otherwise it is considered that no speech wake-up has been performed, and the process returns to S2 to repeat the loop acquisition. in This indicates the minimum threshold for matching user voiceprints. This indicates the upper limit threshold of the spectrum of interfering human voices. This represents the global comprehensive wake-up trust score threshold.
[0033] S6. Real-time detection of the terminal's internet network status, execution of end-to-cloud adaptive hierarchical wake-up determination, obtaining the final wake-up determination result, and completing internet voice wake-up; S61. Detect the terminal's internet link. If it is in any of the following situations: network disconnection, latency greater than 1000ms, or packet loss, perform local wake-up (S62); otherwise, perform cloud-based secondary high-precision verification wake-up (S63). S62. Dynamically lower the global comprehensive wake-up confidence score threshold to obtain the offline relaxed threshold. If the comprehensive wake-up confidence score is not less than the offline relaxed threshold, the wake-up is deemed to be effective and jumps to S7. Otherwise, it is deemed to be interference speech and it is considered that no speech wake-up has been performed. The feature segment is discarded directly and S2 is returned to collect the next frame of speech in a loop. S63. Perform differential lightweight compression on the desensitized FBank feature matrix, acoustic environment parameters, voiceprint matching index, and comprehensive wake-up credibility score, and upload it to the cloud; The cloud decrypts the uploaded data, restores the standardized feature matrix, inputs the features into a large-scale noise-resistant Transformer verification model in the cloud, and outputs the global wake-up confidence score in the cloud. ; S64. Retrieve the interference audio library from the cloud, perform a secondary verification to determine whether the current segment belongs to the broadcast or video playback type of recorded wake-up word interference, and generate a cloud-based interference penalty score. ; S65. Calculate the joint confidence score between the cloud and the endpoint, as follows: ; in , and All represent weighting coefficients; S66, The cloud matches scene-specific thresholds based on the environment tags reported by the device. And perform a wake-up determination: like Once the wake-up is successful, the cloud sends a wake-up confirmation command and the corresponding scenario's optimized algorithm parameter package. like If the system determines that the wake-up is false, it will only upload the current acoustic environment log, without sending any parameters. The terminal will discard the feature and return to S2.
[0034] S7. Dynamically iterate and optimize scene parameters by combining historical parameters from Internet voice wake-up.
[0035] S71. Update the local acoustic scene sample library; Verify the validity of the acoustic features of the wake-up speech, remove impulse noise and truncated fragments, and store the SNR, reverberation duration, average interference energy and environmental label corresponding to this wake-up into the local circular cache. Cluster the cache data to distinguish three basic acoustic scenarios: quiet bedroom, noisy living room, and high reverberation vehicle, and update the current device's resident scenario label; S72. Based on the current clustering scenario and cloud-deployed parameters, adjust the core thresholds and weighting coefficients of the entire process in layers; S73. When the terminal is in a normal network state, compress the local scene statistics log and upload the log using a low-priority asynchronous background upload method. After receiving the log, the cloud updates the global scene parameter library and iteratively optimizes the standard thresholds of various scenes for subsequent distribution to similar acoustic environment devices. S74. After completing parameter updates and log storage, clear the current speech feature cache and temporary scoring variables. The terminal returns to S2 to continuously collect raw speech frames and start the next round of wake-up detection loop.
[0036] This method can be applied to complex, noisy environments such as home kitchens. The scenario includes multiple superimposed noises such as steady-state noise from the range hood, dynamic noise from water flow, interference from television voices, and indoor reverberation, while the terminal is connected to the network normally.
[0037] Multiple composite interferences are suppressed synchronously by a joint loss function, and the VAD threshold is dynamically raised to filter continuous background noise. Invalid noise is quickly filtered out by DTW coarse screening, and the CRNN model accurately extracts wake-up word features. Human voices in TV broadcasts are filtered by voiceprint matching and interference spectrum identification. In the network state, only desensitized features are uploaded to complete the secondary fine verification in the cloud. High confidence thresholds for high-noise scenarios are issued in the cloud, and terminal parameters are iteratively updated to achieve accurate wake-up, error-free triggering, and no missed wake-up in noisy environments.
[0038] Therefore, the present invention employs an Internet voice wake-up method in a complex acoustic environment, which has the following advantages: 1. An innovative parallel processing method with joint constraints for multiple interferences is introduced to replace the traditional serial step-by-step filtering. By using a multi-dimensional joint loss function to simultaneously optimize noise reduction, echo removal, reverberation removal, and speech fidelity, the method addresses industry pain points such as feature distortion, music noise, and far-field wake-up failure under complex acoustic interference.
[0039] 2. The signal-to-noise ratio linked dynamic VAD algorithm is adopted to overcome the rigidity of fixed thresholds. It can adapt to quiet, far-field, strong noise, and transient interference scenarios in real time, reducing the probability of missed detection and false detection from the source.
