A noise reduction control system and method for a bluetooth headset

Abstract

The application provides a noise reduction control system and method for a Bluetooth headset, the noise reduction control system for the Bluetooth headset comprising: a server, a control processing module and a communication module; the control processing module is used for generating a first ambient noise identification model and a first noise elimination model; when the Bluetooth headset is working, the current ambient noise type is determined according to the first noise identification model; and the optimal anti-noise is generated by selecting a matched noise elimination model from the first noise elimination model according to the current ambient noise type and played. Through the application, the noise type in the current environment can be intelligently and efficiently identified, and the corresponding noise elimination model can be flexibly and accurately selected for noise reduction.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of communication technology, in particular to a noise reduction control system and method for a Bluetooth earphone. BACKGROUND

[0002] Bluetooth is a low-cost, high-capacity short-range wireless communication specification, and Bluetooth earphones are earphones that apply Bluetooth technology to hands-free earphones, allowing users to be free from annoying connection lines and achieve free and flexible communication. In real life, there are many scenes where the environment noise is very chaotic and loud when using Bluetooth earphones, which seriously affects the user experience of Bluetooth earphones, and the existing noise reduction scheme for Bluetooth earphones cannot flexibly and intelligently control the noise reduction. SUMMARY

[0003] The present application is based on the above problems, and proposes a noise reduction control system and method for a Bluetooth earphone, which can intelligently and efficiently identify the noise type in the current environment, and flexibly and accurately select the corresponding noise cancellation model for noise reduction.

[0004] Therefore, one aspect of the present application proposes a noise reduction control system for a Bluetooth earphone, comprising:

[0005] a server, a control processing module, and a communication module;

[0006] The control processing module is configured to:

[0007] obtain historical activity data and earphone historical use data of a first user of the Bluetooth earphone from the server through the communication module;

[0008] obtain historical location data and historical behavior data recorded based on time from the historical activity data;

[0009] determine a plurality of first places according to the historical location data;

[0010] obtain first environmental noise data corresponding to different first places, and extract first spectral features, first time domain features, and first spatial features of various first environmental noise data;

[0011] train a first neural network using the first spectral features, the first time domain features, and the first spatial features to obtain a first environmental noise identification model capable of identifying different types of environmental noise;

[0012] obtain first head accompaniment data, and obtain a first noise cancellation model according to the first environmental noise data and the first head accompaniment data;

[0013] integrate the trained first noise recognition model and the first noise cancellation model into the Bluetooth earphone;

[0014] When the Bluetooth earphone is working, determine the current environmental noise type according to the first noise recognition model;

[0015] According to the current environmental noise type, select a matched noise cancellation model from the first noise cancellation model to generate an optimal anti-noise for playing.

[0016] Optionally, in the step of obtaining the first head accompaniment data and obtaining the first noise cancellation model according to the first environmental noise data and the first head accompaniment data, the control processing module is configured to:

[0017] According to the first user's ear canal data, construct an artificial ear canal simulator;

[0018] According to the first environmental noise data, the earphone historical use data and the historical behavior data, generate first simulation sound data;

[0019] Input the first simulation sound data into the artificial ear canal simulator to obtain the first head accompaniment data;

[0020] Construct first scene sound data mixed with different types of first environmental noise data and first target sound data;

[0021] Construct first anti-noise data using the first head accompaniment data;

[0022] Use the first scene sound data and the first anti-noise data to train a second neural network to obtain a first noise cancellation model.

[0023] Optionally, in the step of constructing first scene sound data mixed with different types of first environmental noise data and first target sound data, the control processing module is configured to:

[0024] Obtain the first spectral feature, the first time domain feature and the first space domain feature of the first environmental noise data;

[0025] Collect speech samples, determine a first vocabulary and a first language from the speech samples according to a first preset rule; obtain the first target sound data from real person pronunciation or text speech synthesis according to the first vocabulary and the first language, and perform feature extraction on the first target sound data to obtain second spectral feature, second time domain feature and second space domain feature;

[0026] According to different scene requirements, the first frequency spectrum feature, the first time domain feature and the first space domain feature, and the second frequency spectrum feature, the second time domain feature and the second space domain feature are used to determine different types of second environmental noise data and second target sound data corresponding to the different scene requirements respectively.

[0027] The second environmental noise data of different types and the second target sound data are mixed in a corresponding proportion to obtain the first scene sound data corresponding to different scenes.

[0028] Optionally, in the step of constructing the first anti-noise data by using the first head accompaniment sound data, the control processing module is configured to:

[0029] The mixed first scene sound data is sampled, features are extracted, and corresponding anti-noise sample data is generated according to a preset anti-noise algorithm;

[0030] The first scene sound data and the anti-noise sample data are organized into a training data set for training a deep learning model to obtain an anti-noise generation rule;

[0031] The first anti-noise data is generated according to the anti-noise generation rule, the first scene sound data and the first head accompaniment sound data.

[0032] Optionally, a 3-axis or 6-axis digital gyroscope is configured in the main control chip of the Bluetooth earphone; two or more microphones are respectively configured on the two ear shells of the Bluetooth earphone; a plurality of controllable small-sized loudspeakers are integrated between the earplugs of the Bluetooth earphone; and the control processing module is configured to:

[0033] The motion trajectory data, rotation speed data and direction change data of the head of the first user are detected by the digital gyroscope;

[0034] The data of each microphone is calibrated, and the phase and amplitude difference of the current environmental noise is detected;

[0035] The head rotation direction and head rotation amplitude are determined by using a preset head motion detection algorithm in combination with the motion trajectory data, the rotation speed data and the direction change data;

[0036] The source direction of the current environmental noise is determined by using a preset noise source positioning algorithm in combination with the phase and amplitude difference;

[0037] According to the head rotation direction, the head rotation amplitude and the source direction, the phase and energy distribution of the anti-noise between the two ears are changed correspondingly by using a preset directional anti-noise generation model, to generate directional first anti-noise to offset the current environmental noise.

[0038] Another aspect of the present application provides a noise reduction control method for a Bluetooth headset, the noise reduction control method for a Bluetooth headset comprising:

[0039] obtaining historical activity data and headset historical use data of a first user of a Bluetooth headset;

[0040] obtaining time-based historical location data and historical behavior data from the historical activity data;

[0041] determining a plurality of first places according to the historical location data;

[0042] obtaining first environmental noise data corresponding to different first places, and extracting first spectral features, first time-domain features, and first spatial features of each type of first environmental noise data;

[0043] training a first neural network using the first spectral features, the first time-domain features, and the first spatial features to obtain a first environmental noise recognition model capable of recognizing different types of environmental noise;

[0044] obtaining first head accompaniment data, and obtaining a first noise cancellation model according to the first environmental noise data and the first head accompaniment data;

[0045] integrating the trained first noise recognition model and the first noise cancellation model into the Bluetooth headset;

[0046] when the Bluetooth headset is working, determining a current environmental noise type according to the first noise recognition model;

[0047] selecting a matching noise cancellation model from the first noise cancellation model according to the current environmental noise type to generate optimal anti-noise for playback.

