A smart sound system and a method for adjusting a sound field of the smart sound system

Abstract

This invention discloses a smart speaker sound field adjustment method and a smart speaker system, comprising the following steps: S1: Initialize scanning and store indoor acoustic baseline data, which serves as a benchmark for subsequent sound field adjustment effects; S2: Based on the baseline data of step S1, collect multi-dimensional environmental and scene data in real time and complete preprocessing, with the preprocessed data providing effective input for instruction generation in step S3; S3: Based on the effective data preprocessed in step S2, generate precise sound field adjustment instructions through the MoE intelligent decision layer, which directly drives the sound field construction action in step S4; This invention relies on a closed-loop architecture of multi-modal perception layer + MoE intelligent decision layer + collaborative execution layer, and the closed-loop adjustment method of S1-S5, which can integrate multi-dimensional acoustic, environmental, and scene data, breaking through the limitation of traditional single-modal adjustment relying on only a single data, enabling sound field adjustment to accurately adapt to different media characteristics, environmental parameters, and scene requirements.

Description

Technical Field

[0001] This invention belongs to the field of acoustic adjustment technology, and particularly relates to a smart speaker sound field adjustment method and a smart speaker system. Background Technology

[0002] With the development of acoustic technology, sound field adjustment technology has been widely applied in various scenarios such as homes, offices, and vehicles, becoming an important means of optimizing the acoustic environment. Existing sound field adjustment technologies mainly collect acoustic data through sensing modules and combine this data with adjustment modules to adjust sound effect parameters. The core architecture typically includes a data acquisition unit and an execution adjustment unit. The data acquisition unit usually uses a single sensor to obtain basic acoustic information such as volume and frequency, while the execution adjustment unit optimizes the sound field by controlling parameters such as speaker volume and frequency response. This meets users' basic acoustic experience needs and plays a fundamental role in improving the acoustic environment in scenarios such as daily entertainment, office communication, and in-vehicle audio-visual entertainment, promoting the widespread application of acoustic adjustment technology in both civilian and commercial fields.

[0003] However, existing sound field adjustment technologies fall short of meeting the demands for comprehensive and high-precision sound field optimization, leaving many areas for improvement. In terms of adaptability, current technologies lack multi-dimensional data fusion capabilities, failing to accurately adapt to different media characteristics, environmental parameters, and scenario requirements, making comprehensive adaptive adjustment difficult. Regarding intelligence and efficiency, the lack of dynamic decision-making mechanisms prevents flexible matching of adjustment strategies based on scenario needs, easily leading to resource waste and delayed adjustment response. In terms of sound field effects, hardware design adaptability is limited, making it difficult to solve problems such as sound wave aliasing and impedance mismatch, resulting in insufficient sound field uniformity and stability. In terms of versatility, the rigid architecture design and poor compatibility between sensors and execution modules prevent flexible coverage of various spatial scenarios, and some solutions have high power consumption, making them unsuitable for low-power applications. These shortcomings significantly fall short of the requirements for comprehensive, intelligent, and highly adaptable sound field adjustment. Summary of the Invention

[0004] In order to overcome the above-mentioned defects of the prior art, the present invention provides a smart speaker sound field adjustment method and a smart speaker system, which solves the problems of insufficient adaptability, intelligence, effect and versatility of the sound field adjustment technology in the prior art, and the difficulty in meeting the requirements of high precision in all dimensions.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] A method for adjusting the sound field of a smart speaker and a smart speaker system, comprising the following steps:

[0007] S1: Initialize the scan and store the indoor acoustic baseline data, which serves as a reference for subsequent sound field adjustment effects;

[0008] S2: Based on the baseline data from step S1, collect multi-dimensional environmental and scene data in real time and complete preprocessing. The preprocessed data provides effective input for the generation of instructions in step S3.

[0009] S3: Based on the effective data preprocessed in step S2, a precise sound field adjustment command is generated through the MoE intelligent decision layer, which directly drives the sound field construction action in step S4.

[0010] S4: Respond to the sound field adjustment command generated in step S3, complete the adaptive sound field construction, and collect the current sound field effect data and transmit it to step S5;

[0011] S5: Receive the sound field effect data transmitted in step S4, combine it with the baseline data in step S1 to complete the adjustment effect evaluation and decision parameter iterative optimization, and the optimized parameters will be adapted to the subsequent S2 to S4 processes to achieve closed-loop adjustment.

[0012] Preferably, in step S1, the multimodal sensing layer scans all acoustic areas in the room, collects the distribution of indoor media and initial environmental parameters, generates and stores baseline data of the indoor acoustic environment.

[0013] Preferably, the preprocessing in step S2 employs a multimodal early fusion strategy, mapping multi-dimensional data to a unified feature space to generate standardized feature vectors; the feature mapping calculation formula is as follows: In the formula The fused standardized feature vector has a dimension of M×1, where M is the dimension of the fused feature vector, and its value ranges from 128 to 512. The weight matrix is ​​fused, with dimensions M×N representing the total dimensions of the original features, i.e. , , The sum of the feature dimensions; the concatenated vector of the three is the original feature concatenation vector with a dimension of N×1; b is the bias vector with a dimension of M×1; This represents matrix multiplication; in step S3, the formula for calculating the phase compensation of the loudspeaker unit in the micro WFS aliasing compensation subnetwork is as follows: ;

[0014] In the formula $$ The phase compensation value of the kth loudspeaker unit. Let k be the distance from the k-th loudspeaker unit to the reference sound field point. The wavelength of the sound wave is used; the feedback data in step S5 includes sound field uniformity data and user interaction feedback data, wherein the user interaction feedback data includes volume adjustment operation data and mode switching operation data.

