Intelligent audio signal processing and transmission system of Bluetooth sound equipment

Abstract

The invention provides an intelligent audio signal processing and transmission system of Bluetooth sound equipment, and relates to the technical field of Bluetooth sound equipment systems. The intelligent audio signal processing and transmission system of the Bluetooth sound equipment comprises a signal acquisition module which is used for acquiring an environment audio signal, a noise signal and a feedback signal for playing audio; the intelligent signal processing module is integrated with an algorithm grading and dynamic switch unit; the Bluetooth transmission module is used for realizing audio signal wireless transmission between the sound source equipment and the sound equipment and is adaptive to multi-protocol coding and decoding; the power amplification and playing module is used for performing power amplification on the processed audio signal and playing the audio signal through a loudspeaker; and the user feedback module is used for receiving a feedback instruction triggered by the user due to hearing discomfort, and synchronously recording the timestamp of a feedback moment, the frequency band of the corresponding audio clip, the frequency intensity and the current scene parameter. Personalized audio adaptation is realized through user feedback and machine learning, the operation threshold is reduced, and the use comfort of different users is improved.

Description

Technical Field

[0001] This invention relates to the field of Bluetooth speaker system technology, specifically to an intelligent audio signal processing and transmission system for Bluetooth speakers. Background Technology

[0002] Bluetooth speakers, as portable audio playback devices, have been widely used in various scenarios such as homes, outdoors, and commutes due to the convenience of wireless transmission. As users' demands for sound quality have increased, existing Bluetooth speakers have gradually integrated intelligent audio processing functions such as spatial audio, multi-level noise reduction, and dynamic EQ adjustment. By improving the computing power of chips to support the operation of complex algorithms, sound quality optimization is achieved.

[0003] However, in existing technologies, there is an irreconcilable contradiction between improving computing power and ensuring sound quality. To meet the needs of adapting to multiple scenarios, the complexity of audio processing algorithms is constantly increasing. Low-power chips are prone to audio compression when running multiple tasks in parallel, resulting in hidden sound quality losses such as reduced dynamic range and loss of high-frequency details. These losses are not easily detected in the early stages, but they will significantly affect the listening experience.

[0004] Meanwhile, the audio output parameters of existing Bluetooth speakers are mostly set based on a general auditory model, without considering individual differences in auditory tolerance. Some extreme frequencies generated to enhance sound quality may cause auditory stimulation for certain users. Because different users have different tolerances to frequency intensity, uniform parameter limits cannot suit all users, easily leading to auditory discomfort for some users.

[0005] Furthermore, the existing systems operate on relatively fixed algorithms, making it difficult to dynamically adjust audio processing strategies based on user feedback. Even though some devices support manual EQ parameter adjustment, users need professional acoustic knowledge, resulting in a high barrier to entry and preventing personalized adaptation without manual intervention. These issues collectively hinder the further development of intelligent audio processing technology for Bluetooth speakers and the improvement of their user experience. Summary of the Invention

[0006] To address the shortcomings of existing technologies, this invention provides an intelligent audio signal processing and transmission system for Bluetooth speakers, which solves the problems of implicit sound quality loss caused by the increase in computing power of existing Bluetooth speakers, as well as the inability of general audio parameters to adapt to individual differences in hearing tolerance and the lack of user feedback for adjustment.

[0007] Technical solution To achieve the above objectives, the present invention provides the following technical solution: an intelligent audio signal processing and transmission system for a Bluetooth speaker, comprising: The signal acquisition module is used to acquire environmental audio signals, noise signals, and feedback signals from played audio, and output raw acquisition data containing frequency and intensity characteristics. The intelligent signal processing module integrates an algorithm hierarchical and dynamic switching unit, which is used to hierarchically schedule audio processing algorithms according to scene parameters and system load. The Bluetooth transmission module is used to enable wireless transmission of audio signals between the audio source device and the speaker, and is compatible with multiple protocol codecs. The power amplification and playback module is used to amplify the processed audio signal and play it through a speaker; The user feedback module is used to receive feedback commands triggered by users due to hearing discomfort, and synchronously record the timestamp of the feedback time, the frequency segment of the corresponding audio segment, the frequency intensity, and the current scene parameters. The machine learning adaptation module is used to train a personalized hearing tolerance model based on user feedback data. It dynamically adjusts the frequency intensity threshold of the audio signal according to the model output to achieve a balance between sound quality optimization and hearing comfort.

