Method for realizing personalized sound effect adjustment of loudspeaker through collaborative filtering algorithm

Abstract

The invention discloses a loudspeaker personalized sound effect adjustment method realized through a collaborative filtering algorithm. The method comprises the following steps: S1, constructing a user behavior data set and an audio feature data set; s2, preprocessing the data set to obtain user features and audio features; s3, constructing and training a multi-branch neural network model, and generating personalized recommendation; s4, using a user-based collaborative filtering algorithm to generate sound effect setting recommendations liked by other users and similar to the sound effect preference of the current user; s5, using a project-based collaborative filtering algorithm to generate sound effect setting recommendation of the audio similar to the current audio characteristics; s6, dynamically adjusting the recommendation weight, and generating diversified recommendation sound effect setting options; and S7, collecting feedback of the user to the recommended sound effect setting, and performing model optimization according to the latest user data and the feedback. According to the method, the multi-branch neural network and the collaborative filtering algorithm are combined, personalized sound effect recommendation is realized, and the method has the advantages of high precision and dynamic adaptability.

Description

Technical Field

[0001] The present invention relates to the technical field of loudspeaker sound effect adjustment, and in particular to a method for adjusting personalized loudspeaker sound effects through a collaborative filtering algorithm. Background Art

[0002] In the development of modern audio equipment, users' demands for sound quality and personalized sound effects are constantly increasing. Smart speakers, headphones, car audio and other devices not only need to provide high-quality hardware experience, but also need to have intelligent sound effect adjustment functions to meet the listening preferences of different users. Traditional sound effect adjustment methods mainly rely on manual adjustment of equalizers and volume controls. These methods require users to have certain audio knowledge and require users to actively set them up, lacking the characteristics of intelligence and automation.

[0003] In recent years, sound optimization technologies based on machine learning and artificial intelligence have been gradually applied to high-end audio devices. These technologies usually analyze the user's historical listening data and preferences to automatically adjust the sound parameters in the hope of providing the best listening experience. For example, some systems use simple filters and signal processing algorithms, such as mean filtering and adaptive filtering, to enhance audio signals. However, these technologies mainly focus on the signal processing level and fail to make full use of the user's personalized data and behavioral information, resulting in insufficient personalization of sound adjustment. In addition, such methods usually rely on the user's historical data for recommendations, lack dynamic adaptation capabilities, and cannot respond to changes in user preferences in real time.

[0004] Collaborative filtering algorithm is a technology widely used in recommendation systems. It analyzes users' historical behavior data and preferences, finds similarities between users, and then recommends content that similar users like. This algorithm has achieved good results in video recommendation, product recommendation and other fields. However, in the field of sound effect adjustment, the application of collaborative filtering algorithm is still relatively limited. Existing sound effect recommendation systems mainly rely on user-based collaborative filtering, which recommends sound effect settings by calculating the similarity between users. Although this method can reflect the common preferences of users to a certain extent, it ignores the characteristics and personalized needs of the audio content itself, resulting in the recommended sound effect settings may not fully meet the specific preferences of individual users.

[0005] In the existing sound effect recommendation systems, the application of multi-branch neural networks has gradually attracted attention. Multi-branch neural networks can process different types of input data at the same time, such as user features and audio features, to generate more accurate recommendation results. This method can capture more complex user behavior patterns and audio features through the multi-level feature extraction capabilities of neural networks. However, current applications mostly combine multi-branch neural networks with a single recommendation method, lacking the comprehensive use of user and audio features, resulting in insufficient performance of the recommendation system when faced with diverse user needs. In addition, the training of multi-branch neural networks is usually performed offline, making it difficult to update model parameters in real time to adapt to the dynamic changes in user needs.

[0006] In existing recommendation systems, dynamic weight adjustment mechanisms have not been fully utilized. Most systems use fixed weights to balance different recommendation methods, ignoring the real-time changes in user preferences. As user behaviors and preferences continue to change, fixed weights may result in recommendation results that do not accurately reflect the user's current needs. The dynamic weight adjustment mechanism can dynamically adjust the weights of each recommendation method based on the user's immediate feedback, allowing the system to adapt to user changes in real time, thereby providing more personalized sound effect recommendations.

[0007] Therefore, how to provide a method for personalized sound effect adjustment of a speaker implemented through a collaborative filtering algorithm is an urgent problem to be solved by those skilled in the art. Summary of the invention

[0008] One purpose of the present invention is to propose a method for personalized sound effect adjustment of a speaker through a collaborative filtering algorithm. The present invention combines a multi-branch neural network and a collaborative filtering algorithm, and realizes personalized sound effect recommendation through deep feature extraction and collective wisdom fusion. The system uses a multi-branch neural network to extract user and audio features, and combines user-based and project-based collaborative filtering algorithms to generate accurate and personalized sound effect setting recommendations. At the same time, a dynamic weight adjustment mechanism and online learning function are introduced to enable the system to respond to user feedback in real time and dynamically adjust the recommendation strategy, thereby improving recommendation accuracy and user satisfaction. This method has high accuracy, dynamic adaptability and comprehensiveness, and significantly improves the user's listening experience.

