A voice wake-up false trigger multi-stage verification method, system, medium and device

Abstract

The application discloses a voice wake-up false triggering multi-stage verification method, system, medium and equipment, and belongs to the technical field of artificial intelligence and voice signal processing. The method obtains an end-side audio stream, sequentially performs three-stage verification of acoustic feature filtering, wake-up model triggering judgment and automatic speech recognition semantic matching, generates a wake-up candidate and locks a wake-up judgment window, obtains a transcription text and a transcription confidence, obtains a real wake-up or a false triggering judgment according to a matching result of the transcription text and a preset wake-up word set and a homonym set, performs quality gating on a false triggering segment, labels a segment meeting a condition as a hard negative sample, and writes the hard negative sample into a sample library; and after the newly added hard negative sample is combined with an original training set and the end-side model is fine-tuned and then redeployed, a false triggering detection and model self-optimization closed loop is formed. The application can automatically deposit false triggering hard negative samples without manual annotation and without separating the audio from the end, reduces annotation cost, improves privacy compliance, and continuously reduces the false triggering rate.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and speech signal processing technology, and in particular to a method, system, medium and device for multi-level verification of voice wake-up false triggering. Background Technology

[0002] In products such as smart speakers, conferencing equipment, and in-vehicle voice assistants, on-device wake-up (OWW) models need to continuously listen for specific wake words locally on the device. Due to limitations in computing power and model size, lightweight OOWW models often set relatively loose trigger thresholds to maintain recall in noisy environments. However, this can also lead to false positives in scenarios such as television noise, multi-person conversations, and homophones. In other words, the device may mistakenly enter an interactive state even when the user has not spoken the wake word, causing the screen to light up unexpectedly, interrupting the user, waking up the cloud link, and consuming power and data.

[0003] To address the issue of false voice wake-up triggers, common industry implementations generally fall into three categories: First, manual capture and annotation. Wake-up event audio is exported from device logs, and annotators listen to it, judge its accuracy, manually slice and label it, and then send it back for training. This is costly, time-consuming, and the original audio needs to be exposed to the annotators, making it difficult to meet privacy compliance requirements. Second, pure thresholds or rule-based blacklists. These rely solely on OWW output scores and fixed thresholds, or add static lists of easily confused words for rejection. A single score threshold cannot distinguish between genuine wake-up and highly similar false triggers, and homophone blacklists require manual maintenance and are difficult to cover open domain combinations. Third, cloud-based secondary confirmation. After triggering on the device, the entire audio segment is uploaded to the cloud for secondary judgment by a large-scale Automatic Speech Recognition (ASR) or voiceprint model. This method has high accuracy but relies on off-device uploads, leading to privacy and latency issues. Furthermore, cloud-based rejections often only affect the current interaction and are not systematically accumulated as negative samples in the training set.

[0004] The above solutions all have limitations: the manual solution lacks an automated discrimination and labeling chain; the pure threshold and rule solution cannot close the "whether it is a wake word" determination when there is a lack of high-level semantic verification because the lightweight model on the edge does not have semantic understanding capabilities; the cloud solution lacks a "discrimination-accumulation-training-deployment" closed loop because it is positioned as a single-point discrimination. Summary of the Invention

[0005] To overcome the shortcomings of existing technologies, such as the lack of automated discrimination and labeling links due to manual solutions, and the inability of pure threshold and rule-based solutions to close the "whether it is a wake word" determination in the absence of high-level semantic verification due to the lack of semantic understanding capabilities in lightweight edge models.

[0006] In a first aspect, the present invention provides a multi-level verification method for false voice wake-up, comprising the following steps: Acquire the audio stream collected from the terminal side, and sequentially perform acoustic feature pre-filtering, terminal side voice wake-up model trigger determination, and automatic speech recognition semantic matching on the audio stream to generate wake-up candidates and lock the wake-up determination window, and obtain the transcribed text and transcription confidence of the wake-up determination window. Based on the matching results between the transcribed text and the preset set of wake words and homophones, a judgment result is obtained. The judgment result includes real wake-up and false triggering. Quality gating is performed on the audio segments judged as false triggering, and the false triggering audio segments that meet the quality gating conditions are marked as hard negative samples and written into the hard negative sample library. The newly added hard negative samples are merged with the original training set to fine-tune the terminal voice wake-up model, and the fine-tuned model is redeployed to the terminal to form a closed loop of false trigger detection and model self-optimization.