[0048] Optionally, the step of obtaining first head accompaniment data and obtaining a first noise cancellation model according to the first environmental noise data and the first head accompaniment data comprises:

[0049] constructing an artificial ear canal simulator according to ear canal data of the first user;

[0050] generating first simulated sound data according to the first environmental noise data, the headset historical use data, and the historical behavior data;

[0051] inputting the first simulated sound data into the artificial ear canal simulator to obtain the first head accompaniment data;

[0052] constructing first scene sound data mixed with different types of first environmental noise data and first target sound data;

[0053] construct first anti-noise data by using the first head accompaniment data;

[0054] train a second neural network by using the first scene sound data and the first anti-noise data to obtain a first noise cancellation model.

[0055] Optionally, the step of constructing the first scene sound data mixed with the first environmental noise data and the first target sound data of different types comprises:

[0056] obtaining the first spectral feature, the first time domain feature and the first space domain feature of the first environmental noise data;

[0057] collecting speech samples, determining a first vocabulary and a first language from the speech samples according to a first preset rule, obtaining the first target sound data from a real person pronunciation or text-to-speech according to the first vocabulary and the first language, and performing feature extraction on the first target sound data to obtain a second spectral feature, a second time domain feature and a second space domain feature;

[0058] determining second environmental noise data and second target sound data of different types corresponding to different scene requirements by using the first spectral feature, the first time domain feature and the first space domain feature and the second spectral feature, the second time domain feature and the second space domain feature respectively according to the different scene requirements;

[0059] mixing the second environmental noise data and the second target sound data of different types according to corresponding proportions to obtain corresponding first scene sound data under different scenes.

[0060] Optionally, the step of constructing the first anti-noise data by using the first head accompaniment data comprises:

[0061] sampling the mixed first scene sound data, extracting features and generating corresponding anti-noise sample data according to a preset anti-noise algorithm;

[0062] organizing the first scene sound data and the anti-noise sample data into a training data set for training a deep learning model to obtain an anti-noise generation rule;

[0063] generating the first anti-noise data according to the anti-noise generation rule, the first scene sound data and the first head accompaniment data.

[0064] Optionally, a 3-axis or 6-axis digital gyroscope is configured in the master chip of the Bluetooth earphone; two or more microphones are configured on the two ear shells of the Bluetooth earphone; a plurality of controllable small speakers are integrated between the earbuds of the Bluetooth earphone; and the noise reduction control method for the Bluetooth earphone further comprises:

[0065] The motion trajectory data, rotation speed data and direction change data of the head of the first user are detected by the digital gyroscope;

[0066] The data of each microphone is calibrated, and the phase and amplitude difference of the current environmental noise is detected;

[0067] The head rotation direction and head rotation amplitude are determined by using a preset head motion detection algorithm in combination with the motion trajectory data, rotation speed data and direction change data;

[0068] The source direction of the current environmental noise is determined by using a preset noise source positioning algorithm in combination with the phase and amplitude difference;

[0069] The phase and energy distribution of the anti-noise between the two ears are changed accordingly according to the head rotation direction, head rotation amplitude and source direction by using a preset directional anti-noise generation model, and directional first anti-noise is generated to offset the current environmental noise.

[0070] The technical scheme of the application comprises: obtaining the historical activity data and earphone historical use data of the first user of the Bluetooth earphone; obtaining the historical location data and historical behavior data recorded based on time from the historical activity data; determining a plurality of first places according to the historical location data; obtaining first environmental noise data corresponding to different first places, and extracting first spectral features, first time domain features and first space domain features of various first environmental noise data; training a first neural network by using the first spectral features, first time domain features and first space domain features to obtain a first environmental noise recognition model capable of recognizing different types of environmental noise; obtaining first head surround sound data, and obtaining a first noise cancellation model according to the first environmental noise data and the first head surround sound data; integrating the trained first noise recognition model and the first noise cancellation model into the Bluetooth earphone; when the Bluetooth earphone is working, determining the current environmental noise type according to the first noise recognition model; selecting a matched noise cancellation model from the first noise cancellation model to generate optimal anti-noise for playing according to the current environmental noise type. The scheme of the application can not only intelligently and efficiently recognize the noise type in the current environment, but also flexibly and accurately select the corresponding noise cancellation model for noise reduction. BRIEF DESCRIPTION OF DRAWINGS

[0071] Figure 1 is a schematic block diagram of a noise control system for a Bluetooth headset provided by an embodiment of the present application;

[0072] Figure 2 is a flow chart of a noise control method for a Bluetooth headset provided by an embodiment of the present application. DETAILED DESCRIPTION

[0073] In order to more clearly understand the above objectives, features and advantages of the present application, further specific details of the present application will be given with reference to the accompanying drawings and specific embodiments. It should be noted that the embodiments of the present application and the features in the embodiments can be combined with each other without conflict.

[0074] In the following description, a large number of specific details are set forth in order to facilitate a thorough understanding of the present application, however, the present application can also be implemented in other ways different from those described herein, and therefore, the scope of protection of the present application is not limited by the specific embodiments disclosed below.

[0075] The terms "first", "second", and the like in the specification and claims of the present application and the above-described drawings are used to distinguish different objects, and are not used to describe a specific order. In addition, the terms "include" and "have" and any variations thereof are intended to cover non-exclusive inclusion. For example, a process, method, system, product or device including a series of steps or units is not limited to the listed steps or units, but can optionally include steps or units not listed, or can optionally include other steps or units inherent to the process, method, product or device.

[0076] Reference herein to "an embodiment" means that a particular feature, structure, or characteristic described in connection with the embodiment can be included in at least one embodiment of the present application. The appearance of the phrase in various places in the specification does not necessarily all refer to the same embodiment, nor is it necessarily independent or alternative embodiments to each other. It is explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0077] The noise control system and method for a Bluetooth headset according to some embodiments of the present application will be described below with reference to Figures 1 to 2 .

[0078] As shown in Figure 1 , an embodiment of the present application provides a noise control system for a Bluetooth headset, comprising: a server, a control processing module, a communication module;

[0079] The control processing module is configured to:

[0080] obtaining, by the communication module, historical activity data of a first user of a Bluetooth earphone and earphone historical use data from the server;

[0081] obtaining historical location data and historical behavior data recorded based on time from the historical activity data;

[0082] determining a plurality of first places according to the historical location data;

[0083] obtaining first environmental noise data (including but not limited to airport noise, street noise, coffee shop noise, office noise, etc.) corresponding to different first places, and extracting first spectral features, first time domain features and first spatial domain features of various first environmental noise data;

[0084] training a first neural network by using the first spectral features, the first time domain features and the first spatial domain features to obtain a first environmental noise recognition model capable of recognizing different types of environmental noise;

[0085] obtaining first head accompaniment data, and obtaining a first noise cancellation model according to the first environmental noise data and the first head accompaniment data (to generate a head accompaniment closer to the real anti-noise, and achieve a more personalized noise reduction effect);

[0086] integrating the trained first noise recognition model and the first noise cancellation model into the Bluetooth earphone;

[0087] when the Bluetooth earphone is working, determining a current environmental noise type according to the first noise recognition model;

[0088] generating an optimal anti-noise from the first noise cancellation model according to the current environmental noise type to play. That is, generating a personalized anti-noise according to the user's head accompaniment data to play, and continuously optimizing the model according to the use feedback.

