Sound effect recommendation method and device, computer readable storage medium and wearable intelligent equipment
By generating a first vector and using target divergence to filter sound effect templates, the problem of discrepancies between sound effect recommendations and user expectations in existing technologies is solved, enabling personalized sound effect selection and improving recommendation efficiency and brand recognition.
Patent Information
- Application Number
- CN202512033248.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-12-30
- Publication Date
- 2026-04-10
AI Technical Summary
Existing sound effect recommendation methods are difficult to dynamically adjust based on individual preferences and the current scenario, resulting in discrepancies between recommended sound effects and user expectations, and failing to provide sound effect choices that better suit individual habits.
The first vector is generated by integrating behavioral feature data and environmental feature data. The current scene and its confidence level are accurately identified by combining multi-source perception data. The target sound effect template is selected by using the scene priority matrix, and the target sound effect is determined by the target divergence selection to ensure that the sound effect meets the brand specifications in terms of timbre, attack, and loudness.
It achieves the goal of providing sound effect options that better suit individual habits while ensuring scene adaptability and brand standards, improving recommendation efficiency and avoiding false recommendations, and ensuring that sound effects meet user expectations.
Smart Images
Figure CN121833991A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the field of wearable smart device technology, and further to a sound effect recommendation method and apparatus, a computer-readable storage medium, and a wearable smart device. Background Technology
[0002] In recent years, wearable smart devices have typically used sound to deliver notifications, reminders, interactive feedback, or entertainment content to users, making sound effects one of the core mediums of human-computer interaction. Appropriate sound effect design can not only enhance the user experience but also play a positive role in scenarios such as education, safety warnings, and emotional soothing. For example, soft prompts can be played in classroom mode to avoid distractions, while rhythmic sound effects can be used in sports scenarios to enhance motivation.
[0003] Currently, existing sound effect recommendation methods are commonly used for in-vehicle audio systems. For example, they obtain sound effect recommendations based on user-preset sound effect preferences and the popularity of sound effects. However, the sound effects recommended by this method still differ from user expectations, making it difficult to provide users with sound effect choices that better suit their individual habits. Summary of the Invention
[0004] To address the aforementioned technical issues, this application provides a sound effect recommendation method and apparatus, a computer-readable storage medium, and a wearable smart device, which can provide users with sound effect selections that better suit their individual habits.
[0005] Firstly, this application provides a sound effect recommendation method for wearable smart devices, comprising: Based on the acquired behavioral feature data and environmental feature data, a first vector is determined; based on multi-source perception data, the current scene and first confidence level are determined; based on the current scene, first confidence level, and scene priority matrix, multiple target sound effect templates are determined; based on the first vector, current scene, and multiple target sound effect templates, target divergence is determined, and based on target divergence, target sound effects are determined; behavioral feature data is used to characterize the user's operation on different sound effects, the first vector is used to characterize the user's sound effect preference, and target divergence is used to characterize the degree of distortion between the target sound effect template and the first vector.
[0006] The above sound effect recommendation method generates a first vector representing user sound effect preferences by fusing behavioral and environmental feature data. It then accurately identifies the current scene and its initial confidence level using multi-source perceptual data. Next, it filters multiple target sound effect templates that match the current scene based on a scene priority matrix. Finally, it selects the optimal target sound effect based on the target divergence (i.e., distortion level) between the first vector and each target sound effect template. This method, while ensuring scene adaptability and brand compliance, learns user preferences, effectively avoiding discrepancies between recommended sound effects and user expectations. Furthermore, the confidence level mechanism suppresses false recommendations caused by low-reliability preferences, providing users with sound effect choices that better suit their individual habits.
[0007] In one implementation, the target sound effect is determined based on the target divergence, specifically including: determining the target sound effect parameter vector based on the target divergence; and determining the target sound effect based on the target sound effect parameter vector and the sound effect library.
[0008] In one implementation, the target sound effect is determined based on the target divergence, specifically including: determining the target sound effect parameter vector based on the target divergence; generating the sound effect to be verified based on the target sound effect parameter vector and constraint parameters; and performing a weighted summation based on the color consistency, attack matching degree, and loudness deviation penalty of the sound effect to be verified. When the weighted summation result is not less than the first threshold, the sound effect to be verified is taken as the target sound effect.