[0007] Optionally, the specific steps for pre-filtering acoustic features are as follows: Short-time acoustic features are calculated frame by frame for the audio stream, and audio frames that meet the preset acoustic threshold are sent to the end-side voice wake-up model for triggering determination; wherein the short-time acoustic features include at least one of short-time energy, zero-crossing rate and signal-to-noise ratio.

[0008] Optionally, the specific steps for triggering the terminal-side voice wake-up model are as follows: The activation score is calculated by segmenting the audio that has been pre-filtered by the acoustic features. When the activation score is not lower than the preset trigger threshold in a number of consecutive segments and the time interval between the last trigger exceeds the preset cooldown time, the wake-up candidate is generated, and the wake-up determination window is locked based on the trigger time and the preset retention time before and after.

[0009] Optionally, the specific steps for matching the transcribed text based on a preset set of wake words and a set of homophones to obtain the determination result are as follows: The transcribed text and transcription confidence are obtained from the wake-up determination window using an automatic speech recognition engine; wherein the homophone set is automatically generated by the pinyin or phoneme sequence of the wake word according to the edit distance and iteratively expanded by combining historical false trigger texts; the case where the wake word set is hit is determined to be a real wake-up; the case where the homophone set is hit in other cases is determined to be a false trigger; the case where the transcription confidence meets the threshold is also determined to be a false trigger.

[0010] Optionally, the specific steps for performing quality gating on audio segments determined to be falsely triggered and labeling falsely triggered audio segments that meet the conditions as hard negative samples are as follows: Quality gating is performed on wake-up determination windows determined to be falsely triggered. The quality gating is a combination of the following conditions: the duration of the audio segment is within a preset duration range; the signal-to-noise ratio of the segment is not lower than a preset signal-to-noise ratio threshold; the activation score at the trigger time is within a preset score range; and the transcription confidence is not lower than a preset confidence threshold. Falsely triggered audio segments that meet all the conditions of the quality gating are labeled as hard negative samples and written into the hard negative sample library. Falsely triggered audio segments determined by the homophonic set matching are preferentially labeled as hard negative samples.

[0011] Optionally, the interactive disclosure method of the wake-up candidate includes one of the following: after generating the wake-up candidate, the device immediately enters the interactive state and performs the automatic speech recognition semantic matching in the background, and exits the interactive state when it is determined to be a false trigger; or after generating the wake-up candidate, the device remains in a silent standby state, and enters the interactive state only when the automatic speech recognition semantic matching determines that it is a real wake-up.

[0012] Optionally, the specific steps for fine-tuning the end-side voice wake-up model after merging the newly added hard negative samples with the original training set are as follows: The newly added hard negative samples are merged with the original training set according to a preset merging ratio. During the training process, a higher loss weight is applied to the negative samples than to the positive samples. The terminal voice wake-up model is then fine-tuned to obtain an updated model, which is then redeployed to the terminal to replace the original model.

[0013] Secondly, the present invention provides a multi-level verification system for voice wake-up false triggering, comprising: A multi-level verification module is used to acquire the audio stream collected from the terminal side, and sequentially perform acoustic feature pre-filtering, terminal voice wake-up model trigger judgment and automatic speech recognition semantic matching on the audio stream to generate wake-up candidates and lock the wake-up judgment window, and obtain the transcribed text and transcription confidence of the wake-up judgment window. The false trigger detection and sample accumulation module is used to obtain a detection result based on the matching results of the transcribed text with a preset set of wake words and a set of homophones. The detection result includes... Real wake-up and false triggering are identified, and quality gating is performed on audio segments determined to be false triggering. False triggering audio segments that meet the quality gating conditions are marked as hard negative samples and written into the hard negative sample library. The model self-optimization module is used to fine-tune the terminal voice wake-up model by merging the newly added hard negative samples with the original training set, and then redeploy the fine-tuned model to the terminal, forming a closed loop of false trigger detection and model self-optimization.

[0014] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a multi-level verification method for false voice wake-up.

[0015] Fourthly, the present invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; The processor, when executing a program stored in memory, implements the steps of a multi-level verification method for false triggering of voice wake-up.