[0009] According to an embodiment of the present invention, a method for adjusting personalized sound effects of a speaker implemented by a collaborative filtering algorithm comprises the following steps:

[0010] S1. Collect user listening preference data and historical behavior data to build a user behavior dataset; collect audio feature data to build an audio feature dataset;

[0011] S2. Preprocessing the user behavior dataset and the audio feature dataset, including data cleaning, normalization, and feature extraction, to obtain user features and audio features;

[0012] S3. Build and train a multi-branch neural network model, including a user feature branch, an audio feature branch, a feature fusion layer, and an output layer, and generate personalized recommendations;

[0013] S4. Using a user-based collaborative filtering algorithm, the similarity between users is calculated based on the historical data of user sound effect settings, and a recommendation of sound effect settings that are similar to the current user's sound effect preferences is generated;

[0014] S5. Using an item-based collaborative filtering algorithm, the similarity between audio items is calculated according to audio features, and a sound effect setting recommendation is generated for audio with similar features to the current audio;

[0015] S6. Integrate the personalized recommendation generated by the multi-branch neural network and the recommendation results of the collaborative filtering algorithm, dynamically adjust the recommendation weight, and generate a variety of recommended sound effect setting options;

[0016] S7. Collect user feedback on recommended sound effect settings, regularly update the multi-branch neural network and collaborative filtering modules, and optimize the model based on the latest user data and feedback.

[0017] Furthermore, the S3 specifically includes:

[0018] S31, input user feature vector U = [u1, u2, ..., u n ], where u i Represents a single feature of a user. The temporal pattern in the user feature is extracted through a one-dimensional convolutional neural network. The convolution operation is defined as:

[0019] U c =LeakyReLU(W c *U+b c );

[0020] Among them, U c represents the feature representation after the convolution operation, W c represents the convolution kernel, b c represents the bias vector, * represents the convolution operation, and LeakyReLU represents the activation function, which is used to better handle negative information;

[0021] S32, U c Long short-term memory network with adaptive gating mechanism, which adds adaptive gates to the standard LSTM to dynamically adjust the flow of information:

[0022] g t =σ(Wg [h t-1 ,U c ]+b g );

[0023] f t =σ(W f [h t-1 ,U c ]+b f )⊙g t ;

[0024] i t =σ(W i [h t-1 ,U c ]+b i )⊙(1-g t );

[0025] c t =f t ⊙c t-1 +i t ⊙tanh(W c [h t-1 ,U c ]+b c );

[0026] o t =σ(W o [h t-1 ,U c ]+b o );

[0027] h t =o t ⊙tanh(c t );

[0028] Among them, g t Represents the adaptation gate vector, which is used to dynamically adjust the output of the forget gate and the input gate, which can improve the selective filtering of information, especially for changes in specific user characteristics. g represents the weight matrix of the adaptation gate, b g Represents the bias vector of the adaptation gate, which is used to dynamically adjust the output of other gates, f t represents the forget gate, i t represents the input gate vector, c t represents the cell state vector, o t represents the input gate vector, h t represents the hidden state vector, W f Represents the weight matrix of the forget gate, W i represents the weight matrix of the input gate, W c The weight matrix representing the cell state, W orepresents the weight matrix of the input gate, ⊙ represents the element multiplication operation, σ represents the activation function, and the implicit representation U′ of the user feature is obtained;

[0029] S33, input audio feature vector A = [a1, a2, ..., a m ], where a i Represents a single feature of the audio, and extracts the global correlation between audio features through a multi-head self-attention mechanism. For each head k:

[0030]

[0031] Among them, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k Represents the dimension of the key vector. After connecting the multiple results, the implicit representation A′ of the audio feature is obtained through a fully connected network:

[0032] A′=Concat(head1,head2,…,head h )W O ;

[0033] Among them, head k represents the output of the kth head, W O represents the output transformation matrix;

[0034] S34. Introduce the attention mechanism to calculate the importance weights of user features and audio features, and connect the outputs U′ and A′ of the user feature branch and the audio feature branch to form a fused feature vector F:

[0035] F = [U′, A′];

[0036] The importance weight α of each feature is calculated through the attention network i :

[0037]

[0038] Among them, f i represents the representation of feature i, and w represents the weight vector of the attention network;

[0039] The weighted fusion feature vector F′ is obtained by weighted summation:

[0040]

[0041] S35, input the weighted fusion feature vector F′ to the output layer, and generate personalized recommended sound effect setting parameters P through the fully connected network:

[0042] P=W o1 F′+b o1 ;

[0043] Among them, W o1 represents the weight matrix of the output layer, b o1 Represents the bias vector of the output layer;

[0044] S36. Introduce the user experience evaluation score E into the loss function, and combine the weighted influence of the user's historical behavior on the current sound effect preference to define the comprehensive loss function L:

[0045]

[0046] Where N is the number of samples, p i represents the sound effect setting parameters predicted by the model, represents the user's actual preference parameter, E represents the user experience evaluation score, M represents the number of user historical behaviors, and w j represents the influence weight of the jth historical behavior on the current preference, h j represents the jth historical behavior score of the user, λ1 and λ2 represent the importance coefficients for adjusting different loss terms:

[0047] Calculate the historical behavior score h j :

[0048] h j =σ(W h z j +b h );

[0049] Among them, z j The feature vector representing the jth historical behavior, W h represents the weight matrix, b h represents the bias vector;

[0050] S37. Generate personalized recommendation result R1 through the trained multi-branch neural network model.

[0051] Furthermore, the S4 specifically includes:

[0052] S41. Based on the historical sound effect setting data of the user, construct a user-sound effect setting matrix M=[m ij ], where m ij Represents user P i Set S for sound effects j preference ratings;

[0053] S42, normalizing the user-sound effect setting matrix, normalizing the score of each user to a standard normal distribution with a mean of 0 and a standard deviation of 1;

[0054] S43, calculate the multi-dimensional similarity between users, the similarity calculation uses weighted cosine similarity, combined with the dynamic weight of the sound effect feature vector W(t) = [w1(t), w2(t), ..., w n (t)] and the time decay factor α(t):

[0055]

[0056] Among them, sim(P i ,P k ) represents user P i and user P k The similarity between ij Represents user P i Set S for sound effects j Rating, m kj Represents user P k Set S for sound effects j Rating, w j (t) indicates the sound effect setting S j The dynamic feature weight at time t, α(t) represents the time decay factor, and n represents the number of sound effect settings;

