A deep learning-based vehicle-mounted sound effect intelligent matching system

Abstract

This invention discloses a deep learning-based intelligent matching system for in-vehicle audio sound effects, comprising: a data acquisition module for acquiring playback audio data, user operation record data, and in-vehicle status data; a feature parsing module for generating audio feature vectors, preference feature vectors, and environmental feature vectors; a joint encoding module for forming joint feature vectors; a cross-order spline neural network module for generating cross-order spline results; a sparse activation module for generating sparse activation spline results; a parameter generation module for generating frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters; and an output control module for completing sound effect control. This invention achieves multi-feature fusion and cross-order spline matching, resulting in more accurate parameter generation.

Description

Technical Field

[0001] This invention relates to the field of intelligent audio control, and more particularly to an intelligent matching system for in-vehicle audio effects based on deep learning. Background Technology

[0002] As in-vehicle cabins continue to evolve towards intelligence and personalization, in-vehicle audio systems are no longer limited to fixed equalizer adjustments, preset sound field switching, and simple volume control. In actual use, the types of content played vary significantly. Voice navigation, pop music, classical music, movie audio, and radio programs all have different requirements in terms of spectrum distribution, dynamic range, and spatial performance. Furthermore, different drivers have different listening preferences, and vehicle driving conditions, speed changes, window opening / closing status, air conditioning operation, and cabin noise environment all directly affect the actual listening experience. Most existing in-vehicle audio systems use fixed rules, manual preset modes, or a few parameter lookup tables to adjust sound effects. These adjustments are relatively rigid and difficult to adapt synchronously to complex scenarios.

[0003] In existing technologies, one type of solution relies on manually setting multiple sound effect templates for users to switch between manually; another type performs local compensation based on a single environmental variable, such as adjusting gain only for noise or vehicle speed. While these methods can improve the listening experience in some scenarios, they still suffer from problems such as limited adaptability, insufficient parameter linkage, and inadequate utilization of individual preferences. When the playback content changes rapidly, user operating habits differ significantly, or the in-car environment fluctuates continuously, fixed-rule solutions often fail to provide more detailed frequency band adjustment and sound field distribution results, easily leading to abrupt sound effect switching, lag in adjustment, or insufficient matching accuracy.

[0004] Therefore, there is a need for an intelligent matching system for in-vehicle audio effects that can simultaneously combine the characteristics of the played audio, user preference information, and the in-vehicle environment to intelligently generate and dynamically output sound effect parameters. Summary of the Invention

[0005] One objective of this invention is to propose a deep learning-based intelligent matching system for in-vehicle audio sound effects. This invention fully integrates feature parsing mechanisms, joint encoding structures, multi-order spline neural networks, and sparse activation strategies, and details the entire process of in-vehicle audio sound effect matching from data acquisition, feature extraction, joint representation, spline mapping, control point selection to parameter generation and output control. It possesses advantages such as accurate parameter matching, coordinated sound field distribution, fast response speed, and adaptability to changing in-vehicle environments. According to an embodiment of this invention, a deep learning-based intelligent matching system for in-vehicle audio sound effects includes: The data acquisition module collects audio playback data, user operation records, and vehicle status data to form an input dataset. The feature parsing module parses the input dataset to generate audio feature vectors, preference feature vectors, and environmental feature vectors. The joint encoding module encodes these vectors to form a joint feature vector. The cross-order spline neural network module constructs a node sequence based on the joint feature vector, establishes sets of B-spline basis functions of different orders at corresponding positions in the node sequence, generates order activation labels and interval activation labels, selects the spline order for the current calculation based on the order activation label, selects the spline interval for the current calculation based on the interval activation label, and performs a combination calculation on the selected spline basis function set to generate the cross-order spline result. The sparse activation module is used to establish a set of control points at the corresponding positions of the cross-order spline results, select control points to participate in the current calculation according to the order activation identifier and the interval activation identifier, and perform sparse combination calculation on the selected control points to generate sparse activation spline results; the parameter generation module is used to perform mapping on the sparse activation spline results to generate frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters; the output control module is used to write the frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters into the vehicle audio control unit to complete the sound effect control.

[0006] Optionally, the feature parsing module's function of parsing the input data includes the following steps: Read the audio sampling sequence, channel identifier and playback type identifier from the audio data being played. Perform frame segmentation, windowing and spectrum transformation on the audio sampling sequence. Extract the frequency band energy value, spectral centroid value, dynamic range value and rhythm intensity value corresponding to each frame to form an audio feature vector. Read the volume adjustment record, equalizer adjustment record, sound field mode switching record and playback switching record from the user operation record data, and extract the adjustment direction value, adjustment amplitude value, dwell time value and mode selection frequency value in time order to form a preference feature vector; Read the operating status data, cabin environment data and occupant status data from the vehicle status data, perform numerical conversion and identification mapping on each status data, and extract the environmental noise component, spatial enclosure component, operating condition component and occupant distribution component to form an environmental feature vector. The audio feature vector, preference feature vector, and environment feature vector are processed to achieve uniform length and position alignment, resulting in a parsing result corresponding to the same playback time.