[0089] The technical solution of the embodiment can not only intelligently and efficiently recognize the noise type in the current environment, but also flexibly and accurately select the corresponding noise cancellation model for noise reduction.

[0090] It should be understood that, Figure 1 The block diagram of the noise reduction control system for the Bluetooth earphone shown is only illustrative, and the number of modules shown does not limit the protection scope of the present application.

[0091] In the embodiments of the present application, the head-related impulse response (HRIR) refers to the response data of the human head to sound signals, which contains information about the acoustic characteristics and the structure of the human head. When environmental noise enters the ear canal of a person, it will produce complex multiple reflections, attenuation and diffraction in the head, and finally reach the eardrum and cochlea. The HRIR data records all the filtering and deformation of the head during the process of environmental sound entering the ear canal and reaching the eardrum. Specifically, the HRIR data records: 1) the acoustic transmission function of the external auditory canal, such as the attenuation and phase change of the external auditory canal to different frequency noises, etc. 2) the acoustic structure of the skull, such as the sound diffraction caused by the density and shape of the skull, the acoustic effects of various bone joints and cartilages, etc. 3) the effect of head axis and rotation on sound. The path of sound entering and exiting the ear canal is different when the head direction is different, resulting in differences in HRIR. Therefore, the HRIR data contains unique acoustic structure and acoustic channel characteristics of the human head, and using the HRIR data can perform personalized transformation on any noise signal to simulate the sound effect of real noise entering the head. This lays the foundation for generating personalized anti-noise and realizing high-precision active noise control and elimination technology. In the noise reduction application of Bluetooth earphones, collecting and using the HRIR data of users can generate more realistic and head-acoustic-structure-fitting anti-noise, thereby achieving the purpose of improving individual differences and improving noise reduction accuracy.

[0092] The scheme of the present application can be personalized and optimized in combination with user personal information and use scene habits, and realize more intelligent artificial noise elimination. The key of the scheme of the present application is to obtain sufficient feature data to train the recognition model and the generation model, so as to realize high-precision noise classification and matching noise elimination, thereby making the noise reduction effect of Bluetooth earphones more intelligent and personalized. This requires extensive data collection and powerful AI computing capability.

[0093] In some possible embodiments of the present application, a user self-help head-related sound collection tool can be provided to upload the cloud to continuously enrich the head-related sound database, realize user participation and cooperation, and comprehensively collect more detailed head-related sound information, and generate more accurate and personalized anti-noise. This is a scheme that requires technical innovation and breakthrough in all aspects, including data collection equipment, large-scale data set construction, powerful AI training computing platform, precise model training and deployment, etc.

[0094] In some possible embodiments of the present application, an end-to-end noise reduction model based on deep learning is developed, which directly maps environmental noise to anti-noise without manually extracting noise features and designing noise elimination algorithms, so that the AI model automatically learns features and algorithms, and a more intelligent noise reduction scheme can be realized.

[0095] It can be understood that, in order to obtain accurate anti-noise data to establish a more accurate noise cancellation model, in some possible embodiments of the present application, in the step of obtaining first head companion sound data and obtaining a first noise cancellation model according to the first environmental noise data and the first head companion sound data, the control processing module is configured to:

[0096] A large amount of environmental noise data of various types is collected by designing a precise noise collection device, a large number of sample data covering as many scenarios as possible are obtained, and the data classification and characteristics are labeled. User head companion sound data, including ear canal entrance impulse response data, etc., is obtained.

[0097] An artificial ear canal simulator is constructed according to the ear canal data of the first user. For example, a three-dimensional model of the ear canal can be generated according to image data, structural data, etc. of the ear canal of the first user, and corresponding ear canal physiological parameters, physical parameters, biological response parameters (which can be determined according to test data or historical behavior data of the user combined with relevant data / medical data of other users) are added based on the three-dimensional model to establish a virtual artificial ear canal simulator.

[0098] First simulated sound data is generated according to the first environmental noise data, the earphone historical use data and the historical behavior data. For example, corresponding historical playback data, historical scene data, historical noise processing data of the user, etc. can be determined according to the earphone historical use data and the historical behavior data, and then combined with the first environmental noise data, sound data meeting the requirements can be selected to be fused into the first simulated sound data.

[0099] The first simulated sound data is input into the artificial ear canal simulator to obtain the first head companion sound data.

[0100] First scene sound data mixed with the first environmental noise data and the first target sound data of different types is constructed.

[0101] First anti-noise data is constructed using the first head companion sound data (for example, but not limited to, an artificial head companion sound data is synthesized by training a generative adversarial network (GAN) using the first environmental noise data and the first head companion sound data, and the network is iterated by comparing the artificial head companion sound data with the first head companion sound data to make the artificial head companion sound data more realistic; a deep learning model is trained on a large-scale artificial / original head companion sound data set to learn the mapping relationship model of the head companion sound feature and the optimal anti-noise. The model needs to consider the head shape, sound channel structure, noise characteristics, etc. to generate personalized anti-noise, and obtain the first anti-noise data).

[0102] The first noise cancellation model is obtained by training a second neural network using the first scene sound data and the first anti-noise data (which needs to consider the frequency domain, time domain and spatial features of human voice and noise for comparison, and accurately extract the noise part).

[0103] In order to accurately determine the relationship between the scene and the environmental noise to obtain accurate anti-noise data, in some possible embodiments of the present application, in the step of constructing the first scene sound data mixed with the first environmental noise data and the first target sound data of different types, the control processing module is configured to:

[0104] Obtain the first spectral feature, the first time domain feature and the first spatial feature of the first environmental noise data;

[0105] Collect a voice sample, and determine a first vocabulary and a first language from the voice sample according to a first preset rule;

[0106] Obtain the first target sound data from a real person pronunciation or a text-to-speech synthesis according to the first vocabulary and the first language, and perform feature extraction on the first target sound data to obtain a second spectral feature, a second time domain feature and a second spatial feature;

[0107] According to different scene requirements, the first spectral feature, the first time domain feature and the first spatial feature, and the second spectral feature, the second time domain feature and the second spatial feature are used to determine different types of second environmental noise data and second target sound data corresponding to the different scene requirements from the first environmental noise data and the first target sound data;

[0108] Mix different types of the second environmental noise data and the second target sound data in corresponding proportions to obtain corresponding first scene sound data in different scenes (for example, mix airport noise and conversation voice in a ratio of 3:1 to obtain an airport scene; mix street noise and time voice in a ratio of 10:1 to obtain an outdoor scene; in order to ensure the accuracy and adaptability of the model, it is necessary to set as many variable mixing ratios and scene types as possible).