[0057] S44. For each target user P, select the k users with the highest similarity as the similar user set N(P):

[0058] N(P)={P k |khighestsim(P,P k )};

[0059] Among them, k represents the number of similar users;

[0060] S45: Generate a recommended sound effect setting for the target user based on the sound effect setting history of the similar user set N(P). j , calculate the weighted average rating in a set of similar users:

[0061]

[0062] in, S represents the recommendation for user P j The expected score of sim(P,P k ) represents user P and similar user P k The similarity between kj Represents similar users P k Set S for sound effects j Rating, w j (t) indicates the sound effect setting S j The dynamic feature weight at time t, β kj Represents similar users Pk Set S for sound effects j The personalized weighting coefficient of

[0063] The formula introduces dynamic feature weights, time decay factors and personalized weighting coefficients, which can capture the dynamic characteristics of user preferences changing over time and reflect personalized weights, while improving the model's ability to capture user behavior and sound effect features, thereby providing more accurate and personalized recommendations.

[0064] S46. By sorting the expected scores of all sound effect settings, a recommendation R2 of sound effect settings preferred by other users who have similar sound effect preferences to the current user is generated.

[0065] Furthermore, the S5 specifically includes:

[0066] S51, collect the multi-dimensional attributes of all sound effect settings, expressed as a sound effect setting feature vector G = [g1, g2, ..., g m ], where g i A single property representing a sound effect setting;

[0067] S52, normalizing the sound effect setting feature vector G, normalizing the value range of all attributes to [0, 1] to eliminate the influence of dimension;

[0068] S53, constructing a sound effect setting-attribute matrix B based on the sound effect setting attribute vector G = [b ij ], where b ij Indicates sound setting X i The value of the jth attribute is used to calculate the similarity between each pair of sound effect settings;

[0069] S54, for each target sound effect setting X, selecting l sound effect settings with the highest similarity as a similar sound effect setting set Y(X), where l represents the number of similar sound effect settings;

[0070] S55: Based on the similar sound effect setting set Y(X) and the user preference data, generate a recommended sound effect setting for the target sound effect setting, and i Calculates a weighted score among a set of similar sound settings

[0071] S56. Generate a sound effect setting recommendation R3 for audio with similar features to the current audio by sorting the expected scores of all sound effect settings.

[0072] Furthermore, the S6 specifically includes:

[0073] S61, calculating the score s of the recommended sound effect settings in the sets R1, R2 and R3 i , for the recommendation score of multi-branch neural network

[0074]

[0075] Among them, N1 represents the set of all users, sim(P,P k ) represents the target user P and user P k The similarity between k,i Represents user P k Set S for sound effects i The actual rating of

[0076] S62. User-based collaborative filtering recommendation score

[0077]

[0078] S63. Score for item-based collaborative filtering recommendation

[0079]

[0080] S64. Calculate the fusion recommendation score:

[0081]

[0082] Among them, s final Indicates sound effect settings S i The final recommendation score, ω1, ω2 and ω3 are weight coefficients, satisfying ω1+ω2+ω3=1;

[0083] S65, set all sound effects to S i According to the final recommendation score final Sorting is performed, and the sound effects with the highest scores are set as the final recommendation result set.

[0084] Furthermore, the S7 specifically includes:

[0085] S71. Collect users’ immediate feedback on the recommended sound effect settings through user interface design and interaction;

[0086] S72. Based on the real-time changes of user feedback data, the multi-branch neural network and collaborative filtering model are updated online, and the parameters of the multi-branch neural network are fine-tuned using incremental learning technology to ensure that the model can quickly adapt to changes in user preferences; then, the user-sound effect setting matrix and similarity matrix of the collaborative filtering model are updated, and the loss function in the online update process is:

[0087]

[0088] Among them, N3 represents the number of copies, represents the sound setting score predicted by the model, r i represents the actual feedback score of the user, α represents the weight balance parameter, T represents the user's historical time window, cos(f i,t ,f j,t ) represents the sound effect feature f i,t and f j,t Cosine similarity at time t, λ represents the regularization coefficient, Δθ represents the update amount of the model parameter, and the regularization term λ∥Δθ∥ 2 Used to avoid overfitting and ensure the generalization ability of the model;

[0089] S73. By analyzing the user's long-term behavior data and short-term feedback, monitor the changing trend of user preferences, and use multi-level sliding window technology to calculate the changes in user preferences in different time periods.

[0090]

[0091] Among them, Preference t represents the user’s preference state at time t, represents the user's preference change within the time window [T1, T2], Preference t+1 represents the user's preference state at time t+1, and Decay(t) represents the time decay function, which is used to enhance the focus on recent preferences;

[0092] S74. Regularly conduct A / B testing and comparative experiments to evaluate the performance improvement of the recommendation system. The test content includes recommendation accuracy, user satisfaction and system response speed indicators.

[0093] The beneficial effects of the present invention are:

[0094] First, the introduction of multi-branch neural networks enables the system to process user features and audio features simultaneously, achieving a deep understanding of user personalized preferences. Through one-dimensional convolutional neural networks and adaptive gated long short-term memory networks, we can effectively extract users' temporal behavior patterns and preference trends, while using multi-head self-attention mechanisms to capture the global correlation between audio features. This deep feature extraction capability ensures that the recommendation system can generate more accurate personalized recommendation results and improve the user's listening experience.

[0095] Secondly, the combination of user-based and item-based collaborative filtering algorithms fully utilizes the collective wisdom of users and the characteristics of audio content. By calculating the multi-dimensional similarity between users and between sound effect settings, we are able to recommend sound effect settings that are similar to the user's preferences, as well as sound effect settings for similar audio. This dual recommendation mechanism not only takes into account the preferences of individual users, but also draws on the commonalities of similar users and audio to provide more comprehensive recommendations. In addition, through the dynamic weight adjustment mechanism, we can dynamically adjust the weights of each recommendation method based on the user's immediate feedback to ensure the diversity and adaptability of the recommendation results. This flexible weight adjustment mechanism enables the system to respond to changes in user preferences in real time and maintain the efficiency and accuracy of recommendations.