[0008] Preferably, the algorithm grading and dynamic switching unit divides the audio processing algorithm into three levels: high-level, medium-level, and low-level. Each level corresponds to different computing power consumption and algorithm combinations, wherein: The advanced algorithm suite includes a full HRTF spatial audio algorithm, a multi-level ANC noise reduction algorithm, a 10-band EQ adjustment algorithm, and a dynamic range extension algorithm, which are suitable for quiet indoor scenes and lossless audio source playback, with computing power consumption meeting DSP load >60%; The intermediate algorithm combination includes a basic noise reduction algorithm, a 5-band EQ adjustment algorithm, and a dynamic range compression algorithm, which is suitable for commuting / outdoor scenarios and ordinary audio sources, and the computing power consumption meets 30%-60% of the DSP load. The low-level algorithm combination retains only echo cancellation, automatic gain control, and basic encoding / decoding algorithms, making it suitable for voice call / standby scenarios and low bitrate audio sources, with computing power consumption meeting the requirement of DSP load <30%. The algorithm grading and dynamic switching unit automatically triggers a smooth switching of algorithm levels by monitoring environmental noise, sound source type, system load and user behavior in real time. The switching process uses 10-20ms cross fade-in and fade-out processing.

[0009] Preferably, the user feedback module includes a triggering unit and an information recording unit; the triggering unit supports three methods: physical button triggering, APP button triggering, and voice command triggering, and provides three levels of feedback options: "mild discomfort / obvious discomfort / unbearable discomfort"; After receiving the feedback command, the information recording unit automatically extracts the frequency and intensity characteristics of the audio segment 1-2 seconds before and after the feedback time, and calculates the peak frequency using the following spectrum analysis formula: fpeak = argmaxk|X(k)| × (Fs / N) In the formula, X(k) is the Fourier transform result of the audio signal, k is the frequency point index, Fs is the sampling frequency, N is the Fourier transform frame length, and fpeak is the peak frequency; at the same time, the current playback mode, volume level and algorithm activation status are recorded to form a complete feedback data set.

[0010] Preferably, the machine learning adaptation module includes a data preprocessing unit, a model training unit, and a real-time optimization unit; The data preprocessing unit performs noise reduction on the feedback data group, associates and labels the feedback level, corresponding frequency band, frequency intensity and scene parameters, and constructs a sample dataset. The samples with user feedback of "discomfort" are labeled with a tolerance score of 0, and the samples with no feedback are labeled with a tolerance score of 1. The model training unit employs a lightweight random forest model, using frequency band f, frequency intensity L, and scene parameters S as input features, and tolerance score Y as the output label. During model training, the parameters are optimized using the following loss function: L=(1 / n)×Σi=1n(Yi-Ŷi)² In the formula, Yi is the actual tolerance score of the i-th sample, Ŷi is the model predicted score of the i-th sample, and n is the number of samples. The personalized hearing tolerance model is obtained by minimizing the loss function L. In the initial stage, the basic model parameters are set based on the ISO7731 hearing safety standard, and the model is automatically updated after collecting 5-10 user feedbacks. Before the audio signal enters the power amplification module, the real-time optimization unit obtains the frequency characteristics and intensity of the output signal through real-time spectrum analysis and compares it with the tolerance threshold output by the personalized hearing tolerance model. If the intensity of a certain frequency band is close to the tolerance threshold (difference < 5dB) or belongs to a sensitive frequency band, a narrowband notch filter is used to suppress the intensity of that frequency band. The suppression formula is as follows: Aout=Ain×[1-α×exp(-|f-fs| / f0)] In the formula, Aout is the frequency intensity after compression, Ain is the original frequency intensity, α is the compression coefficient with a value range of 0.6-0.9, fs is the user-sensitive frequency, f is the current frequency point, and f0 is the frequency attenuation coefficient (with a value of 50-200Hz). At the same time, the overall listening experience is balanced through gain compensation of adjacent frequency bands.