[0016] The beneficial effects of this invention are: through three-level cascaded verification and semantic matching, the ability to detect false triggers in open domain noise and homophonic scenarios is significantly improved; false trigger slices are automatically precipitated as hard negative samples through quality gating, eliminating the need for manual listening and annotation and reducing iteration costs; the closed loop of detection-gating precipitation-merging training-redeployment enables the model to continuously self-optimize as it is distributed online. Attached Figure Description

[0017] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0018] Figure 1 These are flowcharts from some of the embodiments; Figure 2 These are system block diagrams from some embodiments; Figure 3 This is another flowchart in some embodiments. Detailed Implementation

[0019] The following will clearly and completely describe the concept, specific structure, and technical effects of the present invention in conjunction with embodiments and accompanying drawings, so as to fully understand the purpose, features, and effects of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, not all of them. Other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are all within the scope of protection of the present invention. Furthermore, all connections / linkages involved in the patent do not simply refer to direct contact between components, but rather to the ability to form a better connection structure by adding or reducing connecting accessories according to specific implementation conditions. The various technical features in this invention can be combined interactively without contradicting each other.

[0020] This invention provides a multi-level verification method for false voice wake-up, comprising the following steps: S1. Acquire the audio stream collected from the terminal side, and sequentially perform acoustic feature pre-filtering, terminal side voice wake-up model trigger determination, and automatic speech recognition semantic matching on the audio stream to generate wake-up candidates and lock the wake-up determination window, and obtain the transcribed text and transcription confidence of the wake-up determination window. S2. Based on the matching results between the transcribed text and the preset wake word set and homophone set, a determination result is obtained. The determination result includes real wake-up and false triggering. Quality gating is performed on the audio segments determined to be false triggering. The false triggering audio segments that meet the quality gating conditions are marked as hard negative samples and written into the hard negative sample library. S3. After merging the newly added hard negative samples with the original training set, fine-tune the terminal voice wake-up model, and redeploy the fine-tuned model to the terminal to form a closed loop of false trigger detection and model self-optimization.

[0021] Specifically, the audio stream acquired from the device side is used to provide the original speech signal source for subsequent multi-level verification (device side refers to the OOWW model and verification logic running locally on the device, typically in smart speakers, conferencing equipment, and in-vehicle voice assistants). The audio stream undergoes three levels of verification sequentially: the first level is acoustic feature pre-filtering, which filters out obvious non-speech or low-quality signals such as silence and white noise before the audio enters the wake-up model to reduce the false trigger floor; the second level is the device-side speech wake-up model trigger determination, where a lightweight neural network (On-device Wake Word, OOWW) activates and scores the audio output block by block and generates wake-up candidates based on thresholds; the third level is Automatic Speech Recognition (ASR), which transcribes the locked window to obtain text and confidence levels, expanding false trigger discrimination from a single score to a closed-loop process of "acoustics—score—semantics". After three levels of concatenation, wake-up candidates are generated and a wake-up determination window (i.e., an audio segment containing the context before and after the trigger, denoted as W_ctx) is locked. The transcribed text and transcription confidence of this window are obtained, providing semantic basis for subsequent determination. Step two: the determination result is obtained based on the matching result of the transcribed text with the preset wake-up word set and the homophone set derived from the wake-up word. The determination result includes real wake-up and false trigger, thus advancing the determination of "whether it is a wake-up word" from the score dimension to the semantic content dimension. Quality gating is performed on the audio segments determined to be false triggers. The false trigger audio segments that meet the quality gating conditions are labeled as hard negative samples (i.e., difficult negative samples that are near the OOW discrimination boundary and are easily misclassified as positive) and written into the hard negative sample library, so that the false trigger events are automatically transformed into difficult negative samples that can be used for training, without the need for manual listening and labeling. Step 3: Add the new hard negative samples (i.e. audio segments that exceed the OOWW threshold but whose semantics are not the target wake word) to the original training set and fine-tune the terminal voice wake-up model. Then, redeploy the fine-tuned model to the terminal to form a closed loop of false trigger detection and model self-optimization, so that the false trigger rate can continue to decrease with the online real distribution.