[0109] In some possible embodiments of the present application, in the step of constructing the first anti-noise data using the first head accompaniment data, the control processing module is configured to:

[0110] The first scene sound data after mixing is sampled, features are extracted, and anti-noise samples are generated according to a preset anti-noise algorithm (such as using traditional noise cancellation algorithms, such as Spectral Subtraction and Wiener Filter, to preprocess the mixed audio and generate anti-noise samples; or designing an "Ideal Binary Mask" that simulates the human ear, estimating the energy distribution of speech and noise through time-frequency analysis, and generating a dataset for training; or using a generative adversarial network (GAN) to input a complex scene, automatically learn to generate corresponding ideal anti-noise, and then iteratively update the network with real data). (This anti-noise needs to cancel out the noise part of the input as much as possible while preserving the clear speech part).

[0111] The first scene sound data and the anti-noise sample data are organized into a training dataset, which is used to train a deep learning model to obtain anti-noise generation rules.

[0112] The first anti-noise data is generated based on the anti-noise generation rules, the first scene sound data, and the first head-related audio data; (the model's input is a complex input scene, and the output is ideal anti-noise).

[0113] In this embodiment of the invention, to improve the model's generalization ability, it is necessary to provide as much rich and varied training data as possible. Therefore, it is necessary to continuously design new scenarios, adjust the proportions, and collect datasets of more noisy environments and various languages ​​for iterative training.

[0114] Understandably, where possible, it is still necessary to manually evaluate the noise reduction effect of the model output and use the feedback to adjust the model structure and parameters.

[0115] In some possible embodiments of the present invention, the main control chip of the Bluetooth headset is equipped with a 3-axis or 6-axis digital gyroscope; two or more microphones are respectively arranged on the two ear shells of the Bluetooth headset; multiple controllable small speakers are integrated between the ear tips of the Bluetooth headset; the control processing module is configured as follows:

[0116] The digital gyroscope is used to detect the motion trajectory data, rotation speed data, and direction change data of the first user's head.

[0117] The data of each microphone is calibrated, and the phase and amplitude differences of the current ambient noise are detected;

[0118] The head rotation direction and head rotation amplitude are determined by using a preset head motion detection algorithm in combination with the motion trajectory data, the rotation speed data, and the direction change data.

[0119] determine the source direction of the current environmental noise by using a preset noise source positioning algorithm in combination with the phase and amplitude difference;

[0120] generate directional first anti-noise to offset the current environmental noise by using a preset directional anti-noise generation model to correspondingly change the phase and energy distribution of the anti-noise between the two ears according to the head rotation direction, the head rotation amplitude and the source direction.

[0121] In the embodiments of the present application, a gyroscope and a microphone array are integrated in a Bluetooth earphone to detect the head movement direction of the user and the noise source direction, and then the direction of the anti-noise is adjusted accordingly to achieve higher precision noise cancellation. Specifically, a 3-axis or 6-axis digital gyroscope is integrated in the main control chip of the Bluetooth earphone to detect the movement trajectory, speed and direction change of the user's head; two or more microphones are configured on the two ear shells of the Bluetooth earphone, and the data of each microphone is accurately calibrated to detect the phase and amplitude difference of the current environmental noise, so as to determine the main source direction of the noise (three or more microphones are used to collect environmental audio information in different directions, and the positional difference between the microphones is used to achieve more accurate noise source positioning and beam generation, thereby achieving better active noise reduction effect); a head movement detection algorithm and a noise source positioning algorithm are developed: the head movement detection algorithm combines the gyroscope data to determine the head rotation direction and amplitude; the noise source positioning algorithm combines the noise signals received by different microphones to determine the approximate source direction of the noise; a directional anti-noise generation model and algorithm are trained: according to the head movement direction and the noise source direction, the phase and energy distribution of the anti-noise between the two ears are correspondingly changed to generate directional anti-noise to offset external noise; in addition to configuring high-performance Bluetooth chips, gyroscopes, microphones and the like, a plurality of controllable small speakers are also integrated between the earbuds to receive directional anti-noise signals and accurately play back; through machine learning, each algorithm and model is continuously optimized, feedback is used to improve the generation effect of directional anti-noise, and more accurate and efficient active noise control is achieved.

[0122] In some possible embodiments of the present application, the noise reduction effect of the algorithm can be improved and the signal delay can be reduced from the following aspects:

[0123] 1. A deep neural network is used to learn the data characteristics of environmental noise, a more accurate noise model is established, and targeted noise suppression is realized on this basis. Deep learning algorithm can achieve good results on unstructured data, and is very suitable for noise analysis in complex environment.

[0124] 2. Convolutional neural network is good at extracting time-frequency features in audio signals, CNN can be used to analyze environmental audio, more accurately extract noise features, and then use these features to realize noise classification and suppression.

[0125] 3. By simulating the analysis process of human ear to sound, using reinforcement learning algorithm to let Bluetooth headset can learn to distinguish noise and useful signal like human ear, and to a certain extent simulate the auditory cognitive process of human brain.

[0126] 4. Using dense connection neural network to establish highly nonlinear mapping relationship between input audio signal and output signal (de-noised signal), using a large amount of training data to let the network can automatically learn the best noise suppression scheme.

[0127] 5. Using GPU and other parallel operation chips to realize parallel computation of algorithm, speeding up the training and inference process of network, so as to minimize the delay while ensuring the effect.

[0128] 6. Adaptive algorithm: according to the type and intensity of environmental noise, dynamically select the optimal noise suppression algorithm, or adjust the network parameters online, to achieve the best noise reduction effect, and expand the adaptive range by using multiple algorithms.

[0129] In summary, the application of advanced machine learning and deep learning algorithms to the field of noise reduction of Bluetooth headset, using parallel architecture and adaptive mechanism to reduce delay, is the key to realize high intelligent noise reduction.

[0130] In some possible embodiments of the present application, the noise reduction effect of Bluetooth headset can be improved from the following aspects:

[0131] 1. Environmental noise classification: using machine learning algorithm to classify and identify noise in different environments, such as classifying indoor noise into keyboard typing noise, speaking noise, TV noise, etc.; classifying outdoor noise into traffic noise, wind noise, etc.; then selecting a targeted noise reduction scheme according to the identification result. For example, in the keyboard typing noise environment, an algorithm for suppressing low-frequency wideband noise can be selected; in the speaking noise environment, an algorithm for identifying and extracting human voice features can be selected.

[0132] 2. Optimal scheme selection: according to the type and intensity of environmental noise, intelligently select the optimal noise reduction scheme or adjust the algorithm hyperparameters to achieve the best noise reduction effect. For example, in a less noisy environment, a relatively simple noise reduction scheme can be selected to reduce computational complexity and delay; in a complex environment with a lot of noise, a more powerful deep learning active noise reduction scheme can be selected.

[0133] In addition, different noise reduction schemes can also be selected according to user's custom settings and usage habits. For example, some users are not sensitive to human voice, and a scheme that suppresses human voice noise more aggressively can be selected.