[0015] Preferably, a smart audio system includes a multimodal perception layer, a MoE intelligent decision-making layer, and a collaborative execution layer;

[0016] The multimodal perception layer is used to collect and preprocess multi-dimensional environmental and scene data, which includes at least indoor medium acoustic impedance data, spatial environment parameters, and user usage scenario data.

[0017] The MoE intelligent decision layer is used to receive preprocessed data and output sound field adjustment commands.

[0018] The collaborative execution layer is used to respond to instructions to construct an optimized sound field and to feed back adjustment effect data to the MoE intelligent decision layer.

[0019] Preferably, indoor acoustic impedance data is obtained through an acoustic sensor array, spatial environmental parameters are obtained by connecting to external smart home devices through a smart home linkage interface, and user scenario data is obtained by an audio content recognition unit.

[0020] Preferably, the MoE intelligent decision-making layer includes a gating network and four specialized expert sub-networks. Specifically, the four expert sub-networks are a cross-medium impedance matching sub-network, a micro WFS aliasing compensation sub-network, an audiovisual environment collaboration sub-network, and a scene sound effect optimization sub-network. After receiving preprocessed data, the gating network dynamically calls 1 to 3 matching expert sub-networks through activation probability distribution calculations, and jointly outputs sound field adjustment commands. The acoustic impedance matching coefficient of the cross-medium impedance matching sub-network is calculated using the following formula: ;

[0021] In the formula To match the acoustic impedance to the target, The output acoustic impedance of the loudspeaker diaphragm. The acoustic impedance of the target medium indoors is given; meanwhile, the activation probability distribution of the gated network is calculated using the SoftMax function, with the specific calculation formula as follows:

[0022] In the formula Let be the activation probability of the i-th expert subnetwork, with a value ranging from 0 to 1; The input score for the i-th expert subnetwork is... equal + ,in Here is the weight matrix of the gated network, with dimension 1. K represents the total number of expert subnetworks, and D represents the feature dimension of the preprocessed data; is the bias vector of the gated network with dimension K×1; x is the feature vector of the preprocessed data; j is the index of the expert subnetwork, K equals 4; exp is the natural exponential function; the values ​​of i and j are both in the range of 1 to 4.

[0023] Preferably, the collaborative execution layer includes a miniaturized WFS array, a cross-medium impedance matching module, a smart home linkage unit, and a feedback module; wherein the miniaturized WFS array consists of 3 to 4 distributed small speakers, the cross-medium impedance matching module is used to adjust the vibration mode and phase of the speaker diaphragm, and the feedback module feeds back the collected sound field effect data to the MoE intelligent decision layer to achieve iterative optimization of decision parameters.

[0024] Preferably, the acoustic sensor array of the multimodal sensing layer is a microphone array, which, combined with a beamforming algorithm, enables acoustic impedance detection in the frequency range of 0.5 to 100 kHz.

[0025] Preferably, the gated network of the MoE intelligent decision layer adopts a feedforward neural network architecture; the cross-medium impedance matching subnetwork is constructed based on the acoustic-vibration coupling metasurface technology model; and the micro WFS aliasing compensation subnetwork incorporates a lightweight algorithm to solve the truncation error of the small loudspeaker array.

[0026] Preferably, the miniaturized WFS array is packaged with a flexible waveguide coupling structure and has a bending radius of no more than 2 mm; the multimodal sensing layer also includes a millimeter-wave auxiliary sensing module for capturing user position information to improve sound field tracking accuracy.

[0027] The technical effects and advantages of the intelligent speaker sound field adjustment method and intelligent speaker system of this invention are as follows:

[0028] 1. This invention relies on a closed-loop architecture of multimodal perception layer + MoE intelligent decision layer + collaborative execution layer, and closed-loop adjustment method of S1-S5. It can integrate multi-dimensional data of acoustics, environment and scene, break through the limitation of traditional single-modal adjustment that only relies on a single data, and enable sound field adjustment to accurately adapt to different medium characteristics, environmental parameters and scene requirements, and solve the problem of mismatch between sound field and usage scene from the root.

[0029] 2. This invention utilizes the MoE intelligent decision-making layer architecture, where the gated network can dynamically call 1-3 expert sub-networks to work collaboratively. Combined with SoftMax activation probability calculation, it achieves accurate matching of expert sub-networks, avoiding the waste of resources from the continuous operation of all expert sub-networks. At the same time, the application of multimodal early fusion strategies and core algorithm formulas ensures the accuracy and efficiency of decision parameter calculation, enabling rapid generation and response of sound field adjustment commands.

[0030] 3. This invention utilizes a miniaturized WFS array and flexible waveguide coupling packaging design to adapt to installation requirements of different spatial configurations. Combined with a cross-medium impedance matching module and phase compensation algorithm, it effectively solves problems such as aliasing and impedance mismatch during sound wave propagation, significantly improving the uniformity of the sound field. At the same time, the closed-loop feedback optimization process defined in claim 2 can continuously iterate and optimize decision parameters to ensure long-term stability of the sound field effect.

[0031] 4. The modular architecture design of this invention enables the multimodal perception layer to flexibly adapt to different types of sensors and linkage devices. The MoE intelligent decision layer can adapt to new scenario requirements through the expansion of expert sub-networks. The miniaturized and flexible design of the collaborative execution layer can adapt to various spaces such as home, office, and vehicle, greatly improving the applicability of the system. At the same time, the lightweight gating network design can reduce system power consumption and adapt to the needs of low-power usage scenarios. Attached Figure Description

[0032] Figure 1 This invention provides a method for adjusting the sound field of a smart speaker and a flowchart of the smart speaker method.