[0009] The above sound effect recommendation method, after determining the target sound effect parameter vector through target divergence, can, on the one hand, directly combine with the sound effect library to quickly match compliant and user-preferred target sound effects, improving recommendation efficiency; on the other hand, it also supports generating sound effects to be verified based on the target sound effect parameter vector and constraint parameters, and introduces color consistency, attack matching degree and loudness deviation penalty for weighted summation. Only when the weighted result is not less than the first threshold is the sound effect to be verified adopted as the target sound effect. This ensures that the sound effects synthesized by wearable smart devices conform to brand specifications in terms of timbre, attack, and loudness, and avoids personalized strategies from damaging brand recognition.
[0010] In one implementation, a first vector is determined based on the acquired behavioral feature data and environmental feature data, specifically including: performing feature encoding on the behavioral feature data and environmental feature data respectively; and inputting the feature-encoded behavioral feature data and environmental feature data into a hierarchical Bayesian structure model to obtain the first vector. The behavioral feature data after feature encoding includes sound effect selection probability, sound effect skip rate, average dwell time, and volume adjustment gradient. The environmental feature data after feature encoding includes equivalent sound level, frequency band energy, time period encoding, and location label.
[0011] In one implementation, the method further includes: inputting the feature-encoded behavioral feature data and environmental feature data into a gradient boosting tree model to obtain a second vector, which is used to characterize the nonlinear feature interaction between the feature-encoded behavioral feature data and environmental feature data; inputting the first vector and the second vector into an attention-gated fusion model to obtain a second confidence level; and increasing the proportion of new sound effect trials when the second confidence level is less than a second threshold.
[0012] The above sound effect recommendation method inputs the feature-encoded behavioral feature data and environmental feature data into a hierarchical Bayesian structure model and a gradient boosting tree model, respectively. It not only uses the hierarchical Bayesian structure to generate a first vector representing the user's sound effect preference, but also uses the gradient boosting tree model to capture the complex nonlinear feature interaction between behavioral and environmental features to obtain a second vector. Furthermore, it uses an attention-gated fusion model to fuse the first and second vectors to obtain a second confidence score. When the second confidence score is less than a second threshold, the system automatically increases the proportion of new sound effect trials, thereby improving the learning efficiency of unexposed sound effect combinations.
[0013] One implementation also includes: real-time detection of abnormal situations; when an abnormal situation is detected or multiple detections fail, the default sound effect is used as the target sound effect.
[0014] In one implementation, the method further includes: uploading aggregated statistical metrics to the cloud so that the cloud can generate a group prior based on the aggregated statistical metrics, federated average, and differential privacy noise; receiving the group prior from the cloud and updating the local model and scene priority matrix based on the group prior.
[0015] Secondly, this application provides a sound effect recommendation device for use in wearable smart devices, comprising: The preference inference module is configured to determine a first vector based on the acquired behavioral feature data and environmental feature data; the decision-making module is configured to: determine the current scene and a first confidence level based on multi-source perception data; and determine multiple target sound effect templates based on the current scene, the first confidence level, and the scene priority matrix; the sound effect generation module is configured to determine the target divergence based on the first vector, the current scene, and multiple target sound effect templates, and determine the target sound effect based on the target divergence; the behavioral feature data is used to characterize the user's operation on different sound effects, the first vector is used to characterize the user's sound effect preference, and the target divergence is used to characterize the degree of distortion between the target sound effect template and the first vector.
[0016] Thirdly, this application provides a wearable smart device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any of the above-described sound effect recommendation methods.
[0017] In one implementation, wearable smart devices include smartwatches.
[0018] Fourthly, this application also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the above-described sound effect recommendation methods.
[0019] Fifthly, this application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of any of the above-described sound effect recommendation methods.
[0020] Compared with the prior art, the present invention has at least one of the following beneficial effects: 1. This method generates a first vector representing user sound effect preferences by fusing behavioral and environmental feature data. It then accurately identifies the current scene and its initial confidence level using multi-source perceptual data. Based on a scene priority matrix, it selects multiple target sound effect templates that match the current scene. Finally, it optimizes the selection of the target sound effect based on the target divergence (i.e., distortion level) between the first vector and each target sound effect template. This method, while ensuring scene adaptability and brand compliance, enables the learning of user preferences, effectively avoiding discrepancies between recommended sound effects and user expectations. Furthermore, the confidence level mechanism suppresses false recommendations caused by low-reliability preferences, providing users with sound effect choices that better suit their individual habits.