[0007] Optionally, the generation order activation identifier and interval activation identifier of the cross-order spline neural network module specifically include: The feature components in the joint feature vector are written into the node construction process according to the preset input order. The numerical value of each feature component, the difference between adjacent components and the local distribution position are read to form a node position sequence. Based on the node position sequence, the nodes are divided into several continuous node intervals. The interval length, the value of the change within the interval, and the value of the difference between the interval boundaries are counted for each node interval to form an interval analysis sequence. Establish first-order B-spline basis functions, second-order B-spline basis functions and third-order B-spline basis functions at corresponding positions in each node interval, and read the basis function response value, interval coverage value and boundary fitting value of each order B-spline basis function in the current node interval to form an order determination sequence. Compare the response results of each order in the order determination sequence, extract the target order number at the corresponding position, and form an order activation identifier; Perform interval comparison on the basis function response results corresponding to each node interval, extract the target interval number at the corresponding position, and form the interval activation identifier; The order activation flag and the interval activation flag are written in the order of node position to form the spline activation flag sequence corresponding to the current joint feature vector.

[0008] Optionally, the generation of cross-order spline results by the cross-order spline neural network module specifically includes: Read the target order number corresponding to each node position in the order activation identifier, and extract the target order basis function sequence that is consistent with each target order number from the first-order B-spline basis function set, the second-order B-spline basis function set, and the third-order B-spline basis function set. Read the target interval number corresponding to each node position in the interval activation identifier, extract the target interval basis function subset that is consistent with each target interval number from each target order basis function sequence, and form an interval spline calculation sequence; Each feature component in the joint feature vector is written into the corresponding position of the interval spline calculation sequence according to the node position order. The input value, node spacing value and interval span value corresponding to each feature component are read. The basis function response value corresponding to each feature component in the basis function subset of the target interval is calculated to form the basis function response sequence. Read the basis function response values ​​and interval action values ​​corresponding to the basis function subsets of each target interval, perform product calculation and accumulation calculation on the response values ​​corresponding to each position in the basis function response sequence, and form the interval spline output values ​​corresponding to each node position; The interval spline output values ​​corresponding to each node position are grouped according to the target order number to form a first-order spline output sequence, a second-order spline output sequence and a third-order spline output sequence. Read the output values ​​corresponding to each position in the first-order spline output sequence, the second-order spline output sequence, and the third-order spline output sequence, and perform alignment according to the node position order to form a cross-order output alignment sequence; Perform sequential superposition calculations on the output values ​​at each position in the cross-order output alignment sequence to form a cross-order spline intermediate sequence; Read the node continuity values, interval connection values, and boundary change values ​​corresponding to the intermediate sequence of the cross-order spline, perform continuous position merging calculation on the intermediate sequence of the cross-order spline, and generate the cross-order spline result.

[0009] Optionally, the sparse activation module establishes a set of control points at the corresponding positions of the cross-order spline results, and filters the control points participating in the current calculation according to the order activation identifier and the interval activation identifier, specifically including: Read the spline output values ​​corresponding to each node position in the cross-order spline results, and extract the node position numbers corresponding to each spline output value; Read the order activation flag and interval activation flag corresponding to each node position, and extract the target order number and target interval number corresponding to each node position; Concatenate the node position number, target order number and target interval number in sequence to generate the control point number corresponding to each node position. The same node position corresponds to one control point number, and different node positions correspond to different control point numbers. Write the control point number into the control point registration sequence, and associate and record the spline output value, target order number, target interval number and node position number corresponding to each control point number to form an initial control point set; The initial control point set is classified according to the target order number, forming a first-order control point subset, a second-order control point subset, and a third-order control point subset; Control points for the corresponding intervals are extracted from the first-order control point subset, the second-order control point subset, and the third-order control point subset according to the target interval number, forming an interval-selected control point sequence; Perform a number check on each control point in the interval screening control point sequence, retain control points whose node position number, target order number and target interval number are consistent, and delete control points whose three are inconsistent, thus forming a valid control point sequence. The effective control point sequence is arranged in order of node position to form a set of control points participating in the current calculation.

[0010] Optionally, the step of performing sparse combination calculations on the selected control points to generate sparse activation splines specifically includes: Read the control point number, node position number, spline output value and interval assignment number corresponding to each control point in the effective control point sequence, and form the control point input sequence according to the node position order; Based on the difference between the node position interval value and the spline output value of each control point in the control point input sequence, the local action value of each control point is extracted. The local action values ​​corresponding to each control point are accumulated according to the node position to form a node local combination sequence; Read the target order number and target interval number corresponding to each node position in the local combination sequence of nodes. Retain the local combination values ​​of nodes with the same target order number and target interval number. Delete the local combination values ​​of nodes with different target order numbers or target interval numbers to form a sparse retention sequence. The retained values ​​corresponding to each node position in the sparse retained sequence are merged according to the node position order to form a sparse combined sequence. Perform node continuity checks and interval consistency checks on sparse combination sequences, and output sparse activation splines.