[0033] Figure 2 This invention presents a method for adjusting the sound field of a smart speaker and a flowchart of a smart speaker system. Detailed Implementation

[0034] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0035] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Moreover, the terms "include," "contain," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that includes a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "includes..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0036] refer to Figure 1-2 This invention discloses a multi-modal perception fusion-based all-dimensional adaptive sound field adjustment method and system. It adopts a closed-loop architecture of "multi-modal perception layer + MoE intelligent decision layer + collaborative execution layer". Through multi-dimensional data fusion and acquisition, dynamic calling of expert sub-networks by gating network, precise calculation of core algorithm and "S1-S5" closed-loop feedback optimization process, it realizes all-dimensional adaptive adjustment of sound field, solves the problems of insufficient adaptability, intelligence, effect and universality of existing technologies, and improves the accuracy and stability of sound field adjustment.

[0037] General technical foundation:

[0038] All embodiments adopt a closed-loop system architecture, wherein: the multimodal perception layer includes at least a microphone array, a smart home linkage interface, and an audio content recognition unit, and may optionally be equipped with a millimeter-wave auxiliary perception module; the MoE intelligent decision layer includes a gated network (feedforward neural network architecture) and four expert sub-networks (cross-medium impedance matching sub-network, miniature WFS aliasing compensation sub-network, audiovisual environment collaboration sub-network, and scene sound effect optimization sub-network); the collaborative execution layer includes a miniaturized WFS array (3-4 distributed small speakers, flexible waveguide coupling package, bending radius ≤2mm), a cross-medium impedance matching module, and a feedback module.

[0039] Each embodiment executes the "S1-S5" closed-loop method process. The preprocessing in step S2 adopts a multimodal early fusion strategy, and the feature mapping formula, acoustic impedance matching formula, SoftMax activation probability formula, and phase compensation formula involved strictly follow the limited form of this invention.

[0040] Example 1

[0041] This embodiment provides a smart speaker sound field adjustment method and a smart speaker system for home theater scenarios (scenarios requiring high-fidelity sound effects). Specific implementation details include:

[0042] Purpose of implementation:

[0043] It is suitable for home theaters with a space of 20㎡, meets the high-fidelity sound requirements when playing action movies and other film and television content, achieves uniform sound field distribution, eliminates echo and distortion, and enhances the user's movie viewing experience.

[0044] Implementation System:

[0045] The system adopts a closed-loop architecture of "multimodal perception layer + MoE intelligent decision-making layer + collaborative execution layer", with the following specific configuration:

[0046] (1) Multimodal perception layer: 8-array microphone array (equipped with beamforming algorithm, detection frequency range 0.5-100kHz) for collecting indoor medium acoustic impedance data; integrated smart home linkage interface to connect with air conditioner and projector to obtain spatial environment parameters; configured with audio content recognition unit for determining scene type; equipped with millimeter wave auxiliary perception module for capturing user real-time location.

[0047] (2) MoE Intelligent Decision Layer: It includes a gating network and four specialized expert sub-networks (cross-medium impedance matching sub-network, micro WFS aliasing compensation sub-network, audiovisual environment collaboration sub-network, and scene sound effect optimization sub-network); among which the gating network adopts a 3-layer feedforward neural network with an input dimension of 256 (matching the feature dimension of preprocessed data), and can dynamically call 1-3 expert sub-networks and jointly output sound field adjustment commands.

[0048] (3) Collaborative execution layer: includes miniaturized WFS array, cross-medium impedance matching module, smart home linkage unit and feedback module; the miniaturized WFS array consists of 4 distributed small speakers, which are encapsulated with flexible waveguide coupling and have a bending radius of 1.5mm; the cross-medium impedance matching module is used to adjust the vibration mode and phase of the speaker diaphragm; the feedback module is used to collect sound field effect data and feed it back to the MoE intelligent decision layer.

[0049] Implementation steps:

[0050] S1: Initialize the scan of the entire living room area (20㎡), collect the distribution of indoor media (wooden floor, walls, fabric sofa) and initial environmental parameters (sound pressure level of 35dB in quiet state), generate acoustic baseline data and store it.

[0051] S2: Based on the baseline data of S1, real-time acquisition of multi-dimensional data: microphone array detection of indoor medium acoustic impedance Z2=3.2×10 6 Pa·s / m³, smart home linkage obtains temperature 25℃ and humidity 50%, audio content recognition unit determines the scene as "action movie playback"; in the preprocessing stage, a multimodal early fusion strategy is adopted, and the fusion feature vector x is calculated through the feature mapping formula, where the acoustic impedance feature dimension is 64, the environmental parameter feature dimension is 32, the scene feature dimension is 64, the original feature total dimension is 160, the fusion weight matrix dimension is set to 256×160 (fusion feature dimension 256, in the range of 128-512), the bias vector dimension is 256×1, and finally a fusion feature vector of 256×1 is generated.