[0021] 2. After determining the target sound effect parameter vector through target divergence, on the one hand, it can directly combine with the sound effect library to quickly match compliant and user-preferred target sound effects, improving recommendation efficiency; on the other hand, it also supports generating sound effects to be verified based on the target sound effect parameter vector and constraint parameters, and introduces color consistency, pitch matching degree and loudness deviation penalty for weighted summation. Only when the weighted result is not less than the first threshold is the sound effect to be verified adopted as the target sound effect, thereby ensuring that the sound effects synthesized by wearable smart devices conform to brand specifications in terms of timbre, pitch, and loudness, and avoiding personalization strategies from damaging brand recognition.
[0022] 3. By inputting the encoded behavioral feature data and environmental feature data into the hierarchical Bayesian structure model and the gradient boosting tree model respectively, the system not only generates a first vector representing the user's sound effect preference using the hierarchical Bayesian structure, but also captures the complex nonlinear feature interaction between behavioral and environmental features to obtain a second vector through the gradient boosting tree model. Furthermore, the first and second vectors are fused through an attention-gated fusion model to obtain a second confidence level. When the second confidence level is less than the second threshold, the system automatically increases the proportion of new sound effect trials, thereby improving the learning efficiency of unexposed sound effect combinations. Attached Figure Description
[0023] The preferred embodiments will now be described in a clear and easy-to-understand manner, in conjunction with the accompanying drawings, to further explain the above-mentioned characteristics, technical features, advantages, and implementation methods of the present invention.
[0024] Figure 1 A flowchart of a sound effect recommendation method provided in an embodiment of this application is shown; Figure 2 This document illustrates a flowchart of a method for determining a target sound effect according to an embodiment of this application. Figure 3 This invention illustrates another flowchart for determining a target sound effect according to an embodiment of the present application; Figure 4 This document illustrates a flowchart of a process for processing feature-encoded behavioral feature data and environmental feature data, as provided in an embodiment of this application. Figure 5 This diagram illustrates a structural block diagram of a sound effect recommendation device provided in an embodiment of this application. Figure 6 A structural block diagram of a wearable smart device provided in an embodiment of this application is shown. Detailed Implementation
[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the specific implementation methods of the present invention will be described below with reference to the accompanying drawings. Obviously, the drawings described below are merely some embodiments of the present invention. For those skilled in the art, other drawings and other implementation methods can be obtained based on these drawings without any creative effort.
[0026] To keep the drawings concise, each figure only schematically shows the parts relevant to the invention, and these do not represent the actual structure of the product. Furthermore, to facilitate understanding, in some figures, only one of components with the same structure or function is schematically depicted, or only one is labeled. In this document, "one" not only means "only one," but can also mean "more than one."
[0027] It should also be further understood that the term “and / or” as used in this application specification and the appended claims means any combination of one or more of the associated listed items and all possible combinations, and includes such combinations.
[0028] In this document, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation," "connection," and "linking" should be interpreted broadly. For example, they can refer to fixed connections, detachable connections, or integral connections; they can refer to mechanical connections or electrical connections; they can refer to direct connections or indirect connections through an intermediate medium; and they can refer to the internal communication between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.
[0029] Furthermore, in the description of this application, the terms "first," "second," etc., are used only to distinguish descriptions and should not be construed as indicating or implying relative importance.
[0030] It should be noted that the above embodiments can be freely combined as needed. The above are merely preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
[0031] In recent years, wearable smart devices have typically used sound to deliver notifications, reminders, interactive feedback, or entertainment content to users, making sound effects one of the core mediums of human-computer interaction. Appropriate sound effect design can not only enhance the user experience but also play a positive role in scenarios such as education, safety warnings, and emotional soothing. For example, soft prompts can be played in classroom mode to avoid distractions, while rhythmic sound effects can be used in sports scenarios to enhance motivation.
[0032] Currently, a common method for recommending sound effects in vehicles involves obtaining a set of candidate sound effects based on the user's preset sound effect preferences and other features. Then, a weighted sum is calculated based on the user's preset sound effect preferences and features such as sound effect popularity, leading to a sound effect recommendation. However, the sound effects recommended by this method still differ from the user's expectations, making it difficult to provide sound effect choices that better suit individual user habits.
[0033] In another method of sound effect recommendation applied to wearable smart devices, the sound effect themes in wearable smart devices are usually fixed templates that cannot be dynamically adjusted according to individual preferences and the current scene, making it difficult to provide users with sound effect choices that better suit their individual habits.
[0034] Therefore, this application provides a sound effect recommendation method to solve the technical problems existing in the prior art.