[0022] By using three-level cascaded verification and semantic matching, the ability to detect false triggers in open domain noise and homophonic scenarios is significantly improved; false trigger slices are automatically precipitated as hard negative samples through quality gating, eliminating the need for manual listening and annotation and reducing iteration costs; the closed loop of detection-gating precipitation-merging training-redeployment enables the model to continuously self-optimize as it is distributed online.

[0023] In addition, acoustic feature pre-filtering can be omitted or simplified (e.g., only retaining voice activity detection), without affecting the two-level core implementation of "end-side voice wake-up model trigger determination + automatic speech recognition semantic matching"; in addition to the general engine, automatic speech recognition can adopt end-side lightweight speech recognition or dedicated keyword / phoneme discrimination small models; hard negative sample backflow supports both end-side personalized fine-tuning and unified retraining and distribution after multi-end convergence; parameters such as the before and after retention time of the wake-up determination window, trigger threshold, and cooldown time can be adjusted according to the product's trade-off between latency and false triggering.

[0024] In some embodiments, the specific steps of acoustic feature pre-filtering are as follows: Short-time acoustic features are calculated frame by frame for the audio stream, and audio frames that meet the preset acoustic threshold are sent to the end-side voice wake-up model for triggering determination; wherein the short-time acoustic features include at least one of short-time energy, zero-crossing rate and signal-to-noise ratio.

[0025] Specifically, short-term acoustic features are calculated frame-by-frame for the audio stream to filter out obvious non-speech or low-quality signals before the audio enters the on-device voice wake-up model, reducing subsequent invalid calculations and false triggering. In one implementation, the frame length is 20 to 30 milliseconds, and the sampling sequence x(i) (i=0,…,L) of the nth frame is used. 1) Calculate the short-time energy E(n) = Σx(i)²; if E(n) is lower than the silence threshold E_min, the frame will not enter the second level. Calculate the zero-crossing rate ZCR(n) to identify and reject frames that are close to white noise, have an abnormally high zero-crossing rate, and low energy. Take a moving average of the energy of recent silent frames to obtain the noise floor N_bg, and estimate the signal-to-noise ratio SNR≈10·log10(E(n) / N_bg); when the SNR is lower than the preset threshold SNR_min, the frame will not enter the second level. Optionally, the Voice Activity Detection (VAD) model is used to output the speech probability p_vad, and the frame will only enter the second level when p_vad≥τ_vad. The short-time acoustic features include at least one of short-time energy, zero-crossing rate, and signal-to-noise ratio, which can be used alone or in combination to adapt to different noise environments; audio frames that do not pass the threshold may not be sent to the wake-up model, or although they are sent, their scores may not participate in the triggering logic. Frames that do not pass the threshold in the first stage are not sent to OOWW, thus effectively reducing the amount of subsequent computation on the computing-limited end.

[0026] Pre-filtering effectively reduces the computational load of subsequent wake-up models on the computing-limited end side and lowers the false triggering floor caused by silence, white noise, etc.; the signal-to-noise ratio formula SNR≈10·log10(E(n) / N_bg) is normalized with sliding noise floor, so that the threshold can adapt to the environment.

[0027] In addition, short-term acoustic features can also be obtained from speech probability, spectral centroid, spectral entropy, etc., output by the speech activity detection model; the acoustic threshold can be changed from a fixed value to a floating threshold that dynamically adapts to the ambient noise level; audio frames that fail to pass the threshold may not be sent to the wake-up model, or although they are sent, their scores may not participate in the triggering logic.

[0028] In some embodiments, the specific steps for determining the trigger of the terminal-side voice wake-up model are as follows: The activation score is calculated by segmenting the audio that has been pre-filtered by the acoustic features. When the activation score is not lower than the preset trigger threshold in a number of consecutive segments and the time interval between the last trigger exceeds the preset cooldown time, the wake-up candidate is generated, and the wake-up determination window is locked based on the trigger time and the preset retention time before and after.