[0011] Preferably, the machine learning adaptation module further includes a continuous learning unit, which employs a forgetting mechanism to weaken sensitive labels that have received no feedback for a long time, with the weakening coefficient β decaying over time. β = β0 × exp(-t / τ) In the formula, t is the time since the last feedback (in days), β0 is the initial weakening coefficient 1, and τ is the attenuation constant (valued at 30 days). At the same time, after the model is updated, low-intensity audio clips containing the original sensitive frequency band are actively played to collect secondary user feedback to calibrate the tolerance threshold.

[0012] Preferably, the Bluetooth transmission module is compatible with Bluetooth 5.3 and above, supports the LEAudio protocol and LC3 codec, enhances anti-interference capability through CodedPHY mode, and automatically switches the transmission mode when the detected signal strength is weaker than -85dBm; the intelligent signal processing module adopts a Bluetooth audio chip with an integrated independent DSP, sets the audio processing thread to the highest priority and binds it to the independent DSP core, limits the CPU usage of background tasks to ≤10%, and configures ≥196KB RAM and 4MFlash, and integrates an FFT accelerator to improve spectrum analysis efficiency.

[0013] Preferably, the power amplification and playback module uses a Class D power amplifier chip with integrated gallium nitride, the impedance matching between the power amplifier and the speaker is 4Ω / 8Ω, and it integrates overcurrent protection function, which can withstand 3 times the rated current surge; it also includes a real-time monitoring module, which monitors DSP load, memory usage and audio distortion THD+N through the chip's built-in registers. When the DSP load is >80%, it triggers an algorithm level downgrade, prioritizing the retention of the core audio processing algorithm.

[0014] Preferably, the user feedback module is also associated with the user account, and the personalized hearing tolerance model supports synchronization across multiple devices; it also provides a visual interactive interface to display the user's sensitive frequency range and tolerance threshold, and supports the user to manually adjust the limiting parameters of each frequency band.

[0015] Beneficial effects This invention provides an intelligent audio signal processing and transmission system for Bluetooth speakers. It has the following beneficial effects: This invention provides an intelligent audio signal processing and transmission system for Bluetooth speakers. Through algorithm grading and dynamic switching, it achieves precise algorithm scheduling under different scenarios and loads. Combined with the optimization of computing power allocation of independent DSP cores, it ensures the stability of audio processing during multi-task parallel operation, avoids dynamic range reduction and high-frequency detail loss, and improves the sound quality of audio playback.

[0016] This invention provides an intelligent audio signal processing and transmission system for Bluetooth speakers. Through a closed-loop design of a user feedback module and a machine learning adaptation module, it accurately captures the auditory tolerance thresholds of different users, performs targeted suppression and gain compensation on sensitive frequency bands, and avoids auditory stimulation caused by extreme frequencies. No professional knowledge is required from the user; simple feedback triggers the system to learn autonomously and generate audio processing strategies adapted to individual listening preferences, improving the user comfort for different user groups. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the system flow of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] like Figure 1 As shown, this embodiment of the invention provides an intelligent audio signal processing and transmission system for a Bluetooth speaker, comprising: The signal acquisition module is used to acquire environmental audio signals, noise signals, and feedback signals from played audio, and output raw acquisition data containing frequency and intensity characteristics. The intelligent signal processing module integrates an algorithm hierarchical and dynamic switching unit, which is used to hierarchically schedule audio processing algorithms according to scene parameters and system load. The Bluetooth transmission module is used to enable wireless transmission of audio signals between the audio source device and the speaker, and is compatible with multiple protocol codecs. The power amplification and playback module is used to amplify the processed audio signal and play it through a speaker; The user feedback module is used to receive feedback commands triggered by users due to hearing discomfort, and synchronously record the timestamp of the feedback time, the frequency segment of the corresponding audio segment, the frequency intensity, and the current scene parameters. The machine learning adaptation module is used to train a personalized auditory tolerance model based on user feedback data, and dynamically adjust the frequency intensity threshold of the audio signal according to the model output to achieve a balance between sound quality optimization and auditory comfort. Specifically, in the above specific embodiments, each module achieves data interaction through a high-speed serial bus, with a data transmission rate of no less than 1Mbps to ensure the real-time performance of audio signal processing; the signal acquisition module adopts a 4-channel microphone array, with each microphone configured with an independent ADC conversion unit, a sampling rate set to 44.1kHz or 48kHz, and a quantization bit depth of 24bit, which can accurately capture signals in the full audio frequency band of 20Hz-20kHz, while suppressing noise floor below -80dB, providing high-quality raw data for subsequent signal processing.