[0035] The following explanation is based on the accompanying diagram: Reference Appendix Figure 1 The diagram illustrates a flowchart of a sound effect recommendation method provided in an embodiment of this application. Figure 1 As shown, it includes: S100, determine the first vector based on the acquired behavioral feature data and environmental feature data.
[0036] S110, based on multi-source perception data, determines the current scene and the first confidence level.
[0037] S120 determines multiple target sound effect templates based on the current scene, first confidence level, and scene priority matrix.
[0038] S130: Based on the first vector, the current scene, and multiple target sound effect templates, determine the target divergence, and based on the target divergence, determine the target sound effect. Behavioral feature data is used to characterize the user's actions towards different sound effects, the first vector is used to characterize the user's sound effect preferences, and the target divergence is used to characterize the degree of distortion between the target sound effect template and the first vector.
[0039] Wearable smart devices can periodically collect behavioral and environmental feature data, with a sampling time window of up to 5 seconds. Behavioral feature data includes user actions such as selecting different sound effects, skipping, dwell time, and volume adjustment. Environmental feature data includes ambient noise spectrum energy, signal-to-noise ratio, time period labels, and geographical context (home / school / outdoor). Sensors used to collect environmental feature data can include, but are not limited to, microphone arrays, single microphones, and other voice sensors. This application does not limit the type of sensor used to collect environmental feature data.
[0040] Based on the acquired behavioral and environmental feature data, and combined with feature encoding processing and a hierarchical Bayesian structure model, a first vector (or sound effect preference vector) is obtained. The dimensions of the first vector can include timbre, rhythm, and loudness. Wearable smart devices can classify scenes based on multi-source sensor data. For example, multi-source sensor data can be input into a cascaded structure of a lightweight convolutional network and a hidden Markov model to output the current scene (or scene label) and the first confidence score.
[0041] A scene priority matrix is constructed (this scene priority matrix can be pre-built and stored in the wearable smart device; the device that pre-builds the scene priority matrix can include, but is not limited to, wearable smart devices, mobile phones, computers, tablets, and other smart terminals; this scene priority matrix can also be constructed when the wearable smart device executes the sound effect recommendation process; this scene priority matrix can also be the default setting when the wearable smart device leaves the factory. This application does not limit the source or construction method of the scene priority matrix). The default priority is classroom > night > exercise > outdoor noise > normal, and dynamic adjustment is allowed according to user policies. For example, when a parent enables "night priority", the corresponding weights of the current scene and the night scene are swapped. Specific scene priority matrices can include: Classroom: Priority 5, volume limit 40dBA, initial mute 300ms replaced with haptic reminder, prohibiting the use of high-frequency sharp tones. Night: Priority 4, volume limit 30dBA, envelope rise and fall exceeding 150ms and duration less than 3s, retaining the brand's sound head but reducing high-frequency energy. Sports: Priority 3, volume cap 65 dBA, main frequency 2-4 kHz, initial rise less than 100ms, enhanced attack to ensure brand recognition. Outdoor Noise: Priority 2, volume cap 75 dBA, bandpass filter 1.5-3 kHz, dynamic compression ratio 2:1, can be overlaid with brand rhythm bands. Normal: Priority 1, volume cap 55 dBA, adaptive parameters, fully preserves the brand's sound library. Through a pre-constructed scene priority matrix, this application can automatically adjust the volume cap according to the scene, effectively reducing noise interference in classroom and nighttime scenarios, thereby improving the user experience.
[0042] Wearable smart devices can determine multiple target sound effect templates based on the current scene, the first confidence level, and a scene priority matrix. Specifically, the current scene obtained by the wearable smart device can fall into several categories. For example, if the first confidence level of the obtained current scene is less than a preset confidence threshold, the current scene can be adjusted to "normal" by combining it with the scene priority matrix. Similarly, if the obtained current scene corresponds to multiple scenes, such as a classroom and nighttime, the current scene can be determined as a classroom based on the priority of these two scenes in the scene priority matrix. Other cases are similar and will not be elaborated upon here. After determining the current scene, the multiple target sound effect templates (or brand preset strategy templates) corresponding to the current scene can then be identified.
[0043] Substitute the first vector, the current scene, and multiple target sound effect templates into the policy mapping function f(p,s)=argmin_kKL(p||T_k(s)). Here, s is the current scene, T_k(s) are the multiple target sound effect templates corresponding to the current scene, p is the first vector, and KL is the target divergence. The significance of the policy mapping function is to obtain the target divergence corresponding to each target sound effect template. Since the target divergence reflects the degree of distortion between the target sound effect template and the first vector, the target sound effect template corresponding to the minimum target divergence can be used as the target sound effect parameter vector.