[0096] In addition, the present invention also introduces an online learning mechanism, which continuously updates the multi-branch neural network and collaborative filtering model by collecting and processing real-time feedback information from users. The application of incremental learning technology enables the system to quickly adapt to the changing needs of users and avoids the risk of outdated models. By adjusting the model parameters online, we can ensure that the recommendation results always reflect the user's latest preferences and improve user satisfaction. At the same time, the introduction of sliding window technology and time decay function enables the system to keenly capture the changing trend of user preferences over time and make corresponding adjustments. This accurate monitoring and adaptation ability of user preferences further enhances the intelligence level of the recommendation system. BRIEF DESCRIPTION OF THE DRAWINGS

[0097] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0098] Figure 1 A flowchart of a method for adjusting personalized sound effects of a loudspeaker implemented by a collaborative filtering algorithm proposed by the present invention;

[0099] Figure 2 A schematic diagram of the structure of a multi-branch neural network model of a method for adjusting personalized sound effects of a loudspeaker realized by a collaborative filtering algorithm proposed in the present invention;

[0100] Figure 3 This is a schematic diagram of a fusion recommendation system of a multi-branch neural network and a collaborative filtering algorithm for a method of personalized speaker sound effect adjustment implemented by a collaborative filtering algorithm proposed in the present invention. DETAILED DESCRIPTION

[0101] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, which only illustrate the basic structure of the present invention in a schematic manner, and therefore only show the components related to the present invention.

[0102] refer to Figure 1-3 , a method for adjusting personalized sound effects of a loudspeaker by collaborative filtering algorithm, comprising the following steps:

[0103] S1. Collect user listening preference data and historical behavior data to build a user behavior dataset; collect audio feature data to build an audio feature dataset;

[0104] S2. Preprocessing the user behavior dataset and the audio feature dataset, including data cleaning, normalization, and feature extraction, to obtain user features and audio features;

[0105] S3. Build and train a multi-branch neural network model, including a user feature branch, an audio feature branch, a feature fusion layer, and an output layer, and generate personalized recommendations;

[0106] S4, using a user-based collaborative filtering algorithm, based on the historical data of user sound effect settings, calculate the similarity between users, and generate sound effect setting recommendations that are similar to the current user's sound effect preferences and that other users like;

[0107] S5. Using an item-based collaborative filtering algorithm, the similarity between audio items is calculated according to audio features, and a sound effect setting recommendation is generated for audio with similar features to the current audio;

[0108] S6. Integrate the personalized recommendation generated by the multi-branch neural network and the recommendation results of the collaborative filtering algorithm, dynamically adjust the recommendation weight, and generate a variety of recommended sound effect setting options;

[0109] S7. Collect user feedback on recommended sound effect settings, regularly update the multi-branch neural network and collaborative filtering modules, and optimize the model based on the latest user data and feedback.

[0110] In this implementation, S3 specifically includes:

[0111] S31, input user feature vector U = [u1, u2, ..., u n ], where u i Represents a single feature of a user. The temporal pattern in the user feature is extracted through a one-dimensional convolutional neural network. The convolution operation is defined as:

[0112] U c =LeakyReLU(W c *U+b c );

[0113] Among them, U c represents the feature representation after the convolution operation, W c represents the convolution kernel, b crepresents the bias vector, * represents the convolution operation, and LeakyReLU represents the activation function, which is used to better handle negative information;

[0114] S32, U c Long short-term memory network with adaptive gating mechanism, which adds adaptive gates to the standard LSTM to dynamically adjust the flow of information:

[0115] g t =σ(W g [h t-1 ,U c ]+b g );

[0116] f t =σ(W f [h t-1 ,U c ]+b f )⊙g t ;

[0117] i t =σ(W i [h t-1 ,U c ]+b i )⊙(1-g t );

[0118] c t =f t ⊙c t-1 +i t ⊙tanh(W c [h t-1 ,U c ]+b c );

[0119] o t =σ(W o [h t-1 ,U c ]+b o );

[0120] h t =o t ⊙tanh(c t );

[0121] Among them, g t Represents the adaptation gate vector, which is used to dynamically adjust the output of the forget gate and the input gate, which can improve the selective filtering of information, especially for changes in specific user characteristics. g represents the weight matrix of the adaptation gate, b g Represents the bias vector of the adaptation gate, which is used to dynamically adjust the output of other gates, f t represents the forget gate, it represents the input gate vector, c t represents the cell state vector, o t represents the input gate vector, h t represents the hidden state vector, W f Represents the weight matrix of the forget gate, W i represents the weight matrix of the input gate, W c The weight matrix representing the cell state, W o represents the weight matrix of the input gate, ⊙ represents the element multiplication operation, σ represents the activation function, and the implicit representation U′ of the user feature is obtained;

[0122] S33, input audio feature vector A = [a1, a2, ..., a m ], where a i Represents a single feature of the audio, and extracts the global correlation between audio features through a multi-head self-attention mechanism. For each head k:

[0123]

[0124] Among them, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k Represents the dimension of the key vector. After connecting the multiple results, the implicit representation A′ of the audio feature is obtained through a fully connected network:

[0125] A′=Concat(head1,head2,…,head h )W O ;

[0126] Among them, head k represents the output of the kth head, W O represents the output transformation matrix;

[0127] S34. Introduce the attention mechanism to calculate the importance weights of user features and audio features, and connect the outputs U′ and A′ of the user feature branch and the audio feature branch to form a fused feature vector F:

[0128] F = [U′, A′];