[0040] 3. A two-level lightweight cascaded judgment + multi-dimensional quantization scoring and verification system was constructed, which balances noise resistance accuracy with extremely low computing power overhead, effectively filters similar human voices and broadcast audio interference, and significantly reduces the false wake-up rate in complex scenarios.
[0041] 4. It pioneered an adaptive dual-mode wake-up mechanism that enables high-precision cloud verification when connected to the internet and local availability when disconnected from the internet. At the same time, it transmits only de-identified feature data, which greatly reduces traffic and power consumption and prevents the leakage of original voice privacy.
[0042] 5. The method has the ability to smoothly iterate parameters and can automatically update algorithm parameters according to the real-time acoustic environment. It can adapt to various complex dynamic scenarios without manual debugging, and its generalization ability and stability are significantly better than traditional fixed parameter algorithms.
[0043] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.
Claims
1. A method for Internet voice wake-up in complex acoustic environments, characterized in that, Includes the following steps: S1. Enter the original relevant data, preset the wake word template, and perform initial configuration; S2. After the initial configuration is completed, the acoustic parameters in the environment are collected in real time and joint preprocessing is performed. S3. Based on the preprocessed acoustic parameters, perform dynamic adaptive speech endpoint detection, filter out noise in the acoustic parameters, and obtain effective speech features; S4. Perform a two-level hierarchical judgment on the effective speech features and the preset wake word template to obtain high-confidence speech features; S5. Perform multi-dimensional verification of high-confidence speech features to prevent false wake-up and obtain suspected valid wake-up commands; S6. Real-time detection of the terminal's internet network status, execution of end-to-cloud adaptive hierarchical wake-up determination, obtaining the final wake-up determination result, and completing internet voice wake-up; S7. Dynamically iterate and optimize scene parameters by combining historical parameters from Internet voice wake-up.
2. The method for Internet voice wake-up in a complex acoustic environment according to claim 1, characterized in that, The process of S1 is as follows: S11. Initialize the basic parameters of the local terminal's Internet connection; S12. Preset and save the wake-up word template; S13. Collect the unique voiceprint features of the target user to obtain the user's baseline voiceprint template. The process is as follows: Extract the voiceprint feature vectors of the target user's wake-up voice at different volumes and speeds, and perform mean fusion on the voiceprint feature vectors to generate the user's baseline voiceprint template. S14. Initialize the network and cloud.
3. The method for Internet voice wake-up in a complex acoustic environment according to claim 2, characterized in that, The process of S2 is as follows: S21. The original voice signal is continuously collected through the deployed terminal microphones at a preset sampling rate; S22. Detect four types of acoustic parameters in the original speech signal: noise energy, signal-to-noise ratio, reverberation duration, and echo intensity. S23, Total Loss To constrain the target, signal optimization is performed on the original speech signal to obtain clean speech frames, as follows: ; ; in Indicates echo residual loss, Indicates the combined noise loss. Indicates speech resonance loss, , and All are preset weighted balance coefficients; S24. Propose FBank acoustic features of clean speech frames, desensitize the original speech information, and unify all FBank acoustic features into a standard feature matrix of fixed length to obtain the preprocessed acoustic parameters.
4. The Internet voice wake-up method in a complex acoustic environment according to claim 3, characterized in that, The process of S3 is as follows: S31, Combining the signal-to-noise ratio in S2 Calculate the signal-to-noise ratio linked dynamic threshold The process is as follows: ; in This represents the VAD basic energy threshold in the terminal's basic internet parameters. This represents the scene adjustment coefficient in the terminal's basic internet parameters; S32. Calculate the frame average energy of the FBank acoustic features for each frame, and use it as a quantitative indicator of the speech activity level of the current frame to obtain the single-frame feature energy. S33. Compare the single-frame feature energy with the signal-to-noise ratio linked dynamic threshold, as follows: If single-frame feature energy If the frame is valid, it is marked as a valid speech frame, and the FBank acoustic features of the current frame are retained; otherwise, it is marked as invalid noise and the features of the frame are discarded directly. S34. Set a continuous frame buffer window. If the proportion of valid audio frames in the window exceeds the preset judgment threshold, the entire segment is judged as a valid audio segment and buffered completely; otherwise, it is judged as continuous background noise and discarded entirely. S35. Perform smoothing filtering on all effective speech segments and concatenate them into a complete speech segment feature sequence to obtain effective speech features.
5. The method for Internet voice wake-up in a complex acoustic environment according to claim 4, characterized in that, The process for determining the signal-to-noise ratio-linked dynamic threshold in S31 is as follows: If the current If the value is low, it indicates a noisy environment, in which case the exponential term will be larger. Adaptive elevation; If the current A higher value indicates a quiet indoor environment or a low-noise far-field scene, where the exponential result will be smaller. Adaptive reduction.