[0044] Wearable smart devices can determine target sound effects based on target parameter vectors. For example, they can search a preset sound effect library based on the target parameter vector to obtain the target sound effect; or, they can directly generate the target sound effect based on the target sound effect parameter vector.
[0045] This application's embodiments generate a first vector representing user sound effect preferences by fusing behavioral feature data and environmental feature data. It then accurately identifies the current scene and its first confidence level using multi-source perceptual data. Next, it filters multiple target sound effect templates that match the current scene based on a scene priority matrix. Finally, it selects the optimal target sound effect based on the target divergence (i.e., distortion level) between the first vector and each target sound effect template. This method, while ensuring scene adaptability and brand consistency, enables the learning of user preferences, effectively avoiding discrepancies between recommended sound effects and user expectations. Furthermore, the confidence level mechanism suppresses false recommendations caused by low-reliability preferences, providing users with sound effect choices that better suit their individual habits.
[0046] Reference Appendix Figure 2 This illustrates a flowchart of a method for determining a target sound effect according to an embodiment of this application. Figure 2 As shown, it includes: S200, determine the target sound effect parameter vector based on the target divergence.
[0047] S210, determine the target sound effect based on the target sound effect parameter vector and the sound effect library.
[0048] Continue to refer to the appendix Figure 3 This illustrates another flowchart for determining a target sound effect provided in an embodiment of this application. For example... Figure 3 As shown, it includes: S300 determines the target sound effect parameter vector based on the target divergence.
[0049] S310 generates the sound effect to be verified based on the target sound effect parameter vector and constraint parameters.
[0050] S320 performs a weighted summation based on the color consistency, attack matching, and loudness deviation penalty of the sound effect to be verified.
[0051] S330: When the weighted summation result is not less than the first threshold, the sound effect to be verified is taken as the target sound effect.
[0052] After obtaining the first vector, the current scene, and multiple target sound effect templates, the wearable smart device substitutes the first vector, the current scene, and multiple target sound effect templates into the strategy mapping function f(p,s)=argmin_k KL(p||T_k(s)), thereby obtaining the target sound effect parameter vector.
[0053] In one implementation of determining a target sound effect, sound effect segments that satisfy the target sound effect parameter vector can be retrieved from the sound effect library using a dual threshold of cosine similarity and perceptual hashing, and these sound effect segments are then used as the target sound effect. The sound effect library can be pre-built and stored in a wearable smart device. The device that pre-builds the scene priority matrix can include, but is not limited to, wearable smart devices, mobile phones, computers, tablets, and other smart terminals. The sound effect library can also be built when the wearable smart device executes a sound effect recommendation process; it can also be the default setting when the wearable smart device leaves the factory. This application does not limit the source or construction method of the sound effect library.
[0054] In another approach to determining the target sound effect, it can be generated directly from the target sound effect parameter vector. For example, particle synthesis or wavetable splicing can be performed on the target sound effect parameter vector and constraint parameters (or brand constraint parameter domain), with a dynamic compressor controlling the maximum sound pressure level, limited to ≤40dBA for classroom scenarios and ≤30dBA for nighttime scenarios. Constraint parameters can include, but are not limited to, color, logo attack, and allowable loudness range. After obtaining the sound effect to be verified, a brand consistency score can be performed, which involves a weighted sum of the color consistency, attack matching, and loudness deviation penalty of the sound effect to be verified. When the weighted sum is not less than a first threshold (which can be 0.85 or set according to user needs), the sound effect to be verified is adopted as the target sound effect. The sum of the weights corresponding to color consistency, attack matching, and loudness deviation penalty is 1. When the weighted sum is less than the first threshold, it automatically reverts to the most recently verified sound effect scheme.
[0055] This application may or may not limit the conditions for implementing the target sound effect. For example, this application may limit the generation of the target sound effect directly based on the target sound effect parameter vector when there is no sound effect segment in the sound effect library that satisfies the target sound effect parameter vector.
[0056] After determining the target sound effect parameter vector through target divergence in this embodiment, on the one hand, it can directly combine with the sound effect library to quickly match compliant and user-preferred target sound effects, improving recommendation efficiency; on the other hand, it also supports generating sound effects to be verified based on the target sound effect parameter vector and constraint parameters, and introduces color consistency, pitch matching degree and loudness deviation penalty for weighted summation. Only when the weighted result is not less than the first threshold is the sound effect to be verified adopted as the target sound effect, thereby ensuring that the sound effects synthesized by wearable smart devices conform to brand specifications in terms of timbre, pitch, and loudness, and avoiding personalization strategies from damaging brand recognition.