[0020] The algorithm grading and dynamic switching unit divides audio processing algorithms into three levels: high-level, mid-level, and low-level. Each level corresponds to different computing power consumption and algorithm combinations. The advanced algorithm suite includes a full HRTF spatial audio algorithm, a multi-level ANC noise reduction algorithm, a 10-band EQ adjustment algorithm, and a dynamic range extension algorithm, which are suitable for quiet indoor scenes and lossless audio source playback, with computing power consumption meeting DSP load >60%; The intermediate algorithm combination includes a basic noise reduction algorithm, a 5-band EQ adjustment algorithm, and a dynamic range compression algorithm, which is suitable for commuting / outdoor scenarios and ordinary audio sources, and the computing power consumption meets 30%-60% of the DSP load. The low-level algorithm combination retains only echo cancellation, automatic gain control, and basic encoding / decoding algorithms, making it suitable for voice call / standby scenarios and low bitrate audio sources, with computing power consumption meeting the requirement of DSP load <30%. The algorithm grading and dynamic switching unit automatically triggers a smooth switching of algorithm levels by monitoring environmental noise, sound source type, system load and user behavior in real time. The switching process uses 10-20ms cross-fade-in and fade-out processing. Specifically, in the above specific embodiments, the scene determination logic of the algorithm classification and dynamic switching unit is realized through threshold comparison: environmental noise is collected by the microphone array and then the sound pressure level is calculated. When the sound pressure level is ≤45dB, it is determined to be a quiet indoor scene; when the sound pressure level is >65dB, it is determined to be a commuting / outdoor scene. The audio source type is identified by parsing the encoding format and bitrate of the audio file. FLAC and WAV formats with a bitrate ≥1411kbps are determined to be lossless audio sources, while MP3 and AAC formats with a bitrate ≤320kbps are determined to be ordinary audio sources. The system load is read in real time through the running status register of the DSP core, with a sampling period of 10ms. The cross-fade-in and fade-out processing adopts a linear superposition algorithm. The signal amplitude of the previous algorithm combination linearly decays to 0 with a period of 10-20ms, and the signal amplitude of the next algorithm combination synchronously and linearly increases to the rated value to avoid audio abrupt changes or popping during the switching process.

[0021] The user feedback module includes a triggering unit and an information recording unit; the triggering unit supports three methods: physical button triggering, APP button triggering, and voice command triggering, and provides three levels of feedback options: "mild discomfort / obvious discomfort / unbearable discomfort"; After receiving the feedback command, the information recording unit automatically extracts the frequency and intensity characteristics of the audio segment 1-2 seconds before and after the feedback time, and calculates the peak frequency using the following spectrum analysis formula: fpeak = argmaxk|X(k)| × (Fs / N) In the formula, X(k) is the Fourier transform result of the audio signal, k is the frequency point index, Fs is the sampling frequency, N is the Fourier transform frame length, and fpeak is the peak frequency; at the same time, the current playback mode, volume level and algorithm activation status are recorded to form a complete feedback data set. Specifically, in the above embodiments, the physical button trigger adopts a dual redundancy design of mechanical buttons and touch buttons, with a press response time of ≤50ms; the voice command trigger supports offline wake-up, the wake-up word can be customized, the recognition accuracy is ≥95%, and recognition is only initiated when the user's voice energy is detected to be ≥60dB to avoid false triggering. The Fourier transform frame length N of the information recording unit is set to 2048 points, and the frequency resolution of the spectrum analysis is ≤23Hz, which can accurately locate sensitive frequency bands; the feedback data group is encapsulated in JSON format, including fields such as feedback level, timestamp (accurate to milliseconds), peak frequency, frequency intensity (unit dBSPL), playback mode, volume level (0-100 levels), and algorithm level, etc., and is temporarily stored in local Flash. After accumulating 3 feedback data, they are batch transmitted to the machine learning adaptation module.