[0052] S3: The MoE intelligent decision layer receives the fused feature vector. The gating network calculates the activation probabilities using the SoftMax function: σ(z1)=0.35 (cross-medium impedance adaptation sub-network), σ(z2)=0.4 (miniature WFS aliasing compensation sub-network), σ(z3)=0.2 (audiovisual environment collaboration sub-network), and σ(z4)=0.05 (scene sound effect optimization sub-network). The first three expert sub-networks are dynamically invoked. The cross-medium impedance adaptation sub-network calculates the target matching acoustic impedance using the acoustic impedance matching formula. (Speaker diaphragm output acoustic impedance Z1=2.8×10) 6 Pa·s / m³, final Z = 2.99 × 10 6Pa·s / m³); The miniature WFS aliasing compensation sub-network calculates φ=2πr / λ using the phase compensation formula (the distances from the four speakers to the user's position are 3.0m, 3.2m, 2.8m, and 3.1m, respectively; the main audio frequency is 1kHz; the sound wave wavelength is 0.34m; and the calculated phase compensation values ​​are 55.1rad, 58.8rad, 51.5rad, and 56.9rad, respectively); The combined output includes impedance matching parameters, phase compensation parameters, and a sound field adjustment command with a volume of 60dB.

[0053] S4: The collaborative execution layer responds to the adjustment command, adjusts the speaker diaphragm vibration mode to the state corresponding to the target matching acoustic impedance, and outputs the compensated phase audio signal through 4 WFS speakers to construct a uniform sound field; the feedback module collects the current sound field uniformity data (sound pressure level deviation in each area ≤2dB).

[0054] S5: Combine the baseline data from S1 and the sound field uniformity data from S4 to evaluate the adjustment effect (high-fidelity sound effect meets the standard, no echo, no distortion); collect user interaction feedback data (the volume is appropriate, no adjustment is needed), iteratively optimize the local parameters of the fusion weight matrix, and adapt to the sound effect requirements of different scenes in subsequent action movies.

[0055] Implementation results:

[0056] With a sound field uniformity deviation of ≤2dB, explosions and dialogues in action movies are clearly distinguishable, with no echo or distortion issues, providing a superior viewing experience compared to traditional fixed sound effect adjustment solutions.

[0057] Example 2

[0058] This embodiment provides a smart speaker sound field adjustment method and a smart speaker system for office space scenarios (scenarios requiring clear voice conferencing). Specific implementation details include:

[0059] Purpose of implementation:

[0060] It is suitable for office spaces of up to 15 square meters, meets the needs of clear voice in multi-person voice conference scenarios, suppresses echo interference, ensures uniform sound field for each participant, and improves meeting communication efficiency.

[0061] Implementation System:

[0062] The system adopts a closed-loop architecture of "multimodal perception layer + MoE intelligent decision-making layer + collaborative execution layer", with the following specific configuration:

[0063] (1) Multimodal perception layer: 4-array microphone array (equipped with beamforming algorithm, detection frequency range 0.5-100kHz) for collecting indoor medium acoustic impedance data; integrated smart home linkage interface to connect with air conditioner and smart light to obtain environmental parameters; configured audio content recognition unit for determining "voice conference scene"; equipped with millimeter wave auxiliary perception module for capturing the location of multiple users around the conference table.

[0064] (2) MoE Intelligent Decision Layer: It includes a gating network and four specialized expert sub-networks. The gating network can dynamically call 1-3 expert sub-networks. In this embodiment, the cross-medium impedance adaptation sub-network and the audiovisual environment collaboration sub-network are called.

[0065] (3) Collaborative execution layer: includes miniaturized WFS array, cross-medium impedance matching module, smart home linkage unit and feedback module; the miniaturized WFS array consists of 3 distributed small speakers, which are encapsulated with flexible waveguide coupling and have a bending radius of 2mm, and are installed on the edge of the conference table; the feedback module is used to collect sound field effect data and feed it back to the MoE intelligent decision layer.

[0066] Implementation steps:

[0067] S1: Initialize a 15㎡ office scan, collect indoor media distribution (floor tiles + glass walls) and initial environmental parameters (sound pressure level of 30dB in quiet state), generate acoustic baseline data and store it.

[0068] S2: Based on the baseline data of S1, real-time acquisition of multi-dimensional data: microphone array detection of indoor medium acoustic impedance Z2=4.5×10 6 Pa·s / m³, smart home linkage obtains temperature 23℃ and humidity 45%, audio content recognition unit determines the scene as "voice conference"; in the preprocessing stage, a multimodal early fusion strategy is adopted, and the fusion feature vector x is calculated by feature mapping formula (fusion feature dimension 128, in the range of 128-512).

[0069] S3: The MoE intelligent decision-making layer receives the fused feature vector, and after the gating network calculates the activation probability, it calls the cross-medium impedance adaptation sub-network and the audiovisual environment collaboration sub-network; the cross-medium impedance adaptation network calculates the target matching acoustic impedance using the acoustic impedance matching formula. (Speaker diaphragm output acoustic impedance Z1=2.8×10) 6 Pa·s / m³, final Z = 3.97 × 10 6 Pa·s / m³); combined output of "voice enhancement + uniform sound field" adjustment commands.

[0070] S4: The collaborative execution layer responds to adjustment commands, adjusts the vibration mode of the speaker diaphragm, and outputs audio signals through the WFS speaker to construct a uniform sound field; the feedback module collects the current sound field uniformity data (sound pressure level deviation at each user location ≤ 1.5dB).

[0071] S5: Combine the baseline data from S1 and the sound field uniformity data from S4 to evaluate the adjustment effect (30% improvement in speech clarity); iteratively optimize decision parameters to adapt to the voice transmission needs of office spaces.

[0072] Implementation results:

[0073] The conference audio was clear and free of background noise, and the sound field was uniform for all participants, effectively suppressing office echo interference and improving the communication efficiency of multi-person meetings.