[0134] In summary, by identifying noise characteristics in different environments through machine learning and selecting targeted noise reduction solutions based on the identification results, this artificial intelligence technology can maximize the noise reduction effect of Bluetooth earphones in complex environments.

[0135] In some possible embodiments of the present application, reinforcement learning is a machine learning method that solves complex problems through simulation learning. To improve the noise reduction effect of Bluetooth earphones using reinforcement learning, specifically:

[0136] 1. Establish a human ear hearing model: by learning the human ear anatomy and auditory signal processing mechanism, a mathematical model is established to simulate the frequency domain analysis and cognitive process of the human ear to sound.

[0137] 2. Intelligent environment construction: build a virtual environment that simulates a real auditory environment, and deploy an intelligent Bluetooth earphone module in this environment, which can receive environmental audio input signals.

[0138] 3. Set up a reward mechanism: define the behavior reward mechanism of the intelligent Bluetooth earphone module, for example: correctly separate the human voice signal in a noisy environment, reduce signal distortion, etc., and get positive rewards, and negative rewards in case of failure.

[0139] 4. Intelligent Bluetooth earphone module learning and optimization: the intelligent Bluetooth earphone module uses human voice and noise data in the environment for a large number of simulated "listening" training, constantly learning and optimizing its own strategy (such as noise reduction scheme) to maximize cumulative rewards.

[0140] 5. Export the learning results: finally, the learning results of the intelligent Bluetooth earphone module are applied to the real Bluetooth earphone device as the device's firmware or software to realize the noise reduction function.

[0141] In the embodiments of the present application, through a large number of simulation training, the intelligent Bluetooth earphone module can analyze and identify sound like the human auditory system, and then learn to choose the best noise reduction scheme. This way of simulating biological knowledge can develop a more humanized noise reduction system to some extent.

[0142] In some possible embodiments of the present application, a dense neural network (DNN) is used to implement noise reduction for Bluetooth earphones, specifically:

[0143] 1. Network structure design: design a deep neural network composed of multiple layers of perceptrons, including an input layer, multiple hidden layers, and an output layer; the input layer receives the original audio signal, and the output layer produces the noise-reduced audio signal.

[0144] 2. Large dataset: Collect a large number of noisy audio samples as training data for the network, which needs to include different types of environmental noise and human voice audio; these data are used for the training and learning of the network.

[0145] 3. Network training: Train the neural network on the training data through the backpropagation algorithm, constantly optimize the parameters (weights and biases) of the network to minimize the difference between the output signal and the clear speech signal.

[0146] 4. Test and application: Use some test data to evaluate the noise reduction effect of the network, if it meets the requirements, deploy it on the Bluetooth headset device as a noise reduction system; if the effect is not good, more data needs to be added to retrain the network.

[0147] 5. Iterative optimization: With the expansion of the application environment and the accumulation of user data, it is necessary to use the latest data set to retrain the neural network regularly to continuously optimize and improve its noise reduction effect.

[0148] In this embodiment, by using DNN, the highly nonlinear relationship between input and output can be automatically learned to find the best noise suppression solution.

[0149] In some possible embodiments of the present application, parallel algorithm architecture can effectively reduce the calculation delay in the Bluetooth headset, which can be achieved in the following two aspects:

[0150] 1. GPU acceleration: Integrate GPU (Graphics Processing Unit) or other specialized neural network acceleration chips in the Bluetooth headset, and use their powerful parallel computing capabilities to accelerate the training and inference process of the neural network.

[0151] For example, GPU can be used to perform parallel computation of matrix multiplication, convolution operation, etc. in the network, which are the most time-consuming operations in deep neural networks. GPU can greatly improve the operation efficiency, thereby shortening the running time of the entire network.

[0152] 2. Algorithm parallel optimization: When designing the algorithm architecture of the neural network, use parallel optimization methods to make the network fully utilize the parallel computing resources of the GPU.

[0153] For example, data parallel method can be used to divide large-scale data sets, then perform small batch training on GPU, and finally aggregate parameters; or model parallel method can be used to divide the layers of the neural network, and different layers use different cores on the GPU for synchronous training. Pipeline parallelism can also be used to pipeline the forward propagation and backpropagation processes of the network on the GPU, reducing the overhead caused by intermediate variable transmission.

[0154] In summary, the parallel algorithm architecture can maximize the computing power of parallel devices such as GPUs through optimization of both algorithm design and hardware resources. This also enables some time-consuming deep learning models to be deployed on mobile devices such as Bluetooth earphones, which is one of the important ways to improve the intelligent algorithms of Bluetooth earphones.

[0155] In some possible embodiments of the present application, the adaptive algorithm refers to dynamically adjusting the parameters of the model or selecting different models to achieve the best effect according to environmental changes. The adaptive algorithm is used to realize adaptive noise reduction of the Bluetooth earphone, specifically:

[0156] 1. Constructing multiple noise reduction models: training multiple neural network models with different parameters or structures, which can reduce noise of different types and intensities. For example, one model is specifically for speech noise, one model is specifically for background music noise, and one powerful model is used for high-intensity noise environment.

[0157] 2. Environment recognition: using an environment recognition algorithm to detect the type and intensity of the current noise to select an appropriate noise reduction model; this recognition can be achieved by forward propagating multiple models and selecting the one with the best output.

[0158] 3. Model selection: based on the results of environment recognition, select the noise reduction model that best matches the current environment to process the input audio signal. When the environment changes, the model can also be dynamically switched.

[0159] 4. Parameter adjustment: for a certain neural network model, the hyperparameters (such as learning rate) can also be adjusted online according to environmental changes to achieve the best noise reduction effect, which also provides higher adaptability of the model.

[0160] 5. Multi-model fusion: the output results of different models can also be fused to produce a new stronger model; this ensemble learning method can expand the application range of the model.

[0161] The adaptive algorithm can greatly improve the adaptability of the Bluetooth earphone to environmental changes through the construction of multiple models, parameter adjustment, and model fusion, and achieve the purpose of dynamically adjusting to achieve the best noise reduction effect.

[0162] Please refer to Figure 2 Another embodiment of the present application provides a noise reduction control method for a Bluetooth earphone, which comprises:

[0163] Obtaining historical activity data and earphone historical use data of a first user of the Bluetooth earphone;

[0164] Obtaining historical location data and historical behavior data based on time records from the historical activity data;

[0165] Determining a plurality of first places according to the historical location data;

[0166] Obtaining first environmental noise data (including but not limited to airport noise, street noise, coffee shop noise, office noise, etc.) corresponding to different first places, and extracting first spectral features, first time domain features, and first spatial domain features of various first environmental noise data;

[0167] Training a first neural network using the first spectral features, the first time domain features, and the first spatial domain features to obtain a first environmental noise recognition model capable of recognizing different types of environmental noise;

[0168] Obtaining first head accompaniment data, and obtaining a first noise cancellation model according to the first environmental noise data and the first head accompaniment data (generating human head accompaniment closer to real anti-noise, achieving more personalized noise reduction effect);

[0169] Integrating the trained first noise recognition model and the first noise cancellation model into the Bluetooth earphone;

[0170] When the Bluetooth earphone is working, determining the current environmental noise type according to the first noise recognition model;

[0171] Selecting a matching noise cancellation model from the first noise cancellation model according to the current environmental noise type to generate optimal anti-noise for playback. That is, personalized anti-noise is generated according to user head accompaniment data for playback, and the model is continuously optimized according to use feedback.