[0029] Specifically, activation scores are calculated for audio blocks after acoustic feature pre-filtering, and these scores are used to output block-by-block confidence scores by the end-side speech wake-up model. One example of parameters for this implementation (not limited): sampling rate 16 kHz; 1280 sampling points per block (approximately 80 milliseconds); preset trigger threshold θ = 0.5; number of consecutive blocks K = 1; cooling time T_cool = 2 seconds; pre-trigger retention time N = 2 seconds; post-trigger retention time M = 1 second. When the activation score s_t in several consecutive blocks is not lower than the preset trigger threshold θ, and the time interval since the last trigger exceeds the preset cooling time T_cool, a wake-up candidate is generated to avoid misjudging a single noise spike as a wake-up and to suppress short-term repeated triggering. The wake-up determination window is locked based on the trigger time t and the preset pre- and post-trigger retention times, denoted as W_ctx = [t...]. [N, t+M] provides audio segments containing the context before and after the trigger for subsequent automatic speech recognition, ensuring the integrity of the transcribed text and improving the reliability of third-level semantic matching.

[0030] Block scoring combined with cooling time effectively suppresses isolated noise and continuous repetitive false triggers; the locked context window W_ctx=[t [N, t+M] ensures the integrity of the transcribed text and improves the reliability of third-level semantic matching; the above parameter examples provide practical engineering values.

[0031] In addition, the trigger threshold θ, the number of consecutive blocks K, the cooling time T_cool, and the retention times N and M before and after are all configurable parameters that can be adjusted according to the device's computing power and latency requirements; the comparison of activation scores can be expanded from "not lower than the threshold" to a dual-threshold judgment with hysteresis; in addition to fixed-length blocks, overlapping blocks or dynamic blocks based on voice activity detection can also be used for block segmentation.

[0032] In some embodiments, the specific steps for matching the transcribed text according to a preset set of wake words and a set of homophones to obtain the determination result are as follows: The transcribed text and transcription confidence are obtained from the wake-up determination window using an automatic speech recognition engine; wherein the homophone set is automatically generated by the pinyin or phoneme sequence of the wake word according to the edit distance and iteratively expanded by combining historical false trigger texts; the case where the wake word set is hit is determined to be a real wake-up; the case where the homophone set is hit in other cases is determined to be a false trigger; the case where the transcription confidence meets the threshold is also determined to be a false trigger.

[0033] Specifically, the automatic speech recognition engine is used to obtain the transcribed text T and transcription confidence c from the wake-up decision window W_ctx, converting the audio content into semantically matchable text content. The wake-up word set W is defined as the preset target wake-up words and their equivalent written forms; the homophone set H is defined as candidates automatically generated from the pinyin or phoneme sequences of the wake-up words according to edit distance, and iteratively expanded by combining historical false-trigger text, so that matching can cover open-domain homophones without manually maintaining a static blacklist (edit distance is an algorithm that measures the similarity between two character sequences; in one implementation, edit distance d=1 or 2 is allowed). Based on the matching results, the judgment result is obtained according to the following four categories: First, when the transcribed text matches the wake word set W, the judgment result is determined to be a true wake-up, and the segment is not recorded as a negative sample; Second, when the transcribed text only matches the homophone set H and does not match W, the judgment result is determined to be a false trigger, and the segment is preferentially labeled as a hard negative sample because it is near the discrimination boundary and is therefore a harder negative sample; Third, when the transcribed text neither matches W nor H, the judgment result is determined to be a false trigger, and the segment is labeled as a hard negative sample; Fourth, when the transcription confidence c is lower than the confidence threshold c_min, the judgment result is determined to be uncertain, neither judged as a true wake-up nor recorded, in order to avoid automatic speech recognition errors polluting the training set. This extends the discrimination from the score dimension to the semantic content dimension and distinguishes between three branches: true wake-up, false trigger, and low-confidence uncertainty.

[0034] Semantic matching is performed using the wake word set W and the automatically expanded homophone set H to distinguish between real wake-ups, false triggers, and low-confidence uncertain branches. Compared with pure thresholds or static blacklists, it can cover unseen homophones and avoid low-confidence pollution. The homophone set H is automatically derived by the edit distance algorithm and iteratively expanded with historical false triggers, eliminating the need for manual maintenance of the blacklist.

[0035] In addition, the automatic speech recognition engine can use edge-based lightweight recognition, cloud-based recognition, or a dedicated keyword / phoneme discrimination small model; the edit distance threshold d of the homophone set and the expanded sources (historical mistriggered text, user correction feedback) can all be configured; voiceprint or speaker confirmation can also be introduced as an additional verification dimension for the judgment.