[0074] Example 3

[0075] This embodiment provides a smart speaker sound field adjustment method and a smart speaker system for use in a bedroom setting (a scenario requiring low noise and sleep aid). Specific implementation details include:

[0076] Purpose of implementation:

[0077] It is designed to fit bedroom spaces, meet users' sleep-aid needs, create a low-noise, soft sound field environment, reduce environmental noise interference, and improve the sleep-aid effect. At the same time, it adopts a lightweight design to reduce device power consumption.

[0078] Implementation System:

[0079] The system adopts a closed-loop architecture of "multimodal perception layer + MoE intelligent decision-making layer + collaborative execution layer", with the following specific configuration:

[0080] (1) Multimodal sensing layer: 4-array microphone array for collecting indoor medium acoustic impedance data; integrated smart home linkage interface to connect to humidifier to obtain environmental parameters; configured audio content recognition unit for determining "sleep aid scene" (white noise playback); simplified design, no millimeter wave auxiliary sensing module configured.

[0081] (2) MoE Intelligent Decision Layer: It includes a gating network and four specialized expert sub-networks. The gating network adopts a lightweight feedforward neural network (to reduce power consumption). In this embodiment, the micro WFS aliasing compensation sub-network and scene sound effect optimization sub-network are called.

[0082] (3) Collaborative execution layer: includes miniaturized WFS array, cross-medium impedance matching module, smart home linkage unit and feedback module; the miniaturized WFS array consists of 3 distributed small speakers, which are encapsulated with flexible waveguide coupling and have a bending radius of 1mm, and are installed on both sides of the headboard; the feedback module is used to collect sound field effect data and user interaction feedback data.

[0083] Implementation steps:

[0084] S1: Initialize the scanning of the bedroom space, collect the indoor media distribution (carpet + wallpaper) and initial environmental parameters, generate acoustic baseline data and store it.

[0085] S2: Based on the baseline data of S1, real-time acquisition of multi-dimensional data: microphone array detection of indoor medium acoustic impedance Z2=2.5×10 6 Pa·s / m³, smart home linkage obtains temperature 22℃ and humidity 60%, audio content recognition unit determines the scene as "sleep aid scene" (white noise playback); in the preprocessing stage, a multimodal early fusion strategy is adopted, and the fusion feature vector x is calculated through the feature mapping formula.

[0086] S3: The MoE intelligent decision layer receives the fused feature vector, and after the gating network calculates the activation probability, it calls the micro WFS aliasing compensation sub-network and the scene sound effect optimization sub-network; it adjusts the output phase of the speaker through the phase compensation formula to avoid the sound waves directly hitting the user's ears and generates a "low noise + soft sound field" adjustment command.

[0087] S4: The collaborative execution layer responds to adjustment commands and outputs low-noise, soft white noise signals through the WFS speaker to create a sleep-aiding sound field; the feedback module collects current sound field effect data and user status data.

[0088] S5: Combine the baseline data from S1 and the data collected from S4 to evaluate the adjustment effect; focus on collecting user interaction feedback data such as turning over and volume adjustment, iteratively optimize the white noise frequency matching parameters, and continuously adapt to the user's sleep aid needs.

[0089] Implementation results:

[0090] The sound field is soft and uncompressive, reducing ambient noise by 25% and significantly aiding sleep; the device has low power consumption, making it suitable for long-term standby use in the bedroom.

[0091] Example 4

[0092] This embodiment provides a smart speaker sound field adjustment method and a smart speaker system for use in a children's room scenario (a scenario requiring safety and low power consumption). Specific implementation details include:

[0093] Purpose of implementation:

[0094] It is designed to fit children's room spaces and meet the safe sound requirements for scenarios such as children's story playback. It limits the frequency and volume of sound effects to avoid stimulating children's hearing, and adopts a low power consumption design to be suitable for long-term use.

[0095] Implementation System:

[0096] The system adopts a closed-loop architecture of "multimodal perception layer + MoE intelligent decision-making layer + collaborative execution layer", with the following specific configuration:

[0097] (1) Multimodal perception layer: Low-power 4-array microphone array for collecting indoor medium acoustic impedance data; integrated smart home linkage interface to connect to temperature and humidity meter to obtain environmental parameters; configured audio content recognition unit for determining "children's story scene".

[0098] (2) MoE Intelligent Decision Layer: It includes a gating network and four specialized expert sub-networks. The gating network limits the output sound effect frequency range to 20Hz-16kHz (to protect children's hearing). In this embodiment, three expert sub-networks are called to work together.

[0099] (3) Collaborative execution layer: includes miniaturized WFS array, cross-medium impedance matching module, smart home linkage unit and feedback module; the miniaturized WFS array consists of 3 distributed small speakers, which are packaged with flexible waveguide coupling and have a bending radius of 1.2mm, and are equipped with low power driving module; the feedback module is used to collect sound field effect data and child status feedback data.

[0100] Implementation steps:

[0101] S1: Initialize the scanning of the children's room space, collect the indoor media distribution (wood flooring + fabric toys) and initial environmental parameters, generate acoustic baseline data and store it.

[0102] S2: Based on the baseline data of S1, real-time acquisition of multi-dimensional data: microphone array detection of indoor medium acoustic impedance Z2=3.0×10 6 Pa·s / m³, smart home linkage obtains environmental temperature and humidity parameters, audio content recognition unit determines the scene as "children's story scene"; in the preprocessing stage, a multimodal early fusion strategy is adopted to strengthen the "volume safety threshold" screening (avoid volume exceeding 60dB), and the fusion feature vector x is calculated through feature mapping formula.