6. The method for Internet voice wake-up in a complex acoustic environment according to claim 5, characterized in that, The process of S4 is as follows: S41. First-level template coarse judgment is performed by combining effective speech features with preset wake word templates. The process is as follows: S411. Normalize the effective speech features and align them with the memory timing of the preset wake word template. S412. Calculate the wake word normalized matching score. The process is as follows: ; in The shortest regular distance representing effective speech features. Indicates the maximum frame length of the preset wake word template; S413, if It was determined to be a suspected wake-up fragment, its feature sequence was preserved, and it proceeded to the second-level model for further refinement. This indicates the DTW matching threshold; otherwise, it is judged as irrelevant speech, and it is considered that no voice wake-up has been performed. The feature segment is discarded directly, and the process returns to S2 to collect the next frame of speech. S42. Perform secondary fine-tuning on suspected wake-up segments, as follows: S421. Output the suspected wake-up fragment to the local CRNN network; The S422 local CRNN network sequentially performs three layers of one-dimensional convolution to extract frequency domain noise-resistant features from suspected wake-up segments, two layers of unidirectional GRU to model speech temporal dependencies, and two layers of fully connected layers to map features. Finally, it outputs a binary classification result through Softmax activation to obtain the model wake-up confidence. And high-confidence speech features.
7. The Internet voice wake-up method in a complex acoustic environment according to claim 6, characterized in that, The process of S5 is as follows: S51. Extract the voiceprint feature vector of the current high-confidence speech features, and perform cosine similarity calculation with the locally stored user baseline voiceprint to obtain the target user voiceprint similarity calculation. ; S52. Compare the current speech spectrum features with the interference sample spectrum library to calculate the interference voice spectrum matching calculation. ; S53. Calculate the comprehensive wake-up credibility score of high-confidence speech features. The process is as follows: ; in , and All are preset weighting coefficients; S54, when both conditions are met , , If the high-confidence speech feature at this time is determined to be a suspected valid wake-up command, otherwise it is considered that no speech wake-up has been performed, and the process returns to S2 to repeat the loop acquisition. in This indicates the minimum threshold for matching user voiceprints. This indicates the upper limit threshold of the spectrum of interfering human voices. This represents the global comprehensive wake-up trust score threshold.
8. The Internet voice wake-up method in a complex acoustic environment according to claim 7, characterized in that, The process of S6 is as follows: S61. Detect the terminal's internet link. If it is in any of the following situations: network disconnection, latency greater than 1000ms, or packet loss, perform local wake-up (S62); otherwise, perform cloud-based secondary high-precision verification wake-up (S63). S62. Dynamically lower the global comprehensive wake-up confidence score threshold to obtain the offline relaxed threshold. If the comprehensive wake-up confidence score is not less than the offline relaxed threshold, the wake-up is deemed to be effective and jumps to S7. Otherwise, it is deemed to be interference speech and it is considered that no speech wake-up has been performed. The feature segment is discarded directly and S2 is returned to collect the next frame of speech in a loop. S63. Perform differential lightweight compression on the desensitized FBank feature matrix, acoustic environment parameters, voiceprint matching index, and comprehensive wake-up credibility score, and upload it to the cloud; The cloud decrypts the uploaded data, restores the standardized feature matrix, inputs the features into a large-scale noise-resistant Transformer verification model in the cloud, and outputs the global wake-up confidence score in the cloud. ; S64. Retrieve the interference audio library from the cloud, perform a secondary verification to determine whether the current segment belongs to the broadcast or video playback type of recorded wake-up word interference, and generate a cloud-based interference penalty score. ; S65. Calculate the joint confidence score between the cloud and the endpoint, as follows: ; in , and All represent weighting coefficients; S66, The cloud matches scene-specific thresholds based on the environment tags reported by the device. And perform a wake-up determination: like Once the wake-up is successful, the cloud sends a wake-up confirmation command and the corresponding scenario's optimized algorithm parameter package. like If the system determines that the wake-up is false, it will only upload the current acoustic environment log, without sending any parameters. The terminal will discard the feature and return to S2.
9. The Internet voice wake-up method in a complex acoustic environment according to claim 8, characterized in that, The process for S7 is as follows: S71. Update the local acoustic scene sample library; Verify the validity of the acoustic features of the wake-up speech, remove impulse noise and truncated fragments, and store the SNR, reverberation duration, average interference energy and environmental label corresponding to this wake-up into the local circular cache. Cluster the cache data to distinguish three basic acoustic scenarios: quiet bedroom, noisy living room, and high reverberation vehicle, and update the current device's resident scenario label; S72. Based on the current clustering scenario and cloud-deployed parameters, adjust the core thresholds and weighting coefficients of the entire process in layers; S73. When the terminal is in a normal network state, compress the local scene statistics log and upload the log using a low-priority asynchronous background upload method. After receiving the log, the cloud updates the global scene parameter library and iteratively optimizes the standard thresholds of various scenes for subsequent distribution to similar acoustic environment devices. S74. After completing parameter updates and log storage, clear the current speech feature cache and temporary scoring variables. The terminal returns to S2 to continuously collect raw speech frames and start the next round of wake-up detection loop.