[0129] The importance weight α of each feature is calculated through the attention network i :

[0130]

[0131] Among them, f i represents the representation of feature i, and w represents the weight vector of the attention network;

[0132] The weighted fusion feature vector F′ is obtained by weighted summation:

[0133]

[0134] S35, input the weighted fusion feature vector F′ to the output layer, and generate personalized recommended sound effect setting parameters P through the fully connected network:

[0135] P=W o1 F′+b o1 ;

[0136] Among them, W o1 represents the weight matrix of the output layer, b o1 Represents the bias vector of the output layer;

[0137] S36. Introduce the user experience evaluation score E into the loss function, and combine the weighted influence of the user's historical behavior on the current sound effect preference to define the comprehensive loss function L:

[0138]

[0139] Where N is the number of samples, p i represents the sound effect setting parameters predicted by the model, represents the user's actual preference parameter, E represents the user experience evaluation score, M represents the number of user historical behaviors, and w j represents the weight of the jth historical behavior on the current preference, h j represents the jth historical behavior score of the user, λ1 and λ2 represent the importance coefficients for adjusting different loss terms:

[0140] Calculate the historical behavior score h j :

[0141] h j =σ(W h z j +b h );

[0142] Among them, z j The feature vector representing the jth historical behavior, W h represents the weight matrix, b h represents the bias vector;

[0143] S37. Generate personalized recommendation result R1 through the trained multi-branch neural network model.

[0144] In this implementation, S4 specifically includes:

[0145] S41. Based on the historical sound effect setting data of the user, construct a user-sound effect setting matrix M=[m ij ], where m ij Represents user P iSet S for sound effects j preference ratings;

[0146] S42, normalizing the user-sound effect setting matrix, normalizing the score of each user to a standard normal distribution with a mean of 0 and a standard deviation of 1;

[0147] S43, calculate the multi-dimensional similarity between users, the similarity calculation uses weighted cosine similarity, combined with the dynamic weight of the sound effect feature vector W(t) = [w1(t), w2(t), ..., w n (t)] and the time decay factor α(t):

[0148]

[0149] Among them, sim(P i ,P k ) represents user P i and user P k The similarity between ij Represents user P i Set S for sound effects j Rating, m kj Represents user P k Set S for sound effects j Rating, w j (t) indicates the sound effect setting S j The dynamic feature weight at time t, α(t) represents the time decay factor, and n represents the number of sound effect settings;

[0150] S44. For each target user P, select the k users with the highest similarity as the similar user set N(P):

[0151] N(P)={P k |khighestsim(P,P k )};

[0152] Among them, k represents the number of similar users;

[0153] S45: Generate a recommended sound effect setting for the target user based on the sound effect setting history of the similar user set N(P). j , calculate the weighted average rating in a set of similar users:

[0154]

[0155] in, S represents the recommendation for user P j The expected score of sim(P,P k ) represents user P and similar user P k The similarity betweenkj Represents similar users P k Set S for sound effects j Rating, w j (t) indicates the sound effect setting S j The dynamic feature weight at time t, β kj Represents similar users P k Set S for sound effects j The personalized weighting coefficient of

[0156] The formula introduces dynamic feature weights, time decay factors and personalized weighting coefficients, which can capture the dynamic characteristics of user preferences changing over time and reflect personalized weights, while improving the model's ability to capture user behavior and sound effect features, thereby providing more accurate and personalized recommendations.

[0157] S46. By sorting the expected scores of all sound effect settings, a recommendation R2 of sound effect settings preferred by other users who have similar sound effect preferences to the current user is generated.

[0158] In this implementation manner, S5 specifically includes:

[0159] S51, collect the multi-dimensional attributes of all sound effect settings, expressed as a sound effect setting feature vector G = [g1, g2, ..., g m ], where g i A single property representing a sound effect setting;

[0160] S52, normalizing the sound effect setting feature vector G, normalizing the value range of all attributes to [0, 1] to eliminate the influence of dimension;

[0161] S53, constructing a sound effect setting-attribute matrix B based on the sound effect setting attribute vector G = [b ij ], where b ij Indicates sound setting X i The value of the jth attribute is used to calculate the similarity between each pair of sound effect settings;

[0162] S54, for each target sound effect setting X, selecting l sound effect settings with the highest similarity as a similar sound effect setting set Y(X), where l represents the number of similar sound effect settings;

[0163] S55: Based on the similar sound effect setting set Y(X) and the user preference data, generate a recommended sound effect setting for the target sound effect setting, and i Calculates a weighted score among a set of similar sound settings

[0164] S56. Generate a sound effect setting recommendation R3 for audio with similar features to the current audio by sorting the expected scores of all sound effect settings.

[0165] In this implementation manner, S6 specifically includes:

[0166] S61, calculating the score s of the recommended sound effect settings in the sets R1, R2 and R3 i , for the recommendation score of multi-branch neural network

[0167]

[0168] Among them, N1 represents the set of all users, sim(P,P k ) represents the target user P and user P k The similarity between k,i Represents user P k Set S for sound effects i The actual rating of

[0169] S62. User-based collaborative filtering recommendation score

[0170]

[0171] S63. Score for item-based collaborative filtering recommendation

[0172]

[0173] S64. Calculate the fusion recommendation score:

[0174]

[0175] Among them, s final Indicates sound effect settings S i The final recommendation score, ω1, ω2 and ω3 are weight coefficients, satisfying ω1+ω2+ω3=1;

[0176] S65, set all sound effects to S i According to the final recommendation score final Sorting is performed, and the sound effects with the highest scores are set as the final recommendation result set.