[0103] S3: The MoE intelligent decision layer receives the fused feature vector, and after the gating network calculates the activation probability, it calls 3 expert sub-networks to generate story sound effect adjustment instructions adapted to children (frequency 20Hz-16kHz, volume ≤60dB) in combination with the requirements for children's hearing protection.

[0104] S4: The collaborative execution layer responds to adjustment commands and outputs story sound effects through low-power WFS speakers to build a safe and uniform sound field; the feedback module collects current sound field effect data and child status data (crying, quiet, etc.).

[0105] S5: Combine the baseline data from S1 and the collected data from S4 to evaluate the adjustment effect; prioritize iterative optimization of the story sound effect volume and frequency parameters based on children's status feedback data to ensure that the sound effect is safe and non-stimulating.

[0106] Implementation results:

[0107] The sound effects are safe and non-irritating, meeting the requirements for children's hearing protection; the device power consumption is reduced by 40%, making it suitable for long-term use in children's rooms, and children have a high acceptance of the story sound effects.

[0108] Example 5

[0109] This embodiment provides a smart speaker sound field adjustment method and a smart speaker system for intelligent vehicle scenarios (mobile sound field following demand scenarios), and the specific implementation content includes:

[0110] Purpose of implementation:

[0111] It adapts to the interior space of smart cars, meets the sound field following requirements of in-car music playback scenarios, and ensures that the sound field always follows the position of the driver / passenger during dynamic changes in vehicle speed, maintaining music clarity and avoiding resonance noise.

[0112] Implementation System:

[0113] The system adopts a closed-loop architecture of "multimodal perception layer + MoE intelligent decision-making layer + collaborative execution layer", with the following specific configuration:

[0114] (1) Multimodal perception layer: 6-array microphone array for collecting in-vehicle medium acoustic impedance data; integrated vehicle system linkage interface to obtain environmental parameters such as vehicle speed and in-vehicle temperature; configured audio content recognition unit for determining "in-vehicle music scene"; equipped with millimeter wave auxiliary perception module for real-time capture of driver / passenger position (dynamic tracking).

[0115] (2) MoE Intelligent Decision Layer: It includes a gating network and four specialized expert sub-networks. Due to the dynamic change of the medium acoustic impedance caused by the change of vehicle speed, this embodiment calls all four expert sub-networks to work together to achieve dynamic adaptation.

[0116] (3) Collaborative execution layer: includes miniaturized WFS array, cross-medium impedance matching module, vehicle linkage unit and feedback module; the miniaturized WFS array consists of 4 distributed small speakers, which are encapsulated with flexible waveguide coupling and have a bending radius of 1.8mm. It is installed on the inside of the car door to adapt to the curved space inside the car; the feedback module is used to collect sound field effect data in real time.

[0117] Implementation steps:

[0118] S1: Initialize the scanning of the vehicle interior space, collect the distribution of media (leather seats + glass) and initial environmental parameters, generate acoustic baseline data and store it.

[0119] S2: Based on the baseline data of S1, real-time acquisition of multi-dimensional dynamic data: microphone array detection of the acoustic impedance of the in-vehicle medium Z2=4.0×10 6 Pa·s / m³, the vehicle system is linked to obtain vehicle speed (dynamically changing from 0-120km / h) and in-vehicle temperature, the audio content recognition unit determines the scene as "in-vehicle music", and the millimeter wave module captures the driver / passenger position in real time; in the preprocessing stage, a multimodal early fusion strategy is adopted to update the environmental parameters corresponding to the vehicle speed in real time, dynamically adjust the preprocessed data, and calculate the fused feature vector x through the feature mapping formula.

[0120] S3: The MoE intelligent decision layer receives the fused feature vector, and after the gating network calculates the activation probability, it calls all four expert sub-networks; based on the dynamic changes in the user's position captured by the millimeter-wave module, it updates the speaker-to-user distance parameter in the phase compensation formula in real time, and calculates the adaptation parameters through acoustic impedance matching formula, phase compensation formula, etc., to generate dynamic sound field adjustment commands.

[0121] S4: The collaborative execution layer responds to adjustment commands and outputs adapted music signals through the WFS speakers inside the doors to construct a dynamic following sound field; the feedback module collects current sound field effect data (uniformity, clarity, etc.) in real time.

[0122] S5: Combine the baseline data from S1 and the real-time data collected from S4 to evaluate the adjustment effect; iteratively optimize the decision parameters to ensure that the sound field always follows the user within the vehicle speed range of 0-120km / h, with a uniformity deviation of ≤2dB.

[0123] Implementation results:

[0124] Within a speed range of 0-120km / h, the sound field remains stable, following the driver / passenger's position, with no loss of music clarity and no in-vehicle resonance noise, thus meeting the sound effect requirements of a dynamic in-vehicle environment.

[0125] Comparative Example 1

[0126] This comparative example provides a traditional single-mode sound field adjustment scheme, the specific contents of which include:

[0127] Solution configuration:

[0128] It adopts a "single microphone + fixed sound effect adjustment module" architecture, without a MoE intelligent decision layer and without multimodal data fusion; it only collects volume data through a single microphone and adjusts the speaker volume based on a fixed threshold, without involving acoustic impedance matching, WFS phase compensation and other technologies; there is no closed-loop feedback optimization process.

[0129] Test scenario:

[0130] The home theater scenario was the same as in Example 1, and the test conditions were completely identical.