[0022] The machine learning adaptation module includes a data preprocessing unit, a model training unit, and a real-time optimization unit; The data preprocessing unit performs noise reduction on the feedback data set, associates and labels the feedback level, corresponding frequency band, frequency intensity and scene parameters, and constructs a sample dataset. Among them, the samples with user feedback of "discomfort" are labeled with a tolerance score of 0, and the samples with no feedback are labeled with a tolerance score of 1. The model training unit uses a lightweight random forest model, with frequency band f, frequency intensity L, and scene parameters S as input features, and tolerance score Y as the output label. During model training, the parameters are optimized using the following loss function: L=(1 / n)×Σi=1n(Yi-Ŷi)² In the formula, Yi is the actual tolerance score of the i-th sample, Ŷi is the model predicted score of the i-th sample, and n is the number of samples. The personalized hearing tolerance model is obtained by minimizing the loss function L. In the initial stage, the basic model parameters are set based on the ISO7731 hearing safety standard, and the model is automatically updated after collecting 5-10 user feedbacks. Before the audio signal enters the power amplification module, the real-time optimization unit obtains the frequency characteristics and intensity of the output signal through real-time spectrum analysis. This is compared with the tolerance threshold output by the personalized hearing tolerance model. If the intensity of a certain frequency band is close to the tolerance threshold (difference < 5dB) or belongs to a sensitive frequency band, a narrowband notch filter is used to suppress the intensity of that frequency band. The suppression formula is as follows: Aout=Ain×[1-α×exp(-|f-fs| / f0)] In the formula, Aout is the frequency intensity after compression, Ain is the original frequency intensity, α is the compression coefficient with a value range of 0.6-0.9, fs is the user-sensitive frequency, f is the current frequency point, and f0 is the frequency attenuation coefficient (with a value of 50-200Hz). At the same time, the overall listening experience is balanced through gain compensation of adjacent frequency bands.

[0023] The machine learning adaptation module also includes a continuous learning unit, which uses a forgetting mechanism to weaken sensitive labels that have not received feedback for a long time. The weakening coefficient β decays over time. β = β0 × exp(-t / τ) In the formula, t is the time since the last feedback (unit: days), β0 is the initial weakening coefficient 1, and τ is the attenuation constant (value is 30 days); at the same time, after the model is updated, low-intensity audio clips containing the original sensitive frequency band are actively played to collect secondary user feedback to calibrate the tolerance threshold. Specifically, in the above embodiments, the noise reduction processing of the data preprocessing unit adopts a median filtering algorithm with a window size of 5, which can effectively remove outliers in the feedback data; the sample dataset is divided into training and validation sets in an 8:2 ratio, and 5-fold cross-validation is used during model training to improve generalization ability. The spectrum analysis and intelligent signal processing modules of the real-time optimization unit share the FFT accelerator, and the analysis period is consistent with the audio frame period (20ms / frame); the gain compensation amplitude of adjacent frequency bands is ≤3dB, and the compensation frequency band is ±50-200Hz of the sensitive frequency band to avoid hollow sound caused by suppression of a single frequency band. The active verification process of the continuous learning unit is only executed when the device is in standby mode and the battery is ≥30%. The intensity of the sensitive frequency segment played is 10dB lower than the original tolerance threshold, and the playback duration is 3-5 seconds. After the verification is completed, the device automatically returns to standby mode without affecting normal user use.

[0024] The Bluetooth transmission module is compatible with Bluetooth 5.3 and above, supports the LEAudio protocol and LC3 codec, and enhances anti-interference capabilities through CodedPHY mode. It automatically switches the transmission mode when the detected signal strength is weaker than -85dBm. The intelligent signal processing module uses a Bluetooth audio chip with an integrated independent DSP, sets the audio processing thread to the highest priority and binds it to the independent DSP core, limits the CPU usage of background tasks to ≤10%, and is equipped with ≥196KB RAM and 4MFlash. It also integrates an FFT accelerator to improve spectrum analysis efficiency. Specifically, in the above embodiments, the Bluetooth transmission module's transmit power is adjustable from 0-10dBm, and its receive sensitivity is ≤-95dBm. Supported codec protocols include SBC, AAC, aptXHD, LDAC, and LC3. It can automatically negotiate the optimal encoding method based on the protocol support of the audio source device. The LC3 codec bitrate can be dynamically adjusted within the range of 34-192kbps, reducing transmission power consumption while ensuring sound quality. The intelligent signal processing module uses the Qualcomm QCC5125 chip, whose built-in KalimbaDSP core has a maximum clock speed of 288MHz. 196KB RAM is used to store real-time processed audio data, and 4MFlash is used to store algorithm programs and configuration parameters. The FFT accelerator supports 1024-point and 2048-point real-number FFT operations with a computation latency of ≤1ms, significantly improving the efficiency of spectrum analysis.