[0011] Optionally, the generated frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters specifically include: The sparse activation spline results are written into the parameter mapping process, and the frequency band response values, spatial distribution values ​​and dynamic change values ​​corresponding to each position are extracted to form frequency band adjustment parameters, sound field distribution parameters and dynamic compensation parameters, respectively. The frequency band adjustment parameters, sound field distribution parameters and dynamic compensation parameters are then verified for range and their positions are organized to form a parameter result sequence that can be written into the vehicle audio control unit.

[0012] Optionally, writing the frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters into the vehicle audio control unit specifically includes: Read the frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters, and write them into the corresponding positions of the equalization control unit, sound field control unit, and compensation control unit according to the parameter category; perform position matching checks and numerical boundary checks on each parameter after writing to form a control parameter sequence; send the control parameter sequence to the vehicle audio control unit to complete speaker output adjustment, sound field position adjustment, and dynamic compensation control.

[0013] The beneficial effects of this invention are: (1) By uniformly analyzing and jointly encoding the playback audio data, user operation record data and vehicle status data, the audio features, preference features and environmental features are made to correspond to each other at the same time and location, avoiding adjustment deviations caused by a single data source and improving the matching accuracy of the generated sound effect parameters.

[0014] (2) By constructing a cross-order spline neural network module, a set of B-spline basis functions of different orders is introduced at the same node position. The spline is selected by combining the order activation label and the interval activation label, so that feature changes can be distinguished and processed under different order expression methods, thereby improving the adaptability to complex audio structures and environmental changes.

[0015] (3) By establishing a set of control points based on the results of cross-order splines, and performing control point screening based on the order activation identifier and the interval activation identifier, an effective control point sequence is formed. Then, sparse combination calculation is performed to reduce the participation of irrelevant control points in the calculation, making the spline output process more stable and reducing computational redundancy.

[0016] (4) By mapping the sparse activation spline results to frequency band adjustment parameters, sound field distribution parameters and dynamic compensation parameters, and writing them into the vehicle audio control unit, multi-parameter coordinated adjustment is achieved, so that the corresponding sound effect control results can be output under different playback content, user operation status and vehicle environment conditions, thereby improving the overall sound effect consistency and response efficiency. Attached Figure Description

[0017] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of a method for a deep learning-based intelligent matching system for in-vehicle audio sound effects proposed in this invention. Figure 2 This is a schematic diagram illustrating the implementation process of a cross-order spline neural network module in a deep learning-based intelligent matching system for in-vehicle audio effects proposed in this invention.

[0018] Figure 3 This is a schematic diagram illustrating the sparse activation and parameter generation process of a deep learning-based intelligent matching system for in-vehicle audio sound effects proposed in this invention. Detailed Implementation

[0019] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0020] refer to Figures 1-3 A deep learning-based intelligent matching system for in-vehicle audio sound effects includes: The data acquisition module collects audio playback data, user operation records, and vehicle status data to form an input dataset. The feature parsing module parses the input dataset to generate audio feature vectors, preference feature vectors, and environmental feature vectors. The joint encoding module encodes these vectors to form a joint feature vector. The cross-order spline neural network module constructs a node sequence based on the joint feature vector, establishes sets of B-spline basis functions of different orders at corresponding positions in the node sequence, generates order activation labels and interval activation labels, selects the spline order for the current calculation based on the order activation label, selects the spline interval for the current calculation based on the interval activation label, and performs a combination calculation on the selected spline basis function set to generate the cross-order spline result. The sparse activation module is used to establish a set of control points at the corresponding positions of the cross-order spline results, select control points to participate in the current calculation according to the order activation identifier and the interval activation identifier, and perform sparse combination calculation on the selected control points to generate sparse activation spline results; the parameter generation module is used to perform mapping on the sparse activation spline results to generate frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters; the output control module is used to write the frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters into the vehicle audio control unit to complete the sound effect control.

[0021] In this embodiment, the feature parsing module performs the following steps to parse the input data: Read the audio sampling sequence, channel identifier and playback type identifier from the audio data being played. Perform frame segmentation, windowing and spectrum transformation on the audio sampling sequence. Extract the frequency band energy value, spectral centroid value, dynamic range value and rhythm intensity value corresponding to each frame to form an audio feature vector. Read the volume adjustment record, equalizer adjustment record, sound field mode switching record and playback switching record from the user operation record data, and extract the adjustment direction value, adjustment amplitude value, dwell time value and mode selection frequency value in time order to form a preference feature vector; Read the operating status data, cabin environment data and occupant status data from the vehicle status data, perform numerical conversion and identification mapping on each status data, and extract the environmental noise component, spatial enclosure component, operating condition component and occupant distribution component to form an environmental feature vector. The audio feature vector, preference feature vector, and environment feature vector are processed to achieve uniform length and position alignment, resulting in a parsing result corresponding to the same playback time.