[0177] In this implementation manner, the S7 specifically includes:

[0178] S71. Collect users’ immediate feedback on the recommended sound effect settings through user interface design and interaction;

[0179] S72. Based on the real-time changes of user feedback data, the multi-branch neural network and collaborative filtering model are updated online, and the parameters of the multi-branch neural network are fine-tuned using incremental learning technology to ensure that the model can quickly adapt to changes in user preferences; then, the user-sound effect setting matrix and similarity matrix of the collaborative filtering model are updated, and the loss function in the online update process is:

[0180]

[0181] Among them, N3 represents the number of copies, represents the sound setting score predicted by the model, r i represents the actual feedback score of the user, α represents the weight balance parameter, T represents the user's historical time window, cos(f i,t ,f j,t ) represents the sound effect feature f i,t and f j,t Cosine similarity at time t, λ represents the regularization coefficient, Δθ represents the update amount of the model parameter, and the regularization term λ∥Δθ∥ 2 Used to avoid overfitting and ensure the generalization ability of the model;

[0182] S73. By analyzing the user's long-term behavior data and short-term feedback, monitor the changing trend of user preferences, and use multi-level sliding window technology to calculate the changes in user preferences in different time periods.

[0183]

[0184] Among them, Preference t represents the user’s preference state at time t, represents the user's preference change within the time window [T1, T2], Preference t+1 represents the user's preference state at time t+1, and Decay(t) represents the time decay function, which is used to enhance the focus on recent preferences;

[0185] S74. Regularly conduct A / B testing and comparative experiments to evaluate the performance improvement of the recommendation system. The test content includes recommendation accuracy, user satisfaction and system response speed indicators.

[0186] Embodiment 1:

[0187] In order to verify the feasibility of the present invention in implementation, the present invention is applied to a typical home scenario. Users often use smart speakers to play various types of audio content, including music, podcasts, movie dubbing, etc. Since family members have different preferences for sound effects and listening habits, the traditional manual sound effect adjustment method is difficult to meet everyone's needs. In addition, the user's listening environment and emotional state will also affect the sound effect preference. Therefore, there is a need for an intelligent system that can dynamically adjust and optimize the sound effect settings to improve the user's overall listening experience.

[0188] In this embodiment, user A and user B are members of the same family, and their listening preferences are quite different. User A likes sound effects with clear treble and less bass, while user B prefers the effect of enhanced bass. In order to meet the needs of the two users, the system first collects the user's listening preference data and historical behavior data through the user's smartphone app and the microphone on the audio device. The system records the audio playback records, volume adjustments, equalizer setting adjustments, and other information of user A and user B in the past month. These data are used to construct a user behavior data set. At the same time, the system also collects audio feature data such as spectrum, rhythm, and pitch in real time through the built-in microphone of the audio device for constructing an audio feature data set.

[0189] In order to achieve personalized recommendations, the system first preprocesses the collected user behavior data and audio feature data. The data cleaning step removes outliers and invalid data, and the normalization process standardizes the value range of all data to [0,1] to ensure data consistency. The feature extraction step extracts the user's personalized feature vector from the user behavior data, such as volume preference, frequently listened music type, etc.; and extracts audio feature vectors from the audio feature data, such as spectrum features and pitch features.

[0190] After data preprocessing, the system uses a multi-branch neural network model for personalized recommendations. User features and audio features are input into the user feature branch and audio feature branch respectively, and features are extracted through a one-dimensional convolutional neural network and a multi-head self-attention mechanism. Then, the user features and audio features are fused in the feature fusion layer, and personalized sound effect setting recommendations are generated through the output layer. These recommendations include parameter settings such as equalizer gain adjustment and bass enhancement. In this embodiment, the system recommends settings for increasing high-frequency gain and reducing low-frequency gain for user A, and settings for enhancing bass for user B.

[0191] To further optimize the recommendation effect, the system combines user-based collaborative filtering algorithms and item-based collaborative filtering algorithms. By calculating the similarity between users, the system can find other users with similar sound effect preferences as the current user and recommend sound effect settings that these users like. At the same time, the system also calculates the similarity between audio items and recommends sound effect settings for audio with similar features to the current audio. Through the dynamic weight adjustment mechanism, the system dynamically adjusts the recommendation weight based on real-time feedback from users to ensure the diversification and personalization of the final recommendation results.

[0192] In actual applications, the system asks users about their satisfaction with the current sound effect settings after each audio playback, and allows users to provide adjustment suggestions. Through these instant feedback, the system can quickly update the multi-branch neural network and collaborative filtering model. User A expressed in his feedback that he was very satisfied with the high-frequency enhancement effect, while user B was very satisfied with the bass enhancement. By analyzing the user's feedback data and behavior data, the system updates the model parameters online and optimizes the recommendation algorithm. During the three consecutive months of use, the system was able to continuously adapt to changes in user preferences and provide sound effect settings that meet the user's real-time needs.

[0193] Table 1 User sound effect settings and system performance data of home smart audio system

[0194]

[0195] In the above table, we show the key data and performance indicators of the sound effect recommendation system in actual application. The system provides highly personalized sound effect setting recommendations for each user by analyzing the historical behavior data and instant feedback of user A and user B. User A's preferences are mainly focused on treble clarity and bass reduction, while user B prefers enhanced bass effects. The system is able to accurately capture the difference between these two preferences and provide corresponding setting recommendations, such as recommending a high-frequency gain of +3dB and a low-frequency reduction of -2dB for user A, and a low-frequency gain of +5dB for user B. This shows that the system performs well in handling diverse user needs.

[0196] In terms of volume adjustment, user A usually keeps the volume at a medium level and makes small adjustments occasionally, while user B prefers a higher volume and makes frequent adjustments. The system's response time is always kept below 50 milliseconds, ensuring that the user's volume adjustment needs can be met quickly. At the same time, the system can handle multiple equalizer setting responses, providing users with different sound effect preset options, and improving the diversity of user experience.