[0025] The power amplification and playback module uses a Class D power amplifier chip with integrated gallium nitride. The impedance matching between the power amplifier and the speaker is 4Ω / 8Ω, and it has integrated overcurrent protection function, which can withstand 3 times the rated current surge. It also includes a real-time monitoring module, which monitors DSP load, memory usage and audio distortion THD+N through the chip's built-in registers. When the DSP load is >80%, it triggers the algorithm level downgrade, giving priority to retaining the core audio processing algorithm. Specifically, in the above embodiment, the power amplification module uses the ACM8815 gallium nitride Class D power amplifier chip, which can achieve an output power of 200W under a 4Ω load, with an efficiency of ≥90% and THD+N ≤0.05%. The overcurrent protection function is implemented by detecting the current signal at the power amplifier output terminal. When the detected current exceeds 3 times the rated current, the current limiting protection mechanism is triggered, with a response time ≤10μs. The real-time monitoring module has a monitoring cycle of 5ms, with DSP load monitoring accuracy ≤1%, memory usage monitoring accuracy ≤0.5%, and audio distortion THD+N monitoring range of 0.01%-10%. When the DSP load >80% continuously for more than 3 monitoring cycles, the algorithm is degraded, prioritizing the shutdown of non-core sound field optimization algorithms while retaining noise reduction and basic EQ adjustment algorithms.

[0026] The user feedback module is also linked to the user account, and the personalized hearing tolerance model supports synchronization across multiple devices. It also provides a visual interactive interface to display the user's sensitive frequency range and tolerance threshold, and supports users to manually adjust the limiting parameters of each frequency band. Specifically, in the above embodiments, user accounts are associated with the cloud server via Bluetooth or Wi-Fi. The personalized hearing tolerance model is stored in the cloud in an encrypted format with an AES-256 encryption level to ensure user data security. Multi-device synchronization is achieved through cloud push. When a user logs into their account on a new device, the cloud automatically pushes the model data to the new device, with a synchronization completion time of ≤30 seconds. A visual interactive interface is deployed in the accompanying APP, using a spectrum diagram to intuitively display the tolerance threshold of the 20Hz-20kHz frequency band. Users can manually adjust the maximum output intensity of each frequency band by dragging the threshold curve, with an adjustment step of 0.5dB. The manually adjusted parameters have higher priority than the parameters automatically generated by the model, satisfying the user's personalized listening preferences.

[0027] Working principle: This system achieves efficient processing and personalized transmission and playback of audio signals through the coordinated operation of signal acquisition, intelligent processing, Bluetooth transmission, power amplification, playback output, user feedback and machine learning adaptation.

[0028] The signal acquisition module collects ambient noise signals and audio playback feedback signals through a microphone array. It also receives audio signals sent by the audio source device via a Bluetooth transmission module. After converting the acquired analog signals into digital signals, it outputs raw data containing frequency and intensity characteristics to the intelligent signal processing module.

[0029] The algorithm grading and dynamic switching unit within the intelligent signal processing module receives environmental noise data and audio source type information from the signal acquisition module in real time, as well as system status data such as DSP load and memory usage from the real-time monitoring module. It then determines the current application scenario and system load level, selecting a matching algorithm combination for audio processing. When playing lossless audio sources in a quiet indoor environment, a high-level algorithm combination is activated; when playing ordinary quality audio sources in outdoor commuting scenarios, it switches to a mid-level algorithm combination; and during voice calls or standby mode, a low-level algorithm combination is used. The switching process utilizes cross-fade-in and cross-fade-out processing to ensure audio continuity.

[0030] When a user experiences auditory discomfort, they can trigger feedback via the physical button on the feedback module, the app, or a voice command. The feedback module simultaneously records the timestamp of the feedback moment, the peak frequency and frequency intensity of the corresponding audio segment, and current scene parameters, transmitting the complete feedback data set to the machine learning adaptation module. The machine learning adaptation module preprocesses the feedback data, labels tolerance scores to construct a sample dataset, trains a personalized auditory tolerance model using a random forest model, and optimizes the model parameters using a loss function.