[0057] In one embodiment of this application, a first vector is determined based on the acquired behavioral feature data and environmental feature data, specifically including: performing feature encoding on the behavioral feature data and environmental feature data respectively; inputting the feature-encoded behavioral feature data and environmental feature data into a hierarchical Bayesian structure model to obtain the first vector; wherein the feature-encoded behavioral feature data includes sound effect selection probability, sound effect skip rate, average dwell time and volume adjustment gradient, and the feature-encoded environmental feature data includes equivalent sound level, frequency band energy, time period encoding and location label.
[0058] After obtaining the behavioral and environmental feature data with encoded features, they can be processed using other models. For example, see the attached document. Figure 4 The illustrated flowchart shows a process for processing feature-encoded behavioral feature data and environmental feature data, according to an embodiment of this application. Figure 4 As shown, it includes: S400: Input the feature-encoded behavioral feature data and environmental feature data into the gradient boosting tree model to obtain the second vector. The second vector is used to characterize the non-linear feature interaction between the feature-encoded behavioral feature data and environmental feature data. S410, input the first vector and the second vector into the attention-gated fusion model to obtain the second confidence score; S420: When the second confidence level is less than the second threshold, increase the proportion of new sound effects to be tested.
[0059] Feature encoding is performed on behavioral feature data and environmental feature data respectively. For example, behavioral feature data can be encoded into a vector x_beh=[selection probability, skip rate, average dwell time, volume adjustment gradient], and environmental feature data can be encoded into a vector x_env=[equivalent sound level, frequency band energy, time period encoding, location label].
[0060] Construct a multi-scale feature set, which may include short-term window behavior features, daily cycle preference trends, and group prior distribution. Input the feature-encoded behavioral and environmental feature data into a hierarchical Bayesian structural model. The individual layer of the hierarchical Bayesian structural model uses Dirichlet(α) as the prior and is incrementally updated based on the interaction results: α_i←α_i+Δi. Then, the hierarchical Bayesian structural model can output the first vector.
[0061] The encoded behavioral and environmental feature data are input into a gradient boosting tree model to obtain a second vector (or feature interaction vector). Then, the first and second vectors can be input into an attention-gated fusion model to obtain a preference probability vector and a second confidence level. The second confidence level can be represented as conf = softmax(W·[p;g]), where p is the first vector and g is the second vector. When the second confidence level is less than a second threshold, the proportion of new sound effect trials is increased. That is, when the second confidence level is less than the second threshold, a multi-armed slot machine exploration strategy is triggered, proposing underexposed sound effect combinations with probability ε_explore and recording the exploration gains to update the weights.
[0062] In this embodiment, the behavioral feature data and environmental feature data after feature encoding are respectively input into a hierarchical Bayesian structure model and a gradient boosting tree model. The hierarchical Bayesian structure is used to generate a first vector representing the user's sound effect preference. The gradient boosting tree model is used to capture the complex nonlinear feature interaction between behavior and environmental features to obtain a second vector. Furthermore, the first vector and the second vector are fused through an attention-gated fusion model to obtain a second confidence level. When the second confidence level is less than a second threshold, the system automatically increases the proportion of new sound effect trials, thereby improving the learning efficiency of unexposed sound effect combinations.
[0063] In one embodiment of this application, it further includes: real-time detection of abnormal situations; when an abnormal situation is detected or multiple detection failures occur, a default sound effect is used as the target sound effect.
[0064] This application introduces an anomaly detector to monitor abnormal situations such as microphone blockage and sudden noise increases. When the detection fails N times in a row (default N=3), it will automatically revert to the brand's default sound effect and prompt parents to maintain it.
[0065] In one embodiment of this application, the method further includes: uploading aggregated statistical metrics to the cloud so that the cloud generates a group prior based on the aggregated statistical metrics, federated average, and differential privacy noise; receiving the group prior from the cloud and updating the local model and scene priority matrix based on the group prior.
[0066] After obtaining the target sound effect, the wearable smart device can generate an explanation report (text and icons) describing the preference input, scene recognition results, priority matrix decision, and final sound effect parameters, and upload the explanation report to the parent's end. The parent's end can view the explanation report and configure supervision policies, which can include whitelisting or blacklisting sound effects, scene start / stop, time period restrictions, and maximum volume threshold settings. The wearable smart device can output the target sound effect according to the supervision policy.