[0197] User satisfaction and feedback data further verified the effectiveness of the system. User A expressed 85% satisfaction with the effect of high-frequency enhancement, and user B expressed 90% satisfaction with bass enhancement. Overall, the system's recommendation matching degree increased by 20% and user satisfaction increased by 15%, showing the system's significant advantages in personalized recommendations and improving user experience. In addition, the system is able to update multi-branch neural networks and collaborative filtering models in real time, and adjust model parameters immediately after receiving user feedback. This online learning mechanism ensures that the system can continue to adapt to changes in user preferences.

[0198] The system also collects a large amount of user feedback data, a total of 1,000 users' feedback, covering multi-dimensional information such as satisfaction with sound settings and specific adjustment suggestions. These data not only help the system optimize the recommendation algorithm, but also provide valuable reference for future improvements. In particular, through the sliding window technology, the system can monitor the long-term and short-term preference changes of users and dynamically adjust the recommendation strategy to better match the immediate needs of users.

[0199] In terms of system performance, the table data shows that the system's average response time remains below 100 milliseconds, and it can respond to various user adjustment needs in real time. This efficient response capability, coupled with a dynamic weight adjustment mechanism and a variety of sound effect setting options, enables the system to provide a high level of personalized recommendation services, bringing users a richer and more satisfying listening experience. In general, the data in the table fully demonstrates the performance of the sound effect recommendation system in actual applications, and its advantages of high accuracy, dynamic adaptability, and improved user satisfaction have been fully verified.

[0200] The above description is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical scheme and inventive concept of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A method for adjusting personalized sound effects of a loudspeaker using a collaborative filtering algorithm, characterized in that: The steps include: S1. Collect user listening preference data and historical behavior data to build a user behavior dataset; Collect audio feature data and build an audio feature dataset; S2. Preprocess the user behavior dataset and audio feature dataset, including data cleaning, normalization, and feature extraction, to obtain user features and audio features; S3. Build and train a multi-branch neural network model, including a user feature branch, an audio feature branch, a feature fusion layer, and an output layer, and generate personalized recommendations. S4. Using a user-based collaborative filtering algorithm, the similarity between users is calculated based on the historical data of user sound effect settings, and recommendations for sound effect settings that are similar to the current user's sound effect preferences are generated; S5. Use an item-based collaborative filtering algorithm to calculate the similarity between audio items based on audio features, and generate sound effect setting recommendations for audio with similar features to the current audio; S6. Integrate the personalized recommendations generated by the multi-branch neural network and the recommendation results of the collaborative filtering algorithm, dynamically adjust the recommendation weights, and generate a variety of recommended sound effect setting options; S7. Collect user feedback on recommended sound effect settings, regularly update the multi-branch neural network and collaborative filtering modules, and optimize the model based on the latest user data and feedback.

2. The method for adjusting speaker personalized sound effects by collaborative filtering algorithm according to claim 1, characterized in that: The S3 specifically includes: S31. Input user feature vector U=[u1,u2,…,u n ], where u i Represents a single feature of a user, and extracts the temporal pattern in the user feature through a one-dimensional convolutional neural network. The convolution operation is defined as: AT c =LeakyReLU(W c *U+b c ); Among them, U c Represents the feature representation after convolution operation, W c represents the convolution kernel, b c Represents the bias vector, * represents the convolution operation, and LeakyReLU represents the activation function; S32, U c Long Short-Term Memory Networks via Adaptive Gating Mechanism: g t =σ(W g [h t-1 ,U c ]+b g ); f t =σ(W f [h t-1 ,U c ]+b f )⊙g t ; i t =σ(W i [h t-1 ,U c ]+b i )⊙(1-g t ); c t =f t ⊙c t-1 +i t ⊙tanh(W c [h t-1 ,U c ]+b c ); the t =σ(W o [h t-1 ,U c ]+b o ); h t =o t ⊙tanh(c t ); Among them, g t Represents the adaptation gate vector, W g represents the weight matrix of the adaptation gate, b g represents the bias vector of the adaptation gate, f t represents the forget gate, i t represents the input gate vector, c t represents the cell state vector, o t represents the input gate vector, h t represents the hidden state vector, W f Represents the weight matrix of the forget gate, W i Represents the weight matrix of the input gate, W c The weight matrix representing the cell state, W o Represents the weight matrix of the input gate, ⊙ represents the element multiplication operation, σ represents the activation function, and the implicit representation U of the user feature is obtained ′ ; S33, input audio feature vector A=[a1,a2,…,a m ], where a i Represents a single feature of the audio, and extracts the global correlation between audio features through a multi-head self-attention mechanism. For each head k: Among them, Q represents the query matrix, K represents the key matrix, V represents the value matrix, and d k Represents the dimension of the key vector. After connecting the multi-head results, the implicit representation A of the audio feature is obtained through the fully connected network. ′ : A ′ =Concat(head1,head2,…,head h )W O ; Among them, head k represents the output of the kth head, W O represents the output transformation matrix; S34, introduce the attention mechanism to calculate the importance weights of user features and audio features, and convert the output U of the user feature branch and the audio feature branch into ′ and A ′ Connect to form a fused feature vector F: F=[U ′ ,A ′ ]; Calculate the importance weight α of each feature through the attention network i : Among them, f i represents the representation of feature i, and w represents the weight vector of the attention network; The weighted fusion feature vector F is obtained by weighted summation ′ : S35, weighted fusion feature vector F ′ Input to the output layer, and generate personalized recommended sound effect setting parameters P through the fully connected network: P=W o1 F ′ +b o1 ; Among them, W o1 represents the weight matrix of the output layer, b o1 Represents the bias vector of the output layer; S36. Introduce the user experience evaluation score E into the loss function, and combine the weighted influence of the user's historical behavior on the current sound effect preference to define the comprehensive loss function L: Among them, N represents the number of samples, p i represents the sound effect setting parameters predicted by the model, represents the user's actual preference parameter, E represents the user experience evaluation score, M represents the number of user historical behaviors, and w j represents the influence weight of the jth historical behavior on the current preference, h j represents the user’s j-th historical behavior score, and λ1 and λ2 represent the importance coefficients for adjusting different loss terms: Calculate the historical behavior score h j : h j =σ(W h z j +b h ); Among them, z j The feature vector representing the jth historical behavior, W h represents the weight matrix, b h represents the bias vector; S37. Generate personalized recommendation result R1 through the trained multi-branch neural network model.