[0131] Test results:

[0132] (1) The sound field uniformity deviation is ≥8dB, and the sound distortion in the sofa edge area is severe;

[0133] (2) It cannot adapt to the medium characteristics of wooden flooring + fabric sofa, and there is echo interference;

[0134] (3) Unable to identify movie scene type; the volume of explosion sounds and dialogue sounds in action movies is unbalanced;

[0135] (4) No feedback optimization, users need to frequently adjust the volume manually, resulting in a very poor experience.

[0136] The comparison shows that the "multimodal perception fusion + MoE intelligent decision-making + closed-loop feedback" architecture of the present invention can effectively solve the problems of "non-uniform sound field, poor medium adaptation, weak scene adaptation, and lack of dynamic optimization" in traditional single-modal solutions. The various embodiments cover the differentiated needs of different scenarios, verifying the rationality and necessity of the core technical features (expert sub-network, formula parameters, module configuration) of the present invention, and fully demonstrating the technical advantages of the present invention.

[0137] The necessity of a multimodal perception layer: In each embodiment, scene differentiation adaptation can only be achieved through multi-dimensional data fusion (acoustic impedance + environment + scene). The technical significance of multimodal perception fusion is verified by comparing it with the single-modal data acquisition of the comparative proportion.

[0138] The necessity of the MoE intelligent decision-making layer: In Examples 1-5, the dynamic invocation of expert sub-networks can adapt to the core needs of different scenarios (such as home theaters requiring multi-expert collaboration, and office scenarios requiring streamlined experts to reduce power consumption). Compared with the fixed adjustment mode of the comparison ratio, the technical advantages of the MoE intelligent decision-making architecture are verified.

[0139] The necessity of the closed-loop method: The closed-loop process of S1-S5 ensures continuous optimization of the sound field adjustment effect. Compared with the feedback-free mode of the comparative model, the technical value of the closed-loop method is verified.

[0140] Compared with Examples 1-5 and Comparative Example 1, the core difference between Examples 1-5 and Comparative Example 1 (traditional single-modal sound field adjustment scheme) lies in the systematic nature of the architecture design and the intelligence of the adjustment logic. The former is based on a closed-loop architecture of "multimodal perception layer + MoE intelligent decision layer + collaborative execution layer" and "S1-S5" dynamic optimization method, covering five typical scenarios: home theater, office space, bedroom, children's room, and smart car, to achieve full-dimensional adaptive sound field adjustment; the latter adopts a simple architecture of "single microphone + fixed sound effect adjustment module", without intelligent decision-making and closed-loop optimization capabilities. The two show significant differences in scene adaptation, adjustment effect and user experience.

[0141] At the architectural design level, Examples 1-5 all adopt a multi-module collaborative closed-loop architecture. The multimodal perception layer can integrate multi-dimensional data such as microphone arrays, linkage interfaces, and audio recognition, adapting to changes in media (wooden floors, floor tiles, leather, etc.) and environment (temperature, humidity, vehicle speed) in different scenarios. The MoE intelligent decision-making layer achieves accurate calculations such as acoustic impedance matching and phase compensation by dynamically calling 1-4 expert sub-networks. The flexible WFS array of the collaborative execution layer can adapt to different spatial forms. In contrast, the comparative example only relies on a single microphone to collect volume data, lacks an intelligent decision-making module, and cannot achieve multi-dimensional data fusion and dynamic parameter optimization, resulting in significant architectural limitations.

[0142] In terms of scene adaptability, Examples 1-5 can accurately match differentiated needs such as high fidelity, clear voice, low noise for sleep aid, safety and low power consumption, and dynamic tracking. For example, in a home theater scenario, the sound field uniformity deviation is ≤2dB; in an office scenario, voice clarity is improved by 30%; and in a smart car scenario, the sound field maintains stable tracking within a speed range of 0-120km / h. The comparative examples, however, can only achieve fixed threshold volume adjustment and cannot adapt to the media characteristics and usage requirements of different scenarios. For instance, in a home theater scenario, the sound field uniformity deviation is ≥8dB, resulting in severe echo and distortion; and in an office scenario, echo interference cannot be suppressed.

[0143] In terms of adjustment effect and user experience, Examples 1-5, through core technologies such as multimodal early fusion and SoftMax activation probability calculation, combined with closed-loop feedback iterative optimization, can reduce environmental noise by 25%-40% without manual user intervention; in children's room scenarios, it can also precisely control sound frequency and volume, ensuring safety. The comparison example, lacking a closed-loop feedback mechanism, requires frequent manual volume adjustment by the user and cannot solve problems such as uneven sound field and sound distortion, resulting in a very poor user experience.

[0144] In summary, Examples 1-5 verify the universality and superiority of the "multimodal perception fusion + MoE intelligent decision-making + closed-loop feedback" architecture of the present invention. It can effectively solve the core shortcomings of the comparative model, such as "weak adaptation, poor effect, and inferior experience", and provide a precise, stable, and intelligent sound field adjustment solution for different scenarios, highlighting the technical value and application prospects of the present invention.

[0145] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of protection of the claims.