[0022] Specifically, the framing, windowing, and spectral transformation processes performed on the audio sampling sequence include: Read continuous audio sampling points according to a preset sampling duration, and sequentially truncate the continuous audio sampling points according to a fixed frame length to form several audio frame sequences; read the frame shift values ​​corresponding to adjacent audio frames, and perform sliding truncation on each audio frame sequence to form a framed result sequence containing overlapping sampling segments; perform a length check on each audio frame sequence, and pad the end of audio frames with fewer sampling points than the preset frame length with zero values ​​to form a fixed-length audio frame sequence; generate a weighted value sequence of corresponding length according to the number of sampling points in each fixed-length audio frame sequence, and multiply each sampling point value with the corresponding weighted value point by point to form a windowed audio frame sequence; perform spectral transformation calculation on each windowed audio frame sequence to obtain the spectral result sequence corresponding to each audio frame; read the real and imaginary values ​​corresponding to each frequency position in each spectral result sequence, and perform summation of squares and square root calculations to form the spectral amplitude sequence corresponding to each audio frame; divide each spectral amplitude sequence into intervals according to frequency position to form the frequency band spectral distribution result corresponding to each audio frame.

[0023] Specifically, the frequency band energy value, spectral centroid value, dynamic range value, and rhythm intensity value include: The spectral amplitude sequence corresponding to each audio frame is segmented according to a preset frequency interval to form several frequency band intervals. The spectral amplitude values ​​corresponding to each frequency position within each frequency band interval are summed to form the frequency band energy value. The frequency value and spectral amplitude value corresponding to each frequency position are read, and a multiplicative summation is performed on the frequency value and spectral amplitude value. The summation of the spectral amplitude values ​​corresponding to each frequency position is then divided by the summation of the spectral amplitude values ​​to form the spectral centroid value. The absolute value sequence of time-domain sampling points corresponding to each audio frame is read, the maximum and minimum absolute values ​​are extracted, and the difference between the maximum and minimum absolute values ​​is calculated to form the dynamic range value. The frequency band energy values ​​corresponding to adjacent audio frames are read, and inter-frame difference calculations are performed to form an energy change value sequence. Local peak extraction is performed on the energy change value sequence in chronological order, and the number of peak occurrences within a unit of time is counted to form the rhythm intensity value. The frequency band energy value, spectral centroid value, dynamic range value, and rhythm intensity value corresponding to each audio frame are written in the order of frame position to form an audio feature vector.

[0024] In this embodiment, the generation order activation identifier and interval activation identifier of the cross-order spline neural network module specifically include: The feature components in the joint feature vector are written into the node construction process according to the preset input order. The numerical value of each feature component, the difference between adjacent components and the local distribution position are read to form a node position sequence. Based on the node position sequence, the nodes are divided into several continuous node intervals. The interval length, the value of the change within the interval, and the value of the difference between the interval boundaries are counted for each node interval to form an interval analysis sequence. Establish first-order B-spline basis functions, second-order B-spline basis functions and third-order B-spline basis functions at corresponding positions in each node interval, and read the basis function response value, interval coverage value and boundary fitting value of each order B-spline basis function in the current node interval to form an order determination sequence. Compare the response results of each order in the order determination sequence, extract the target order number at the corresponding position, and form an order activation identifier; Perform interval comparison on the basis function response results corresponding to each node interval, extract the target interval number at the corresponding position, and form the interval activation identifier; The order activation flag and the interval activation flag are written in the order of node position to form the spline activation flag sequence corresponding to the current joint feature vector.

[0025] Specifically, the first-order B-spline basis functions, second-order B-spline basis functions, and third-order B-spline basis functions include: Read the values ​​of the starting and ending nodes of each interval to determine the effective range of each basis function. In constructing the first-order B-spline basis functions, establish a linear basis function within a single interval, ensuring continuous variation of the basis function values ​​between the start and end positions, forming a sequence of first-order basis function values. In constructing the second-order B-spline basis functions, establish a quadratic basis function within the effective range formed by two adjacent intervals, ensuring continuous variation of the basis function values ​​within the effective range, reaching a maximum value at the middle position, forming a sequence of second-order basis function values. In constructing the third-order B-spline basis functions, establish a cubic basis function within the effective range formed by multiple adjacent intervals, ensuring continuous connection of the basis function values ​​at the connection points of adjacent intervals, forming a sequence of third-order basis function values. Arrange the first-order, second-order, and third-order basis function value sequences according to their node positions to form sets of first-order, second-order, and third-order B-spline basis functions, respectively. The first-order B-spline basis function set, the second-order B-spline basis function set, and the third-order B-spline basis function set are written into the basis function registration sequence according to the node positions, forming a multi-order B-spline basis function set that participates in subsequent calculations.

[0026] Specifically, the comparison of the basis function response results for each node interval includes: Read the first-order, second-order, and third-order B-spline basis function response values ​​for each node interval to form a response value group for the current node interval. Within each node interval, perform cumulative calculations on the response value groups to form the cumulative interval response value for each node interval. Arrange the cumulative interval response values ​​for each node interval in order of node position to form a set of interval response values. Compare the cumulative interval response values ​​for each node interval in the set of interval response values ​​and extract the target interval with the largest cumulative interval response value. When multiple interval response cumulative values ​​are the same, read the boundary fitting values ​​of the corresponding intervals and extract the target interval with the larger boundary fitting value. Write the target interval numbers for each node interval in order of node position to form interval activation identifiers.