[0067] Wearable smart devices can also upload aggregated statistical metrics and encrypted log fingerprints to the cloud. The cloud can generate a group prior based on the aggregated statistical metrics, federated averages, and differential privacy noise. Wearable smart devices can update their local models and priority matrices based on the group priors sent from the cloud. Aggregated statistical metrics can include the mean of preference vectors, confidence intervals, and the number of scene triggers, and Laplace noise is injected into the statistics to ensure ε-differential privacy. Wearable smart devices can record the evidence chain for each policy decision (preference input, scene label, brand constraint, and final output) and provide visual explanations. A 24-hour rolling cache is set up, which uses AES-256 encryption and fingerprint verification. The cache is automatically destroyed after expiration, and the hash fingerprint is retained in the cloud for auditing purposes.
[0068] In one embodiment of this application, the method further includes: using incremental Bayesian update and online gradient descent joint correction model based on user feedback buttons, voice commands, or skip actions.
[0069] Reference Appendix Figure 5 The diagram illustrates a structural block diagram of a sound effect recommendation device provided in an embodiment of this application. Figure 5 As shown, the device 500 includes: a preference inference module 510, configured to determine a first vector based on acquired behavioral feature data and environmental feature data; a decision module 520, configured to: determine the current scene and a first confidence level based on multi-source perception data; and determine multiple target sound effect templates based on the current scene, the first confidence level, and a scene priority matrix; and a sound effect generation module 530, configured to determine a target divergence based on the first vector, the current scene, and the multiple target sound effect templates, and to determine a target sound effect based on the target divergence; the behavioral feature data is used to characterize the user's operation on different sound effects, the first vector is used to characterize the user's sound effect preference, and the target divergence is used to characterize the degree of distortion between the target sound effect template and the first vector.
[0070] This application's embodiments generate a first vector representing user sound effect preferences by fusing behavioral feature data and environmental feature data. It then accurately identifies the current scene and its first confidence level using multi-source perceptual data. Next, it filters multiple target sound effect templates that match the current scene based on a scene priority matrix. Finally, it selects the optimal target sound effect based on the target divergence (i.e., distortion level) between the first vector and each target sound effect template. This method, while ensuring scene adaptability and brand consistency, enables the learning of user preferences, effectively avoiding discrepancies between recommended sound effects and user expectations. Furthermore, the confidence level mechanism suppresses false recommendations caused by low-reliability preferences, providing users with sound effect choices that better suit their individual habits.
[0071] Reference Appendix Figure 6 This application also provides a wearable smart device 600, which includes a memory 610, a processor 620, and a computer program stored on the memory 610. The processor 620 executes the computer program to implement the steps of the sound effect recommendation method of any of the above embodiments.
[0072] The memory 610 can be non-volatile memory (NVM), such as, but not limited to, semiconductor non-volatile memory, disk storage, or optical storage. Semiconductor non-volatile memory includes, but is not limited to, read-only memory (ROM) or flash memory, such as mask ROM, programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), NAND flash memory, or NOR flash memory.
[0073] The memory 610 can also be volatile memory, such as random access memory (RAM). RAM includes, for example, static random-access memory (SRAM) or dynamic random-access memory (DRAM). DRAM includes, for example, synchronous dynamic RAM (SDRAM) or double data rate SDRAM (DDR). With the development of technology, DDR includes, but is not limited to, DDR1, DDR2, DDR3, ..., DDR5, and may also include future DDR6.
[0074] Processor 620 is a circuit with signal processing capabilities. In one example, the processor can be a circuit with instruction read and execute capabilities; such as a central processing unit (CPU), microcontroller unit (MCU), microprocessor unit (MPU), graphics processing unit (GPU), or digital signal processor (DSP). In another example, the processor can realize its processing capabilities through the logical relationships of hardware circuits, which can be fixed or reconfigurable; for example, the processor can be a dedicated processor, such as a processor implemented with an application-specific integrated circuit (ASIC), which realizes its processing capabilities through the design of the logical relationships between components within the circuit; or a processor implemented with a programmable logic device (PLD), which realizes its processing capabilities by configuring the logical relationships between logic devices through configuration files; for example, a processor implemented with a field-programmable gate array (FPGA). In another example, the processor can be a hardware circuit designed for artificial intelligence, which can be understood as an ASIC, such as a neural network processing unit (NPU), tensor processing unit (TPU), deep learning processing unit (DPU), etc. This application is not limited to the type of processor.