[0172] The technical scheme of this embodiment can not only intelligently and efficiently recognize the noise type in the current environment, but also accurately select the corresponding noise cancellation model for noise reduction.

[0173] In the embodiments of the present application, the head-related impulse response (HRIR) refers to the response data of the human head to sound signals, which contains information about the acoustic characteristics and the structure of the human head. When environmental noise enters the ear canal of a person, it will produce complex multiple reflections, attenuation and diffraction in the head, and finally reach the eardrum and cochlea. The HRIR data records all the filtering and deformation of the head during the process of environmental sound entering the ear canal and reaching the eardrum. Specifically, the HRIR data records: 1) the acoustic transmission function of the external auditory canal, such as the attenuation and phase change of the external auditory canal to different frequency noises, etc. 2) the acoustic structure of the skull, such as the sound diffraction caused by the density and shape of the skull, the acoustic effects of various bone joints and cartilages, etc. 3) the effect of head axis and rotation on sound. The path of sound entering and exiting the ear canal is different when the head direction is different, resulting in differences in HRIR. Therefore, the HRIR data contains unique acoustic structure and acoustic channel characteristics of the human head, and using HRIR data can perform personalized transformation on any noise signal to simulate the sound effect of real noise entering the head, which lays a foundation for generating personalized anti-noise and realizing high-precision active noise control and elimination technology. In the noise reduction application of Bluetooth earphones, collecting and using the HRIR data of users can generate more realistic and head-acoustic-structure-fitting anti-noise, thereby achieving the purpose of improving individual differences and improving noise reduction accuracy.

[0174] The scheme of the present application can be personalized and optimized in combination with user personal information and use scene habits, and realize more intelligent artificial noise elimination. The key of the scheme of the present application is to obtain sufficient feature data to train the recognition model and the generation model, so as to realize high-precision noise classification and matching noise elimination, thereby making the noise reduction effect of Bluetooth earphones more intelligent and personalized. This requires extensive data collection and powerful AI computing capability.

[0175] In some possible embodiments of the present application, a user self-help head-related sound collection tool can be provided to upload the cloud to continuously enrich the head-related sound database, realize user participation and cooperation, and comprehensively collect more detailed head-related sound information, and generate more accurate and personalized anti-noise. This is a scheme that requires technical innovation and breakthrough in all aspects, including data collection equipment, large-scale data set construction, powerful AI training computing platform, precise model training and deployment, etc.

[0176] In some possible embodiments of the present application, an end-to-end noise reduction model based on deep learning is developed, which directly maps environmental noise to anti-noise without manually extracting noise features and designing noise elimination algorithms, so that the AI model automatically learns features and algorithms, and a more intelligent noise reduction scheme can be realized.

[0177] It can be understood that, in order to obtain accurate anti-noise data to establish a more accurate noise cancellation model, in some possible embodiments of the present application, the step of obtaining the first head companion sound data and obtaining the first noise cancellation model according to the first environmental noise data and the first head companion sound data comprises:

[0178] A large amount of environmental noise data of various types is collected by designing a precise noise collection device, a large number of sample data covering as many scenarios as possible are obtained, and the data classification and characteristics are labeled. The user's head companion sound data, including the impulse response data of the ear canal entrance, is obtained.

[0179] An artificial ear canal simulator is constructed according to the ear canal data of the first user. For example, a three-dimensional model of the ear canal can be generated according to the image data, structural data, etc. of the ear canal of the first user, and corresponding physiological parameters, physical parameters, biological response parameters (which can be determined according to test data or historical behavior data of the user combined with relevant data / medical data of other users) of the ear canal are added based on the three-dimensional model to establish a virtual artificial ear canal simulator.

[0180] First simulated sound data is generated according to the first environmental noise data, the historical use data of the earphone, and the historical behavior data. For example, corresponding historical playback data, historical scene data, and historical noise processing data of the user can be determined according to the historical use data of the earphone and the historical behavior data, and then combined with the first environmental noise data, sound data meeting the requirements can be selected to be fused into the first simulated sound data.

[0181] The first simulated sound data is input into the artificial ear canal simulator to obtain the first head companion sound data.

[0182] First scene sound data is constructed by mixing different types of first environmental noise data and first target sound data.

[0183] First anti-noise data is constructed using the first head companion sound data (for example, but not limited to, an artificial head companion sound data is synthesized by training a generative adversarial network (GAN) using the first environmental noise data and the first head companion sound data, and the network is iterated by comparing the artificial head companion sound data with the first head companion sound data to make the artificial head companion sound data more realistic. A deep learning model is trained on a large-scale artificial / original head companion sound data set to learn the mapping relationship model of the head companion sound features and the optimal anti-noise. The model needs to consider the head shape, sound channel structure, noise characteristics, etc. to generate personalized anti-noise, and obtain the first anti-noise data.

[0184] The first noise cancellation model is obtained by training a second neural network using the first scene sound data and the first anti-noise data (which needs to consider the frequency domain, time domain and spatial features of human voice and noise, and accurately extract the noise part).

[0185] In order to accurately determine the relationship between the scene and the environmental noise to obtain accurate anti-noise data, in some possible embodiments of the present application, the step of constructing the first scene sound data mixed with the first environmental noise data and the first target sound data of different types comprises:

[0186] Obtain the first spectral feature, the first time domain feature and the first spatial feature of the first environmental noise data.

[0187] Collect speech samples, and determine a first vocabulary and a first language from the speech samples according to a first preset rule.

[0188] Obtain the first target sound data from real human pronunciation or text-to-speech according to the first vocabulary and the first language, and perform feature extraction on the first target sound data to obtain a second spectral feature, a second time domain feature and a second spatial feature.

[0189] According to different scene requirements, the first spectral feature, the first time domain feature and the first spatial feature, and the second spectral feature, the second time domain feature and the second spatial feature are used to determine different types of second environmental noise data and second target sound data corresponding to the different scene requirements from the first environmental noise data and the first target sound data.

[0190] Mix different types of the second environmental noise data and the second target sound data in corresponding proportions to obtain corresponding first scene sound data in different scenes (for example, mix airport noise and conversation voice in a ratio of 3:1 to obtain a flight scene; mix street noise and time voice in a ratio of 10:1 to obtain an outdoor scene; in order to ensure the accuracy and adaptability of the model, it is necessary to set as many variable mixing ratios and scene types as possible).