[0027] In this embodiment, the generation of cross-order spline results by the cross-order spline neural network module specifically includes: Read the target order number corresponding to each node position in the order activation identifier, and extract the target order basis function sequence that is consistent with each target order number from the first-order B-spline basis function set, the second-order B-spline basis function set, and the third-order B-spline basis function set. Read the target interval number corresponding to each node position in the interval activation identifier, extract the target interval basis function subset that is consistent with each target interval number from each target order basis function sequence, and form an interval spline calculation sequence; Each feature component in the joint feature vector is written into the corresponding position of the interval spline calculation sequence according to the node position order. The input value, node spacing value and interval span value corresponding to each feature component are read. The basis function response value corresponding to each feature component in the basis function subset of the target interval is calculated to form the basis function response sequence. Read the basis function response values ​​and interval action values ​​corresponding to the basis function subsets of each target interval, perform product calculation and accumulation calculation on the response values ​​corresponding to each position in the basis function response sequence, and form the interval spline output values ​​corresponding to each node position; The interval spline output values ​​corresponding to each node position are grouped according to the target order number to form a first-order spline output sequence, a second-order spline output sequence and a third-order spline output sequence. Read the output values ​​corresponding to each position in the first-order spline output sequence, the second-order spline output sequence, and the third-order spline output sequence, and perform alignment according to the node position order to form a cross-order output alignment sequence; Perform sequential superposition calculations on the output values ​​at each position in the cross-order output alignment sequence to form a cross-order spline intermediate sequence; Read the node continuity values, interval connection values, and boundary change values ​​corresponding to the intermediate sequence of the cross-order spline, perform continuous position merging calculation on the intermediate sequence of the cross-order spline, and generate the cross-order spline result.

[0028] Specifically, performing sequential superposition calculation on the output values ​​at each position in the cross-order output alignment sequence includes: Read the first-order, second-order, and third-order output values ​​corresponding to each node position in the cross-order output alignment sequence; at each node position, first add the first-order and second-order output values ​​to form the first superimposed value; add the first superimposed value to the third-order output value corresponding to the same node position to form the second superimposed value; arrange the second superimposed values ​​corresponding to each node position according to the node position order to form a sequential superimposed sequence; read the superimposed values ​​corresponding to adjacent node positions in the sequential superimposed sequence, and sequentially pass and accumulate the superimposed values ​​of the previous node position and the superimposed values ​​of the next node position to form a continuous superimposed value sequence; write the continuous superimposed value sequence into the output position corresponding to each node position to form the cross-order spline intermediate sequence.

[0029] In this embodiment, the sparse activation module establishes a set of control points at the corresponding positions of the cross-order spline results, and filters the control points participating in the current calculation according to the order activation identifier and the interval activation identifier, specifically including: Read the spline output values ​​corresponding to each node position in the cross-order spline results, and extract the node position numbers corresponding to each spline output value; Read the order activation flag and interval activation flag corresponding to each node position, and extract the target order number and target interval number corresponding to each node position; Concatenate the node position number, target order number and target interval number in sequence to generate the control point number corresponding to each node position. The same node position corresponds to one control point number, and different node positions correspond to different control point numbers. Write the control point number into the control point registration sequence, and associate and record the spline output value, target order number, target interval number and node position number corresponding to each control point number to form an initial control point set; The initial control point set is classified according to the target order number, forming a first-order control point subset, a second-order control point subset, and a third-order control point subset; Control points for the corresponding intervals are extracted from the first-order control point subset, the second-order control point subset, and the third-order control point subset according to the target interval number, forming an interval-selected control point sequence; Perform a number check on each control point in the interval screening control point sequence, retain control points whose node position number, target order number and target interval number are consistent, and delete control points whose three are inconsistent, thus forming a valid control point sequence. The effective control point sequence is arranged in order of node position to form a set of control points participating in the current calculation.

[0030] Specifically, classifying the initial control point set according to the target order number to form first-order control point subsets, second-order control point subsets, and third-order control point subsets includes: Read the control point number, node position number, target order number, and target interval number corresponding to each control point in the initial control point set; Classify and label the control points in the initial control point set according to the target order number of each control point; Extract the control points with target order number one to form a first-order control point subset, and associate and retain the corresponding control point number and node position number; Extract the control points with target order number two to form a second-order control point subset, and associate and retain the corresponding control point number and node position number; Extract the control points with target order number three to form a third-order control point subset, and associate and retain the corresponding control point number and node position number; Arrange the first-order, second-order, and third-order control point subsets in order of node position number to form corresponding control point subset sequences; Write the first-order, second-order, and third-order control point subset sequences into the classification result positions to complete the control point classification process.