[0036] In some embodiments, the specific steps for performing quality gating on audio segments determined to be falsely triggered and labeling falsely triggered audio segments that meet the conditions as hard negative samples are as follows: Quality gating is performed on wake-up determination windows determined to be falsely triggered. The quality gating is a combination of the following conditions: the duration of the audio segment is within a preset duration range; the signal-to-noise ratio of the segment is not lower than a preset signal-to-noise ratio threshold; the activation score at the trigger time is within a preset score range; and the transcription confidence is not lower than a preset confidence threshold. Falsely triggered audio segments that meet all the conditions of the quality gating are labeled as hard negative samples and written into the hard negative sample library. Falsely triggered audio segments determined by the homophonic set matching are preferentially labeled as hard negative samples.

[0037] Specifically, a quality gating process is applied to wake-up windows identified as false triggers to filter out segments of sufficient quality suitable for training from numerous false triggers. The quality gating is a combination of four conditions that together ensure reliable labeling and usable acoustic quality of the input samples: First, the audio segment duration is within a preset duration range, approximately 2 to 4 seconds in one implementation; second, the segment signal-to-noise ratio (SNR) is not lower than a preset SNR threshold (SNR_gate) to eliminate segments with excessively poor acoustic quality; third, the activation score (s_t) at the trigger time is within a preset score range, i.e., s_t∈[θ,1], where θ is a preset trigger threshold with weight 3; fourth, the transcription confidence is not lower than a preset confidence threshold (c_min) to avoid low-confidence automatic speech recognition errors. False trigger audio segments that meet all four conditions are labeled as hard negative samples and written into the hard negative sample library; false trigger audio segments determined by homophonic set matching are preferentially labeled as hard negative samples because they are near the discrimination boundary and have greater training value. Judgment and gating can be completed locally on the device side; during reflow training, one implementation method only uploads acoustic features or desensitized slices to protect privacy.

[0038] Quality gating ensures that the hard negative samples entering the database are both "semantically confirmed as non-wake words" and "acoustic quality meets the standards", improving the efficiency of fine-tuning; prioritizing the precipitation of homophonic false triggers further enhances the targeting of difficult negative samples; all four gating conditions are expressed by quantifiable parameters (duration interval, SNR_gate, score interval [θ,1], c_min), which facilitates engineering calibration.

[0039] In addition, the conditions for quality gating can be added, deleted, or relaxed. For example, only the duration and signal-to-noise ratio can be retained, or the continuity of speech activity can be added. The threshold and interval can be calibrated according to the acoustic environment of the device. The priority labeling rules can also be extended to prioritize according to the "score close to the threshold" or "homophonic editing distance".

[0040] In some embodiments, the interactive disclosure method of the wake-up candidate includes one of the following: after generating the wake-up candidate, the device immediately enters the interactive state and performs the automatic speech recognition semantic matching in the background, and exits the interactive state when it is determined to be a false trigger; or after generating the wake-up candidate, the device remains in a silent standby state and enters the interactive state only when the automatic speech recognition semantic matching determines that it is a real wake-up.

[0041] The interaction disclosure method for wake-up candidates includes one of the following, used to select between "low latency" and "low false triggering" based on product positioning. Implementation A (Low Latency): After generating a wake-up candidate, the device immediately enters a standby state and performs automatic speech recognition semantic matching in the background. If a false trigger is detected, the device exits the standby state and the segment is sent to the sedimentation gating. This is suitable for scenarios sensitive to response speed. Implementation B (Low False Trigger): After generating a wake-up candidate, the device remains in a silent standby state, only entering the interaction state when the automatic speech recognition semantic matching determines it is a genuine wake-up. This is suitable for scenarios sensitive to the cost of false triggering. In both implementations, the false trigger determination only takes effect after semantic matching is completed, avoiding the exposure of false wake-ups.

[0042] The two disclosure methods enable the same multi-level verification framework to adapt to the trade-off between latency and false triggering in different products, and the false triggering judgment only takes effect after semantic matching is completed, avoiding the exposure of false wake-ups to the outside world; Implementation method A trades response speed for minimal latency, while implementation method B trades a brief period of silence for a lower false triggering rate.