[0191] In some possible embodiments of the present application, the step of constructing the first anti-noise data using the first head accompaniment data comprises:

[0192] The first scene sound data after mixing is sampled, features are extracted, and corresponding anti-noise sample data is generated according to a preset anti-noise algorithm (for example, a traditional noise elimination algorithm such as a Spectral Subtraction method, a Wiener Filter method, etc. is used to generate anti-noise samples by preprocessing mixed audio; for example, an 'Ideal Binary Mask' ideal binary mask simulating human ears is designed, energy distribution of voice and noise is estimated through time-frequency analysis, and a data set for training is generated; for example, a GAN generative adversarial network is used, a complex scene is input, and corresponding ideal anti-noise is automatically learned and generated, and then the network is iteratively updated with real data); (the anti-noise needs to counteract the noise part as much as possible and retain the clear voice part)

[0193] The first scene sound data and the anti-noise sample data are organized into a training data set for training a deep learning model to obtain an anti-noise generation rule;

[0194] The first anti-noise data is generated according to the anti-noise generation rule, the first scene sound data and the first head accompaniment sound data; (the input of the model is a complex input scene, and the output is ideal anti-noise)

[0195] In the embodiments of the present application, in order to improve the generalization ability of the model, it is necessary to provide as much diverse training data as possible. Therefore, it is necessary to continuously design new scenes, adjust the proportion, collect more noise environment and language type data sets for iterative training.

[0196] It can be understood that, as far as possible, the anti-noise effect of the model output needs to be evaluated by artificial evaluation, and used as feedback to adjust the model structure and parameters.

[0197] In some possible embodiments of the present application, a 3-axis or 6-axis digital gyroscope is configured in the master control chip of the Bluetooth earphone; two or more microphones are respectively configured on the two ear shells of the Bluetooth earphone; a plurality of controllable small speakers are integrated between the earplugs of the Bluetooth earphone; the noise reduction control method for the Bluetooth earphone further comprises:

[0198] The motion trajectory data, rotation speed data and direction change data of the head of the first user are detected by the digital gyroscope;

[0199] The data of each microphone is calibrated, and the phase and amplitude difference of the current environmental noise is detected;

[0200] The preset head motion detection algorithm is used to determine the head rotation direction and head rotation amplitude in combination with the motion trajectory data, the rotation speed data and the direction change data;

[0201] determine the source direction of the current environmental noise by using a preset noise source positioning algorithm in combination with the phase and amplitude difference;

[0202] generate directional first anti-noise to offset the current environmental noise by using a preset directional anti-noise generation model to correspondingly change the phase and energy distribution of the anti-noise between the two ears according to the head rotation direction, the head rotation amplitude and the source direction.

[0203] In the embodiments of the present application, a gyroscope and a microphone array are integrated in a Bluetooth earphone to detect the head movement direction of the user and the noise source direction, and then the direction of the anti-noise is adjusted accordingly to achieve higher precision noise cancellation. Specifically, a 3-axis or 6-axis digital gyroscope is integrated in the main control chip of the Bluetooth earphone to detect the movement trajectory, rotation speed and direction change of the user's head; two or more microphones are configured on the two ear shells of the Bluetooth earphone, and the data of each microphone is accurately calibrated to detect the phase and amplitude difference of the current environmental noise, so as to determine the main source direction of the noise (three or more microphones are used to collect environmental audio information in different directions, and the positional difference between the microphones is used to achieve more accurate noise source positioning and beam generation, thereby achieving better active noise reduction effect); a head movement detection algorithm and a noise source positioning algorithm are developed: the head movement detection algorithm combines the gyroscope data to determine the head rotation direction and amplitude; the noise source positioning algorithm combines the noise signals received by different microphones to determine the approximate source direction of the noise; a directional anti-noise generation model and algorithm are trained: according to the head movement direction and the noise source direction, the phase and energy distribution of the anti-noise between the two ears are correspondingly changed to generate directional anti-noise to offset external noise; in addition to configuring high-performance Bluetooth chips, gyroscopes, microphones and the like, a plurality of controllable small speakers are also integrated between the earbuds to receive directional anti-noise signals and accurately play back; through machine learning, each algorithm and model is continuously optimized, feedback is used to improve the generation effect of directional anti-noise, and more accurate and efficient active noise control is achieved.

[0204] In some possible embodiments of the present application, the noise reduction effect of the algorithm can be improved and the signal delay can be reduced from the following aspects:

[0205] 1. A deep neural network is used to learn the data characteristics of environmental noise, to establish a more accurate noise model, and to achieve targeted noise suppression on this basis. Deep learning algorithms can achieve good results on unstructured data, and are very suitable for noise analysis in complex environments.

[0206] 2. Convolutional neural networks are good at extracting time-frequency features in audio signals, and CNN can be used to analyze environmental audio to more accurately extract noise features, and then use these features to classify and suppress noise.

[0207] 3. By simulating the analysis process of human ear to sound, using reinforcement learning algorithm to let Bluetooth headset can learn to distinguish noise and useful signal like human ear, and to a certain extent simulate the auditory cognitive process of human brain.

[0208] 4. Using dense connection neural network to establish a highly nonlinear mapping relationship between input audio signal and output signal (de-noised signal), using a large amount of training data to let the network can automatically learn the best noise suppression scheme.

[0209] 5. Using GPU and other parallel operation chips to realize parallel computation of algorithm, speeding up the training and inference process of network, so as to minimize the delay while ensuring the effect.

[0210] 6. Adaptive algorithm: according to the type and intensity of environmental noise, dynamically select the optimal noise suppression algorithm, or adjust the network parameters online, to achieve the best noise reduction effect, and expand the adaptive range by using multiple algorithms.

[0211] In summary, the application of advanced machine learning and deep learning algorithms to the field of noise reduction of Bluetooth headset, using parallel architecture and adaptive mechanism to reduce delay, is the key to realize high-intelligence noise reduction.

[0212] In some possible embodiments of the present application, the noise reduction effect of Bluetooth headset can be improved from the following aspects:

[0213] 1. Environmental noise classification: using machine learning algorithm to classify and identify noise in different environments, such as classifying indoor noise into keyboard typing noise, speaking noise, TV noise, etc.; classifying outdoor noise into traffic noise, wind noise, etc.; then selecting a targeted noise reduction scheme according to the identification result. For example, in the keyboard typing noise environment, an algorithm for suppressing low-frequency wideband noise can be selected; in the speaking noise environment, an algorithm for identifying and extracting human voice features can be selected.

[0214] 2. Optimal scheme selection: according to the type and intensity of environmental noise, intelligently select the optimal noise reduction scheme or adjust the algorithm hyperparameters to achieve the best noise reduction effect. For example, in a less noisy environment, a relatively simple noise reduction scheme can be selected to reduce computational complexity and delay; in a complex environment with a lot of noise, a more powerful deep learning active noise reduction scheme can be selected.

[0215] In addition, different noise reduction schemes can also be selected according to user's custom settings and usage habits. For example, some users are not sensitive to human voice, and a scheme that suppresses human voice noise more aggressively can be selected.

[0216] In summary, by identifying noise characteristics in different environments through machine learning and selecting targeted noise reduction solutions based on the identification results, this artificial intelligence technology can maximize the noise reduction effect of Bluetooth earphones in complex environments.