[0031] After completing basic audio processing, the intelligent signal processing module transmits the signal to the real-time optimization unit of the machine learning adaptation module. The real-time optimization unit obtains the frequency characteristics and intensity of the signal to be output through spectrum analysis, compares it with the tolerance threshold output by the personalized hearing tolerance model, uses narrowband notch filtering to suppress the intensity of frequency bands that are close to the threshold or sensitive, and ensures a balanced listening experience through gain compensation of adjacent frequency bands.

[0032] The personalized optimized audio signal is transmitted to the power amplification module, where it is amplified by a gallium nitride Class D power amplifier chip and played back through impedance-matched speakers. The real-time monitoring module continuously monitors parameters such as DSP load and audio distortion. When the DSP load exceeds a threshold, it triggers an algorithm grading and dynamic switching unit to downgrade the algorithm level. The machine learning adaptation module's continuous learning unit weakens sensitive markers that have received no feedback for a long time through a forgetting mechanism and actively plays low-intensity sensitive frequency segments to collect secondary feedback, enabling iterative updates to the model.

[0033] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A smart audio signal processing and transmission system for a Bluetooth speaker, characterized in that, include: The signal acquisition module is used to acquire environmental audio signals, noise signals, and feedback signals from played audio, and output raw acquisition data containing frequency and intensity characteristics. The intelligent signal processing module integrates an algorithm hierarchical and dynamic switching unit, which is used to hierarchically schedule audio processing algorithms according to scene parameters and system load. The Bluetooth transmission module is used to enable wireless transmission of audio signals between the audio source device and the speaker, and is compatible with multiple protocol codecs. The power amplification and playback module is used to amplify the processed audio signal and play it through a speaker; The user feedback module is used to receive feedback commands triggered by users due to hearing discomfort, and synchronously record the timestamp of the feedback time, the frequency segment of the corresponding audio segment, the frequency intensity, and the current scene parameters. The machine learning adaptation module is used to train a personalized hearing tolerance model based on user feedback data. It dynamically adjusts the frequency intensity threshold of the audio signal according to the model output to achieve a balance between sound quality optimization and hearing comfort.

2. The intelligent audio signal processing and transmission system for a Bluetooth speaker according to claim 1, characterized in that: The algorithm grading and dynamic switching unit divides the audio processing algorithm into three levels: high-level, medium-level, and low-level. Each level corresponds to different computing power consumption and algorithm combinations, wherein: The advanced algorithm suite includes a full HRTF spatial audio algorithm, a multi-level ANC noise reduction algorithm, a 10-band EQ adjustment algorithm, and a dynamic range extension algorithm, which are suitable for quiet indoor scenes and lossless audio source playback, with computing power consumption meeting DSP load >60%; The intermediate algorithm combination includes a basic noise reduction algorithm, a 5-band EQ adjustment algorithm, and a dynamic range compression algorithm, which is suitable for commuting / outdoor scenarios and ordinary audio sources, and the computing power consumption meets 30%-60% of the DSP load. The low-level algorithm combination retains only echo cancellation, automatic gain control, and basic encoding / decoding algorithms, making it suitable for voice call / standby scenarios and low bitrate audio sources, with computing power consumption meeting the requirement of DSP load <30%. The algorithm grading and dynamic switching unit automatically triggers a smooth switching of algorithm levels by monitoring environmental noise, sound source type, system load and user behavior in real time. The switching process uses 10-20ms cross fade-in and fade-out processing.

3. The intelligent audio signal processing and transmission system for a Bluetooth speaker according to claim 1, characterized in that: The user feedback module includes a triggering unit and an information recording unit; the triggering unit supports three methods: physical button triggering, APP button triggering, and voice command triggering, and provides three levels of feedback options: "mild discomfort / obvious discomfort / unbearable discomfort"; After receiving the feedback command, the information recording unit automatically extracts the frequency and intensity characteristics of the audio segment 1-2 seconds before and after the feedback time, and calculates the peak frequency using the following spectrum analysis formula: fpeak = argmaxk|X(k)| × (Fs / N) In the formula, X(k) is the Fourier transform result of the audio signal, k is the frequency point index, Fs is the sampling frequency, N is the Fourier transform frame length, and fpeak is the peak frequency; at the same time, the current playback mode, volume level and algorithm activation status are recorded to form a complete feedback data set.