3. The method for adjusting speaker personalized sound effects by collaborative filtering algorithm according to claim 1, characterized in that: The S4 specifically includes: S41. Based on the user's historical sound effect setting data, construct a user-sound effect setting matrix M=[m ij ], where m ij Represents user P i Set S for sound effects j preference ratings; S42, normalizing the user-sound effect setting matrix to normalize each user's rating to a standard normal distribution with a mean of 0 and a standard deviation of 1; S43. Calculate the multi-dimensional similarity between users. The similarity calculation uses weighted cosine similarity, combined with the dynamic weight of the sound effect feature vector W(t) = [w1(t), w2(t), ..., w n (t)] and the time decay factor α(t): Among them, sim(P i ,P k ) represents user P i and user P k The similarity between ij Represents user P i Set S for sound effects j Rating, m kj Represents user P k Set S for sound effects j Rating, w j (t) indicates the sound effect setting S j The dynamic feature weight at time t, α(t) represents the time decay factor, and n represents the number of sound effect settings; S44. For each target user P, select the k users with the highest similarity as the similar user set N(P): N(P)={P k ∣khighestsim(P,P k )}; Where k represents the number of similar users; S45, based on the sound effect setting history of the similar user set N(P), generate the recommended sound effect setting for the target user. For each sound effect setting S j , calculate the weighted average rating in a set of similar users: in, S is recommended to user P j The expected score of sim(P,P k ) represents user P and similar user P k The similarity between kj Represents similar users P k Set S for sound effects j Rating, w j (t) indicates the sound effect setting S j The dynamic feature weight at time t, β kj Represents similar users P k Set S for sound effects j Personalized weighting coefficient of S46. Generate a recommendation R2 of sound effect settings that are similar to the current user's sound effect preferences and are preferred by other users by sorting the expected scores of all sound effect settings.

4. The method for adjusting speaker personalized sound effects by collaborative filtering algorithm according to claim 1, characterized in that: The S5 specifically includes: S51. Collect the multi-dimensional attributes of all sound effect settings, and express them as the sound effect setting feature vector G = [g1, g2, ..., g m ], where g i A single property representing a sound effect setting; S52, normalize the sound effect setting feature vector G, and normalize the value range of all attributes to [0, 1]; S53, constructing a sound effect setting attribute matrix B based on the sound effect setting attribute vector G = [b ij ], where b ij Indicates sound effect setting X i The value of the jth attribute is used to calculate the similarity between each pair of sound effect settings; S54, for each target sound effect setting X, selecting the l sound effect settings with the highest similarity as a similar sound effect setting set Y(X), where l represents the number of similar sound effect settings; S55: Based on the similar sound effect setting set Y(X) and the user preference data, generate a recommended sound effect setting for the target sound effect setting, and i Calculates a weighted score among a set of similar sound settings S56. Generate a sound effect setting recommendation R3 for audio with similar characteristics to the current audio by sorting the expected scores of all sound effect settings.

5. The method for adjusting speaker personalized sound effects by collaborative filtering algorithm according to claim 1, characterized in that: The S6 specifically includes: S61. Calculate the score s of the recommended sound effect settings in the sets R1, R2, and R3. i , for the recommendation score of multi-branch neural network Among them, N1 represents the set of all users, sim(P,P k ) represents the target user P and user P k The similarity between k,i Represents user P k Set S for sound effects i The actual rating of S62. User-based collaborative filtering recommendation score S63. For item-based collaborative filtering recommendation scores S64. Calculate the fusion recommendation score: Among them, s final Indicates sound effect setting S i The final recommendation score, ω1, ω2 and ω3 are weight coefficients, satisfying ω1+ω2+ω3=1; S65, set all sound effects to S i According to the final recommendation score final Sorting is performed, and the sound effects with the highest scores are set as the final recommendation result set.

6. The method for adjusting personalized speaker sound effects by collaborative filtering algorithm according to claim 1, characterized in that: The S7 specifically includes: S71. Collect users’ immediate feedback on recommended sound effect settings through user interface design and interaction; S72. Based on the real-time changes in user feedback data, update the multi-branch neural network and collaborative filtering model online, use incremental learning technology to fine-tune the parameters of the multi-branch neural network, update the user-sound effect setting matrix and similarity matrix of the collaborative filtering model, and the loss function during the online update process. for: Among them, N3 represents the number of represents the sound setting score predicted by the model, r i represents the user's actual feedback score, α represents the weight balance parameter, T represents the user's historical time window, cos(f i,t ,f j,t ) represents the sound effect feature f i,t and f j,t The cosine similarity at time t, λ represents the regularization coefficient, and Δθ represents the update amount of the model parameter; S73. By analyzing the long-term behavior data and short-term feedback of users, we can monitor the changing trend of user preferences and use multi-level sliding window technology to calculate the changes in user preferences in different time periods. Among them, Preference t represents the user’s preference state at time t, Indicates the user's preference changes within the time window [T1, T2]. t+1 represents the user's preference state at time t+1, and Decay(t) represents the time decay function; S74. Regularly conduct A / B testing and comparative experiments to evaluate the performance improvement of the recommendation system. The test content includes recommendation accuracy, user satisfaction, and system response speed indicators.

Images