[0043] In addition, the two disclosure methods can be dynamically switched or coexisted according to the scenario (for example, use method A in a quiet environment and method B in a noisy environment); after entering the interactive state, exiting can be accompanied by a prompt sound or visual feedback; in the silent standby state, candidates can also be recorded locally for subsequent diagnosis.

[0044] In some embodiments, the specific steps for fine-tuning the edge voice wake-up model after merging the newly added hard negative samples with the original training set are as follows: The newly added hard negative samples are merged with the original training set according to a preset merging ratio. During the training process, a higher loss weight is applied to the negative samples than to the positive samples. The terminal voice wake-up model is then fine-tuned to obtain an updated model, which is then redeployed to the terminal to replace the original model.

[0045] Specifically, newly added hard negative samples are merged with the original training set according to a preset merging ratio to control the proportion of fresh false triggering samples in the training set and prevent the new distribution from overwhelming the original data (in one implementation, the merging ratio ρ∈[0.05,0.3]). During training, a higher loss weight is applied to negative samples than to positive samples, making the model pay more attention to easily confused difficult negative samples (in one implementation, the negative class loss weight λ_neg>1). After fine-tuning the on-device voice wake-up model, an updated model is obtained and redeployed to the on-device to replace the original model, completing the model update step in the closed loop. This allows the new model to continue generating discrimination results online, and the closed loop repeats, with the false triggering rate converging with the true distribution.

[0046] The merging ratio ρ and the negative class weighting λ_neg together improve the suppression effect on false triggers of the same class, and have higher training efficiency than random noise negative samples. Furthermore, redeployment allows the online model to continuously follow the real distribution. ρ∈[0.05,0.3] and λ_neg>1 provide feasible training hyperparameters.

[0047] In addition, the merging ratio ρ and the negative class loss weight λ_neg can be adjusted online according to the severity of the error trigger; in addition to full fine-tuning, the fine-tuning methods can include efficient parameter fine-tuning, distillation, or incremental training; redeployment can be updated locally on the device or uniformly distributed by the server.

[0048] This invention provides a multi-level verification system for voice wake-up false triggering, comprising: A multi-level verification module is used to acquire the audio stream collected from the terminal side, and sequentially perform acoustic feature pre-filtering, terminal voice wake-up model trigger judgment and automatic speech recognition semantic matching on the audio stream to generate wake-up candidates and lock the wake-up judgment window, and obtain the transcribed text and transcription confidence of the wake-up judgment window. The false trigger detection and sample accumulation module is used to obtain a detection result based on the matching results of the transcribed text with a preset set of wake words and a set of homophones. The detection result includes... Real wake-up and false triggering are identified, and quality gating is performed on audio segments determined to be false triggering. False triggering audio segments that meet the quality gating conditions are marked as hard negative samples and written into the hard negative sample library. The model self-optimization module is used to fine-tune the terminal voice wake-up model by merging the newly added hard negative samples with the original training set, and then redeploy the fine-tuned model to the terminal, forming a closed loop of false trigger detection and model self-optimization.

[0049] The present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method described above.

[0050] This invention provides an electronic device, including a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps described above.

[0051] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the embodiments described. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of this application.

Claims

1. A multi-level verification method for false triggering of voice wake-up, characterized in that, Includes the following steps: Acquire the audio stream collected from the terminal side, and sequentially perform acoustic feature pre-filtering, terminal side voice wake-up model trigger determination, and automatic speech recognition semantic matching on the audio stream to generate wake-up candidates and lock the wake-up determination window, and obtain the transcribed text and transcription confidence of the wake-up determination window. Based on the matching results between the transcribed text and the preset set of wake words and homophones, a judgment result is obtained. The judgment result includes real wake-up and false trigger. Quality gating is performed on the audio segments judged as false triggers. The false trigger audio segments that meet the quality gating conditions are marked as hard negative samples and written into the hard negative sample library. The newly added hard negative samples are merged with the original training set to fine-tune the terminal voice wake-up model, and the fine-tuned model is redeployed to the terminal to form a closed loop of false trigger detection and model self-optimization.