[0075] The wearable smart device used in this application embodiment is basically similar to the method embodiment, so the description is relatively simple. For relevant details, please refer to the description of the method embodiment.
[0076] This application also provides a computer-readable storage medium storing a computer program thereon, which, when executed by a processor, implements the steps of the sound effect recommendation method of any of the above embodiments.
[0077] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the sound effect recommendation method of any of the above embodiments.
[0078] It should be noted that the above embodiments can be freely combined as needed. The above are merely preferred embodiments of the present invention. It should be pointed out that for those skilled in the art, several improvements and modifications can be made without departing from the principle of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.
Claims
1. A sound effect recommendation method, applied to wearable smart devices, characterized in that, include: Based on the acquired behavioral and environmental feature data, a first vector is determined; Based on multi-source sensing data, determine the current scene and the first confidence level; Based on the current scene, the first confidence level, and the scene priority matrix, multiple target sound effect templates are determined; Based on the first vector, the current scene, and multiple target sound effect templates, a target divergence is determined, and a target sound effect is determined based on the target divergence. The behavioral feature data is used to characterize the user's operation on different sound effects, the first vector is used to characterize the user's sound effect preference, and the target divergence is used to characterize the degree of distortion between the target sound effect template and the first vector.
2. The sound effect recommendation method according to claim 1, characterized in that, The step of determining the target sound effect based on the target divergence specifically includes: Based on the target divergence, determine the target sound effect parameter vector; The target sound effect is determined based on the target sound effect parameter vector and the sound effect library.
3. The sound effect recommendation method according to claim 1, characterized in that, The step of determining the target sound effect based on the target divergence specifically includes: Based on the target divergence, determine the target sound effect parameter vector; Based on the target sound effect parameter vector and constraint parameters, generate the sound effect to be verified; The weighted sum is calculated based on the color consistency, attack matching, and loudness deviation penalty of the sound effect to be verified. When the weighted summation result is not less than the first threshold, the sound effect to be verified is taken as the target sound effect.
4. The sound effect recommendation method according to claim 1, characterized in that, The step of determining the first vector based on the acquired behavioral feature data and environmental feature data specifically includes: The behavioral feature data and environmental feature data are respectively feature-encoded; The encoded behavioral feature data and environmental feature data are input into a hierarchical Bayesian structure model to obtain the first vector. The behavioral feature data after feature encoding includes sound effect selection probability, sound effect skip rate, average dwell time and volume adjustment gradient, and the environmental feature data after feature encoding includes equivalent sound level, frequency band energy, time period encoding and location label.
5. The sound effect recommendation method according to claim 4, characterized in that, Also includes: The encoded behavioral feature data and environmental feature data are input into the gradient boosting tree model to obtain a second vector, which is used to characterize the nonlinear feature interaction between the encoded behavioral feature data and environmental feature data. The first vector and the second vector are input into the attention-gated fusion model to obtain the second confidence level; When the second confidence level is less than the second threshold, the proportion of new sound effects to be tested is increased.
6. The sound effect recommendation method according to any one of claims 1-5, characterized in that, Also includes: Real-time detection of abnormal situations; When the above-mentioned abnormal situation is detected or multiple detection failures occur, the default sound effect will be used as the target sound effect.
7. The sound effect recommendation method according to any one of claims 1-5, characterized in that, Also includes: Upload aggregated statistical metrics to the cloud so that the cloud can generate a population prior based on the aggregated statistical metrics, federated average, and differential privacy noise; Receive the group prior from the cloud, and update the local model and the scene priority matrix based on the group prior.
8. A sound effect recommendation device, applied to wearable smart devices, characterized in that, include: The preference inference module is configured to determine a first vector based on the acquired behavioral feature data and environmental feature data; The decision-making module is configured to: determine the current scene and a first confidence level based on multi-source perception data; and determine multiple target sound effect templates based on the current scene, the first confidence level, and the scene priority matrix. The sound effect generation module is configured to determine a target divergence based on the first vector, the current scene, and multiple target sound effect templates, and to determine a target sound effect based on the target divergence. The behavioral feature data is used to characterize the user's operation on different sound effects, the first vector is used to characterize the user's sound effect preference, and the target divergence is used to characterize the degree of distortion between the target sound effect template and the first vector.
9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the computer program implements the steps of the sound effect recommendation method according to any one of claims 1-7.
10. A wearable smart device, comprising a memory, a processor, and a computer program stored in the memory, characterized in that, The processor executes the computer program to implement the steps of the sound effect recommendation method according to any one of claims 1-7.