4. The intelligent audio signal processing and transmission system for a Bluetooth speaker according to claim 1, characterized in that: The machine learning adaptation module includes a data preprocessing unit, a model training unit, and a real-time optimization unit. The data preprocessing unit performs noise reduction on the feedback data group, associates and labels the feedback level, corresponding frequency band, frequency intensity and scene parameters, and constructs a sample dataset. The samples with user feedback of "discomfort" are labeled with a tolerance score of 0, and the samples with no feedback are labeled with a tolerance score of 1. The model training unit employs a lightweight random forest model, using frequency band f, frequency intensity L, and scene parameters S as input features, and tolerance score Y as the output label. During model training, the parameters are optimized using the following loss function: L=(1 / n)×Σi=1n(Yi-Ŷi)² In the formula, Yi is the actual tolerance score of the i-th sample, Ŷi is the model predicted score of the i-th sample, and n is the number of samples. The personalized hearing tolerance model is obtained by minimizing the loss function L. In the initial stage, the basic model parameters are set based on the ISO7731 hearing safety standard, and the model is automatically updated after collecting 5-10 user feedbacks. Before the audio signal enters the power amplification module, the real-time optimization unit obtains the frequency characteristics and intensity of the output signal through real-time spectrum analysis and compares it with the tolerance threshold output by the personalized hearing tolerance model. If the intensity of a certain frequency band is close to the tolerance threshold (difference < 5dB) or belongs to a sensitive frequency band, a narrowband notch filter is used to suppress the intensity of that frequency band. The suppression formula is as follows: Aout=Ain×[1-α×exp(-|f-fs| / f0)] In the formula, Aout is the frequency intensity after compression, Ain is the original frequency intensity, α is the compression coefficient with a value range of 0.6-0.9, fs is the user-sensitive frequency, f is the current frequency point, and f0 is the frequency attenuation coefficient (with a value of 50-200Hz). At the same time, the overall listening experience is balanced through gain compensation of adjacent frequency bands.

5. The intelligent audio signal processing and transmission system for a Bluetooth speaker according to claim 4, characterized in that: The machine learning adaptation module also includes a continuous learning unit, which employs a forgetting mechanism to weaken sensitive labels that have received no feedback for a long time, with the weakening coefficient β decaying over time. β = β0 × exp(-t / τ) In the formula, t is the time since the last feedback (in days), β0 is the initial weakening coefficient 1, and τ is the attenuation constant (valued at 30 days). At the same time, after the model is updated, low-intensity audio clips containing the original sensitive frequency band are actively played to collect secondary user feedback to calibrate the tolerance threshold.

6. The intelligent audio signal processing and transmission system for a Bluetooth speaker according to claim 1, characterized in that: The Bluetooth transmission module is compatible with Bluetooth 5.3 and above, supports the LEAudio protocol and LC3 codec, and enhances anti-interference capabilities through CodedPHY mode. It automatically switches the transmission mode when the detected signal strength is weaker than -85dBm. The intelligent signal processing module uses a Bluetooth audio chip with an integrated independent DSP, sets the audio processing thread to the highest priority and binds it to the independent DSP core, limits the CPU usage of background tasks to ≤10%, and configures ≥196KB RAM and 4MFlash. It also integrates an FFT accelerator to improve spectrum analysis efficiency.

7. The intelligent audio signal processing and transmission system for a Bluetooth speaker according to claim 1, characterized in that: The power amplification and playback module uses a Class D power amplifier chip with integrated gallium nitride. The impedance matching between the power amplifier and the speaker is 4Ω / 8Ω, and it has integrated overcurrent protection, which can withstand three times the rated current surge. It also includes a real-time monitoring module, which monitors DSP load, memory usage, and audio distortion THD+N through the chip's built-in registers. When the DSP load is >80%, it triggers an algorithm level downgrade, prioritizing the retention of the core audio processing algorithm.

8. The intelligent audio signal processing and transmission system for a Bluetooth speaker according to claim 1, characterized in that: The user feedback module is also linked to the user account, and the personalized hearing tolerance model supports synchronization across multiple devices. It also provides a visual interactive interface that displays the user's sensitive frequency ranges and tolerance thresholds, and allows users to manually adjust the limiting parameters for each frequency band.

Images