2. The multi-level verification method for false voice wake-up as described in claim 1, characterized in that, The specific steps of the acoustic feature pre-filtering are as follows: Short-time acoustic features are calculated frame by frame for the audio stream, and audio frames that meet the preset acoustic threshold are sent to the end-side voice wake-up model for triggering determination; wherein the short-time acoustic features include at least one of short-time energy, zero-crossing rate and signal-to-noise ratio.

3. The multi-level verification method for false voice wake-up as described in claim 1, characterized in that, The specific steps for triggering the terminal-side voice wake-up model are as follows: The activation score is calculated by segmenting the audio that has been pre-filtered by the acoustic features. When the activation score is not lower than the preset trigger threshold in a number of consecutive segments and the time interval between the last trigger exceeds the preset cooldown time, the wake-up candidate is generated, and the wake-up determination window is locked based on the trigger time and the preset before and after retention time.

4. The multi-level verification method for false voice wake-up as described in claim 1, characterized in that, The specific steps for matching the transcribed text based on a preset set of wake words and a set of homophones to obtain the determination result are as follows: The transcribed text and the transcription confidence are obtained by using an automatic speech recognition engine on the wake-up determination window; wherein the homophone set is automatically generated by the pinyin or phoneme sequence of the wake-up word according to the edit distance and iteratively expanded by combining historical false trigger text; the determination result is determined to be a real wake-up when the wake-up word set is hit; In the remaining cases, if the homophone set is matched, the determination result is determined to be a false trigger; in the remaining cases, if the transcription confidence meets the threshold, the determination result is also determined to be a false trigger.

5. The multi-level verification method for false voice wake-up as described in claim 1, characterized in that, The specific steps for performing quality gating on audio segments determined to be falsely triggered and labeling falsely triggered audio segments that meet the conditions as hard negative samples are as follows: Quality gating is performed on wake-up determination windows determined to be falsely triggered. The quality gating is a combination of the following conditions: the duration of the audio segment is within a preset duration range; the signal-to-noise ratio of the segment is not lower than a preset signal-to-noise ratio threshold; the activation score at the trigger time is within a preset score range; and the transcription confidence is not lower than a preset confidence threshold. Falsely triggered audio segments that meet all the conditions of the quality gating are labeled as hard negative samples and written into the hard negative sample library. Falsely triggered audio segments determined by the homophonic set matching are preferentially labeled as hard negative samples.

6. The multi-level verification method for false voice wake-up as described in claim 1, characterized in that, The interactive disclosure method of the wake-up candidate includes one of the following: after the wake-up candidate is generated, the device immediately enters the interactive state and performs the automatic speech recognition semantic matching in the background, and exits the interactive state when it is determined to be a false trigger; or after the wake-up candidate is generated, the device remains in a silent standby state and enters the interactive state only when the automatic speech recognition semantic matching determines that it is a real wake-up.

7. The multi-level verification method for false voice wake-up as described in claim 1, characterized in that, The specific steps for fine-tuning the edge voice wake-up model by merging the newly added hard negative samples with the original training set are as follows: The newly added hard negative samples are merged with the original training set according to a preset merging ratio. During the training process, a higher loss weight is applied to the negative samples than to the positive samples. The terminal voice wake-up model is then fine-tuned to obtain an updated model, which is then redeployed to the terminal to replace the original model.

8. A multi-level verification system for voice wake-up false triggering, characterized in that, include: A multi-level verification module is used to acquire the audio stream collected from the terminal side, and sequentially perform acoustic feature pre-filtering, terminal voice wake-up model trigger judgment and automatic speech recognition semantic matching on the audio stream to generate wake-up candidates and lock the wake-up judgment window, and obtain the transcribed text and transcription confidence of the wake-up judgment window. The false trigger detection and sample accumulation module is used to obtain a detection result based on the matching results of the transcribed text with a preset set of wake words and a set of homophones. The detection result includes... Real wake-up and false triggering are identified, and quality gating is performed on audio segments determined to be false triggering. False triggering audio segments that meet the quality gating conditions are marked as hard negative samples and written into the hard negative sample library. The model self-optimization module is used to fine-tune the terminal voice wake-up model by merging the newly added hard negative samples with the original training set, and then redeploy the fine-tuned model to the terminal, forming a closed loop of false trigger detection and model self-optimization.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a program stored in memory, implements the steps of the method according to any one of claims 1-7.

Images