[0031] In this embodiment, performing sparse combination calculations on the selected control points to generate sparse activation splines specifically includes: Read the control point number, node position number, spline output value and interval assignment number corresponding to each control point in the effective control point sequence, and form the control point input sequence according to the node position order; Based on the difference between the node position interval value and the spline output value of each control point in the control point input sequence, the local action value of each control point is extracted. The local action values ​​corresponding to each control point are accumulated according to the node position to form a node local combination sequence; Read the target order number and target interval number corresponding to each node position in the local combination sequence of nodes. Retain the local combination values ​​of nodes with the same target order number and target interval number. Delete the local combination values ​​of nodes with different target order numbers or target interval numbers to form a sparse retention sequence. The retained values ​​corresponding to each node position in the sparse retained sequence are merged according to the node position order to form a sparse combined sequence. Perform node continuity checks and interval consistency checks on sparse combination sequences, and output sparse activation splines.

[0032] In this embodiment, the generated frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters specifically include: The sparse activation spline results are written into the parameter mapping process, and the frequency band response values, spatial distribution values ​​and dynamic change values ​​corresponding to each position are extracted to form frequency band adjustment parameters, sound field distribution parameters and dynamic compensation parameters, respectively. The frequency band adjustment parameters, sound field distribution parameters and dynamic compensation parameters are then verified for range and their positions are organized to form a parameter result sequence that can be written into the vehicle audio control unit.

[0033] In this embodiment, writing the frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters into the vehicle audio control unit specifically includes: Read the frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters, and write them into the corresponding positions of the equalization control unit, sound field control unit, and compensation control unit according to the parameter category; perform position matching checks and numerical boundary checks on each parameter after writing to form a control parameter sequence; send the control parameter sequence to the vehicle audio control unit to complete speaker output adjustment, sound field position adjustment, and dynamic compensation control.

[0034] Example 1: To verify the feasibility of this invention in practice, it was applied to an adaptive adjustment scenario for a vehicle cabin audio system. In this scenario, the vehicle audio system needs to handle various playback content, including voice navigation, popular music, movie audio, and radio programs, while also adapting to changes in environmental noise, the number of passengers, and manual adjustments during playback. Traditional audio control methods often use fixed equalizer templates, fixed sound field modes, or compensation based on a single environmental variable. When playback content changes or the vehicle's in-vehicle status continuously changes, problems such as inaccurate frequency band matching, unstable sound field positioning, adjustment lag, and the need for repeated manual adjustments by the user can easily occur. This embodiment addresses this problem by uniformly collecting and processing playback audio data, user operation log data, and vehicle status data, and outputting corresponding sound effect control parameters.

[0035] In practical applications, the data acquisition module reads the audio sampling sequence, channel identifier, and playback type identifier from the played audio data in real time. Simultaneously, it reads volume adjustment records, equalizer adjustment records, sound field mode switching records, and playback switching records from the user operation log. It also reads the vehicle status data, including operating status data, cabin environment data, and occupant status data, to form the input dataset. The feature parsing module processes the input dataset, extracting frequency band energy values, spectral centroid values, dynamic range values, and rhythm intensity values ​​from the played audio data to form an audio feature vector. It extracts adjustment direction values, adjustment amplitude values, dwell time values, and mode selection frequency values ​​from the user operation log to form a preference feature vector. It extracts environmental noise components, spatial enclosure components, operating condition components, and occupant distribution components from the vehicle status data to form an environmental feature vector. Subsequently, the joint encoding module performs unified encoding on the audio feature vector, preference feature vector, and environmental feature vector to form a joint feature vector.

[0036] The cross-order spline neural network module constructs a node sequence based on the joint feature vector, establishes a set of B-spline basis functions of different orders at the corresponding positions of the node sequence, generates order activation labels and interval activation labels, selects the spline order to participate in the current calculation according to the order activation label, and selects the spline interval to participate in the current calculation according to the interval activation label, and outputs the cross-order spline result. The sparse activation module establishes a set of control points at the corresponding positions of the cross-order spline result, filters the control points to participate in the current calculation according to the order activation label and interval activation label, performs sparse combination calculation on the filtered control points, and generates sparse activation spline result. The parameter generation module then generates frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters based on the sparse activation spline result. The output control module writes the parameters into the vehicle audio control unit to realize speaker output adjustment, sound field position adjustment, and dynamic compensation control.

[0037] In the tests, six typical scenarios were selected for verification: voice navigation, pop music, movie audio, radio programs, playback with a full rear passenger seat, and switching between navigation and music. The results showed that when voice navigation was inserted into the playback process, the invention could quickly improve the parameter output corresponding to human voice clarity; when playing pop music or movie audio, the invention could adjust the frequency band distribution based on rhythm intensity, dynamic range, and user preference records; when the number of rear passengers increased, the invention could correct the sound field distribution based on passenger distribution components and spatial closure components; and when navigation and music switched frequently, the invention could still maintain a fast parameter response speed. Overall, the invention outperformed the traditional fixed template control method in terms of parameter matching time, number of secondary user adjustments, and subjective listening experience scores, indicating that the invention can better adapt to the sound effect matching needs in complex in-vehicle environments.