[0146] In conclusion, the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for adjusting the sound field of a smart speaker, characterized in that, Includes the following steps: S1: Initialize the scan and store the indoor acoustic baseline data, which serves as a reference for subsequent sound field adjustment effects; S2: Based on the baseline data from step S1, collect multi-dimensional environmental and scene data in real time and complete preprocessing. The preprocessed data provides effective input for the generation of instructions in step S3. S3: Based on the effective data preprocessed in step S2, a precise sound field adjustment command is generated through the MoE intelligent decision layer, which directly drives the sound field construction action in step S4. S4: Respond to the sound field adjustment command generated in step S3, complete the adaptive sound field construction, and collect the current sound field effect data and transmit it to step S5; S5: Receive the sound field effect data transmitted in step S4, combine it with the baseline data in step S1 to complete the adjustment effect evaluation and decision parameter iterative optimization, and the optimized parameters will be adapted to the subsequent S2 to S4 processes to achieve closed-loop adjustment.

2. The intelligent speaker sound field adjustment method as described in claim 1, characterized in that, In step S1, the multimodal sensing layer scans all acoustic areas in the room, collects the distribution of indoor media and initial environmental parameters, generates and stores baseline data of the indoor acoustic environment.

3. The intelligent speaker sound field adjustment method as described in claim 1, characterized in that, Step S2 preprocessing employs a multimodal early fusion strategy, mapping multi-dimensional data to a unified feature space to generate standardized feature vectors; the feature mapping calculation formula is as follows: In the formula The fused standardized feature vector has a dimension of M×1, where M is the dimension of the fused feature vector, and its value ranges from 128 to 512. The weight matrix is ​​fused, with dimensions M×N representing the total dimensions of the original features, i.e. , , The sum of the feature dimensions; The concatenated vector of the three is the original feature concatenation vector, with a dimension of N×1; b is the bias vector, with a dimension of M×1; This represents matrix multiplication; in step S3, the formula for calculating the phase compensation of the loudspeaker unit in the micro WFS aliasing compensation subnetwork is as follows: ; In the formula $$ The phase compensation value of the kth loudspeaker unit. Let k be the distance from the k-th loudspeaker unit to the reference sound field point. The wavelength of the sound wave is used; the feedback data in step S5 includes sound field uniformity data and user interaction feedback data, wherein the user interaction feedback data includes volume adjustment operation data and mode switching operation data.

4. A smart audio system, characterized in that, It includes a multimodal perception layer, a MoE intelligent decision-making layer, and a collaborative execution layer; The multimodal perception layer is used to collect and preprocess multi-dimensional environmental and scene data, which includes at least indoor medium acoustic impedance data, spatial environment parameters, and user usage scenario data. The MoE intelligent decision layer is used to receive preprocessed data and output sound field adjustment commands. The collaborative execution layer is used to respond to instructions to construct an optimized sound field and to feed back adjustment effect data to the MoE intelligent decision layer.

5. The intelligent audio system as described in claim 4, characterized in that, Indoor acoustic impedance data is obtained through an acoustic sensor array; spatial environmental parameters are obtained by connecting to external smart home devices through a smart home linkage interface; and user scenario data is obtained by an audio content recognition unit.

6. The intelligent audio system as described in claim 4, characterized in that, The MoE intelligent decision-making layer includes a gating network and four specialized expert sub-networks. Specifically, the four expert sub-networks are a cross-medium impedance matching sub-network, a micro WFS aliasing compensation sub-network, an audiovisual environment collaboration sub-network, and a scene sound effect optimization sub-network. After receiving preprocessed data, the gating network dynamically calls 1 to 3 matched expert sub-networks through activation probability distribution calculations, jointly outputting sound field adjustment commands. The acoustic impedance matching coefficient of the cross-medium impedance matching sub-network is calculated using the following formula: ; In the formula To match the acoustic impedance to the target, The output acoustic impedance of the loudspeaker diaphragm. The acoustic impedance of the target medium indoors is given; meanwhile, the activation probability distribution of the gated network is calculated using the SoftMax function, with the specific calculation formula as follows: In the formula Let be the activation probability of the i-th expert subnetwork, with a value ranging from 0 to 1; The input score for the i-th expert subnetwork is... equal + ,in Here is the weight matrix of the gated network, with dimension 1. K represents the total number of expert subnetworks, and D represents the feature dimension of the preprocessed data; is the bias vector of the gated network with dimension K×1; x is the feature vector of the preprocessed data; j is the index of the expert subnetwork, K equals 4; exp is the natural exponential function; the values ​​of i and j are both in the range of 1 to 4.

7. The intelligent audio system as described in claim 4, characterized in that, The collaborative execution layer includes a miniaturized WFS array, a cross-medium impedance matching module, a smart home linkage unit, and a feedback module. The miniaturized WFS array consists of 3 to 4 distributed small speakers. The cross-medium impedance matching module is used to adjust the vibration mode and phase of the speaker diaphragm. The feedback module feeds back the collected sound field effect data to the MoE intelligent decision layer to achieve iterative optimization of decision parameters.

8. The intelligent audio system as described in claim 4, characterized in that, The acoustic sensor array of the multimodal sensing layer is a microphone array, which, combined with a beamforming algorithm, enables acoustic impedance detection in the frequency range of 0.5 to 100 kHz.

9. A smart audio system as described in claim 4, characterized in that, The gated network of the MoE intelligent decision layer adopts a feedforward neural network architecture; the cross-medium impedance matching subnetwork is constructed based on the acoustic-vibration coupling metasurface technology model; and the micro WFS aliasing compensation subnetwork incorporates a lightweight algorithm to solve the truncation error of small loudspeaker arrays.

10. A smart audio system as described in claim 4, characterized in that, The miniaturized WFS array is packaged with a flexible waveguide coupling structure and has a bending radius of no more than 2 mm. The multimodal sensing layer also includes a millimeter-wave auxiliary sensing module, which is used to capture user position information to improve sound field tracking accuracy.

Images