[0038] Table 1: Comparison of Intelligent Matching Effects of Car Audio Systems

[0039] As shown in Table 1, the present invention outperforms the traditional solution in all six typical scenarios. First, consider the parameter matching time: the traditional solution's times are 1.8s, 2.4s, 2.7s, 3.0s, 2.6s, and 3.4s, with an average of 2.65s; the present invention's times are 0.7s, 0.9s, 1.0s, 1.0s, 0.9s, and 1.2s, with an average of 0.95s, a reduction of 1.70s, or approximately 64.2%. This result indicates that the present invention can generate frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters more quickly after changes in playback content and environmental conditions occur. Next, consider the number of secondary adjustments: the traditional solution requires a total of 27 manual adjustments in the six scenarios, while the present invention requires only 8, a reduction of 19, or approximately 70.4%. This indicates that the output of the present invention is closer to actual listening needs and can reduce repeated operations by drivers or passengers. In terms of subjective scoring, the traditional solution averaged 6.70 points, while the present invention averaged 8.75 points, an improvement of 2.05 points. Specifically, in scenario F, which involves frequent switching between navigation and music, the traditional solution scored 6.2 points, while the present invention achieved 8.5 points, an improvement of 2.3 points. In scenario A, the traditional solution scored 7.1 points, while the present invention achieved 8.8 points, an improvement of 1.7 points. Combining parameter matching time and the number of secondary adjustments further illustrates that the present invention does not simply accelerate response speed, but rather provides an adjustment result that better meets the needs of in-vehicle listening in a shorter time, thus exhibiting better real-time performance, matching, and stability overall.

[0040] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based intelligent matching system for in-vehicle audio sound effects, characterized in that, include: The data acquisition module is used to collect audio playback data, user operation record data, and vehicle status data to form an input dataset; The feature parsing module parses the input dataset to generate audio feature vectors, preference feature vectors, and environment feature vectors. The joint encoding module encodes the audio feature vectors, preference feature vectors, and environment feature vectors to form a joint feature vector. The cross-order spline neural network module constructs a node sequence based on the joint feature vector, establishes a set of B-spline basis functions of different orders at the corresponding positions of the node sequence, generates order activation labels and interval activation labels, selects the spline order to participate in the current calculation according to the order activation label, selects the spline interval to participate in the current calculation according to the interval activation label, and performs a combination calculation on the selected spline basis function set to generate cross-order spline results. The sparse activation module is used to establish a set of control points at the corresponding positions of the cross-order spline results, select control points to participate in the current calculation according to the order activation identifier and the interval activation identifier, and perform sparse combination calculation on the selected control points to generate sparse activation spline results; the parameter generation module is used to perform mapping on the sparse activation spline results to generate frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters; the output control module is used to write the frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters into the vehicle audio control unit to complete the sound effect control.

2. The intelligent matching system for vehicle audio sound effects based on deep learning according to claim 1, characterized in that, The feature parsing module performs the parsing of the input data by including the following steps: Read the audio sampling sequence, channel identifier and playback type identifier from the audio data being played. Perform frame segmentation, windowing and spectrum transformation on the audio sampling sequence. Extract the frequency band energy value, spectral centroid value, dynamic range value and rhythm intensity value corresponding to each frame to form an audio feature vector. Read the volume adjustment record, equalizer adjustment record, sound field mode switching record and playback switching record from the user operation record data, and extract the adjustment direction value, adjustment amplitude value, dwell time value and mode selection frequency value in time order to form a preference feature vector; Read the operating status data, cabin environment data and occupant status data from the vehicle status data, perform numerical conversion and identification mapping on each status data, and extract the environmental noise component, spatial enclosure component, operating condition component and occupant distribution component to form an environmental feature vector. The audio feature vector, preference feature vector, and environment feature vector are processed to achieve uniform length and position alignment, resulting in a parsing result corresponding to the same playback time.

3. The intelligent matching system for vehicle audio sound effects based on deep learning according to claim 2, characterized in that, The generation order activation identifier and interval activation identifier of the cross-order spline neural network module specifically include: The feature components in the joint feature vector are written into the node construction process according to the preset input order. The numerical value of each feature component, the difference between adjacent components and the local distribution position are read to form a node position sequence. Based on the node position sequence, the nodes are divided into several continuous node intervals. The interval length, the value of the change within the interval, and the value of the difference between the interval boundaries are counted for each node interval to form an interval analysis sequence. Establish first-order B-spline basis functions, second-order B-spline basis functions and third-order B-spline basis functions at corresponding positions in each node interval, and read the basis function response value, interval coverage value and boundary fitting value of each order B-spline basis function in the current node interval to form an order determination sequence. Compare the response results of each order in the order determination sequence, extract the target order number at the corresponding position, and form an order activation identifier; Perform interval comparison on the basis function response results corresponding to each node interval, extract the target interval number at the corresponding position, and form the interval activation identifier; The order activation flag and the interval activation flag are written in the order of node position to form the spline activation flag sequence corresponding to the current joint feature vector.

4. The intelligent matching system for in-vehicle audio effects based on deep learning according to claim 3, characterized in that, The generation of cross-order spline results by the cross-order spline neural network module specifically includes: Read the target order number corresponding to each node position in the order activation identifier, and extract the target order basis function sequence that is consistent with each target order number from the first-order B-spline basis function set, the second-order B-spline basis function set, and the third-order B-spline basis function set. Read the target interval number corresponding to each node position in the interval activation identifier, extract the target interval basis function subset that is consistent with each target interval number from each target order basis function sequence, and form an interval spline calculation sequence; Each feature component in the joint feature vector is written into the corresponding position of the interval spline calculation sequence according to the node position order. The input value, node spacing value and interval span value corresponding to each feature component are read. The basis function response value corresponding to each feature component in the basis function subset of the target interval is calculated to form the basis function response sequence. Read the basis function response values ​​and interval action values ​​corresponding to the basis function subsets of each target interval, perform product calculation and accumulation calculation on the response values ​​corresponding to each position in the basis function response sequence, and form the interval spline output values ​​corresponding to each node position; The interval spline output values ​​corresponding to each node position are grouped according to the target order number to form a first-order spline output sequence, a second-order spline output sequence and a third-order spline output sequence. Read the output values ​​corresponding to each position in the first-order spline output sequence, the second-order spline output sequence, and the third-order spline output sequence, and perform alignment according to the node position order to form a cross-order output alignment sequence; Perform sequential superposition calculations on the output values ​​at each position in the cross-order output alignment sequence to form a cross-order spline intermediate sequence; Read the node continuity values, interval connection values, and boundary change values ​​corresponding to the intermediate sequence of the cross-order spline, perform continuous position merging calculation on the intermediate sequence of the cross-order spline, and generate the cross-order spline result.

5. The intelligent matching system for vehicle audio sound effects based on deep learning according to claim 4, characterized in that, The sparse activation module establishes a set of control points at the corresponding positions of the cross-order spline results, and filters the control points participating in the current calculation according to the order activation identifier and the interval activation identifier, specifically including: Read the spline output values ​​corresponding to each node position in the cross-order spline results, and extract the node position numbers corresponding to each spline output value; Read the order activation flag and interval activation flag corresponding to each node position, and extract the target order number and target interval number corresponding to each node position; Concatenate the node position number, target order number and target interval number in sequence to generate the control point number corresponding to each node position. The same node position corresponds to one control point number, and different node positions correspond to different control point numbers. Write the control point number into the control point registration sequence, and associate and record the spline output value, target order number, target interval number and node position number corresponding to each control point number to form an initial control point set; The initial control point set is classified according to the target order number, forming a first-order control point subset, a second-order control point subset, and a third-order control point subset; Control points for the corresponding intervals are extracted from the first-order control point subset, the second-order control point subset, and the third-order control point subset according to the target interval number, forming an interval-selected control point sequence; Perform a number check on each control point in the interval screening control point sequence, retain control points whose node position number, target order number and target interval number are consistent, and delete control points whose three are inconsistent to form a valid control point sequence; The effective control point sequence is arranged in order of node position to form a set of control points participating in the current calculation.

6. The intelligent matching system for vehicle audio sound effects based on deep learning according to claim 5, characterized in that, The step of performing sparse combination calculations on the selected control points to generate sparse activation splines specifically includes: Read the control point number, node position number, spline output value and interval assignment number corresponding to each control point in the effective control point sequence, and form the control point input sequence according to the node position order; Based on the difference between the node position interval value and the spline output value of each control point in the control point input sequence, the local action value of each control point is extracted. The local action values ​​corresponding to each control point are accumulated according to the node position to form a node local combination sequence; Read the target order number and target interval number corresponding to each node position in the local combination sequence of nodes. Retain the local combination values ​​of nodes with the same target order number and target interval number. Delete the local combination values ​​of nodes with different target order numbers or target interval numbers to form a sparse retention sequence. The retained values ​​corresponding to each node position in the sparse retained sequence are merged according to the node position order to form a sparse combined sequence. Perform node continuity checks and interval consistency checks on sparse combination sequences, and output sparse activation splines.

7. The intelligent matching system for in-vehicle audio effects based on deep learning according to claim 6, characterized in that, The generated frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters specifically include: The sparse activation spline results are written into the parameter mapping process, and the frequency band response values, spatial distribution values ​​and dynamic change values ​​corresponding to each position are extracted to form frequency band adjustment parameters, sound field distribution parameters and dynamic compensation parameters, respectively. The frequency band adjustment parameters, sound field distribution parameters and dynamic compensation parameters are then verified for range and their positions are organized to form a parameter result sequence that can be written into the vehicle audio control unit.

8. The intelligent matching system for vehicle audio sound effects based on deep learning according to claim 7, characterized in that, The specific steps of writing frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters into the vehicle audio control unit include: Read the frequency band adjustment parameters, sound field distribution parameters, and dynamic compensation parameters, and write them into the corresponding positions of the equalization control unit, sound field control unit, and compensation control unit according to the parameter category; perform position matching checks and numerical boundary checks on each parameter after writing to form a control parameter sequence; send the control parameter sequence to the vehicle audio control unit to complete speaker output adjustment, sound field position adjustment, and dynamic compensation control.

Images