[0217] In some possible embodiments of the present application, reinforcement learning is a machine learning method that solves complex problems through simulation learning. To improve the noise reduction effect of Bluetooth earphones using reinforcement learning, specifically:

[0218] 1. Establish a human ear hearing model: by learning the human ear anatomy and auditory signal processing mechanism, a mathematical model is established to simulate the frequency domain analysis and cognitive process of the human ear to sound.

[0219] 2. Intelligent environment construction: build a virtual environment that simulates a real auditory environment, and deploy an intelligent Bluetooth earphone module in this environment, which can receive environmental audio input signals.

[0220] 3. Set up a reward mechanism: define the behavior reward mechanism of the intelligent Bluetooth earphone module, for example: correctly separate the human voice signal in a noisy environment, reduce signal distortion, etc., and get positive rewards, and negative rewards for mistakes.

[0221] 4. Intelligent Bluetooth earphone module learning and optimization: the intelligent Bluetooth earphone module uses human voice and noise data in the environment for a large number of simulated "listening" training, constantly learning and optimizing its own strategy (such as noise reduction scheme) to maximize cumulative rewards.

[0222] 5. Export the learning results: finally, the learning results of the intelligent Bluetooth earphone module are applied to the real Bluetooth earphone device as the device's firmware or software to realize the noise reduction function.

[0223] In the embodiments of the present application, through a large number of simulation training, the intelligent Bluetooth earphone module can analyze and identify sound like the human auditory system, and then learn to choose the best noise reduction scheme. This way of simulating biological knowledge can develop a more humanized noise reduction system to some extent.

[0224] In some possible embodiments of the present application, a dense neural network (DNN) is used to implement noise reduction for Bluetooth earphones, specifically:

[0225] 1. Network structure design: design a deep neural network composed of multiple layers of perceptrons, including an input layer, multiple hidden layers, and an output layer; the input layer receives the original audio signal, and the output layer produces the noise-reduced audio signal.

[0226] 2. Large dataset: Collect a large number of noisy audio samples as training data for the network, which needs to include different types of environmental noise and human voice audio; these data are used for the training and learning of the network.

[0227] 3. Network training: Train the neural network on the training data through the backpropagation algorithm, constantly optimize the parameters (weights and biases) of the network to minimize the difference between the output signal and the clear speech signal.

[0228] 4. Test and application: Use some test data to evaluate the noise reduction effect of the network, if it meets the requirements, deploy it on the Bluetooth headset device as a noise reduction system; if the effect is not good, more data needs to be added to retrain the network.

[0229] 5. Iterative optimization: With the expansion of the application environment and the accumulation of user data, it is necessary to use the latest data set to retrain the neural network regularly to continuously optimize and improve its noise reduction effect.

[0230] In this embodiment, by using DNN, the highly nonlinear relationship between input and output can be automatically learned to find the best noise suppression solution.

[0231] In some possible embodiments of the present application, parallel algorithm architecture can effectively reduce the calculation delay in the Bluetooth headset, which can be achieved in the following two aspects:

[0232] 1. GPU acceleration: Integrate GPU (Graphics Processing Unit) or other specialized neural network acceleration chips in the Bluetooth headset, and use their powerful parallel computing capabilities to accelerate the training and inference process of the neural network.

[0233] For example, GPU can be used to perform parallel computation of matrix multiplication, convolution operation and other operations in the network, which are the most time-consuming operations in deep neural networks. GPU can greatly improve the operation efficiency and shorten the running time of the entire network.

[0234] 2. Algorithm parallel optimization: When designing the algorithm architecture of the neural network, use parallel optimization methods to make the network fully utilize the parallel computing resources of the GPU.

[0235] For example, data parallel method can be used to divide large-scale data sets, then perform small batch training on GPU, and finally aggregate the parameters; or model parallel method can be used to divide the layers of the neural network, and different layers use different cores on the GPU for synchronous training. Pipeline parallelism can also be used to pipeline the forward propagation and backpropagation processes of the network on the GPU, reducing the overhead caused by intermediate variable transmission.

[0236] In summary, the parallel algorithm architecture can maximize the computing power of parallel devices such as GPUs through optimization of both algorithm design and hardware resources. This also enables some time-consuming deep learning models to be deployed on mobile devices such as Bluetooth earphones, which is one of the important ways to improve the intelligent algorithms of Bluetooth earphones.

[0237] In some possible embodiments of the present application, the adaptive algorithm refers to dynamically adjusting the parameters of the model or selecting different models to achieve the best effect according to environmental changes. The adaptive algorithm is used to achieve adaptive noise reduction of the Bluetooth earphone, specifically as follows:

[0238] 1. Constructing multiple noise reduction models: training multiple neural network models with different parameters or structures, which can reduce noise of different types and intensities. For example, one model is specifically for speech noise, one model is specifically for background music noise, and one powerful model is used for high-intensity noise environment.

[0239] 2. Environment recognition: using an environment recognition algorithm to detect the type and intensity of the current noise to select the appropriate noise reduction model; the recognition can be achieved by forward propagating multiple models to select the one with the best output.

[0240] 3. Model selection: based on the results of environment recognition, select the noise reduction model that best matches the current environment to process the input audio signal, and when the environment changes, the model can also be dynamically switched.

[0241] 4. Parameter adjustment: for a certain neural network model, the hyperparameters (such as learning rate, etc.) can also be adjusted online according to environmental changes to achieve the best noise reduction effect, which also provides higher adaptability of the model.

[0242] 5. Multi-model fusion: the output results of different models can also be fused to produce a new stronger model; this ensemble learning method can expand the application range of the model.

[0243] The adaptive algorithm can greatly improve the adaptability of the Bluetooth earphone to environmental changes through the construction of multiple models, parameter adjustment, and model fusion, and achieve the purpose of dynamically adjusting to achieve the best noise reduction effect.

[0244] It should be noted that, for the foregoing method embodiments, in order to simply describe, they are all expressed as a series of action combinations, but those skilled in the art should know that the present application is not limited to the action sequence described, because according to the present application, certain steps can be performed in other sequences or simultaneously. Secondly, those skilled in the art should know that the embodiments described in the specification all belong to preferred embodiments, and the actions and modules involved are not necessarily required by the present application.

[0245] In the above-described embodiments, the description of each embodiment focuses on different aspects, and the parts not described in detail in a certain embodiment can be referred to the relevant description of other embodiments.

[0246] In several embodiments provided in the present application, it should be understood that the disclosed apparatus can be implemented by other means. For example, the apparatus embodiments described above are only illustrative, for example, the division of the above-mentioned units is only a logical function division, and actual implementation can have another division manner, for example, a plurality of units or components can be combined or integrated into another system, or some features can be ignored or not executed. In addition, the coupling or direct coupling or communication connection between the units or components shown or discussed can be indirect coupling or communication connection through some interfaces, devices or units, which can be electrical or other forms.

[0247] The units described above as separate components can or can not be physically separate, and the components shown as units can or can not be physical units, that is, they can be located in one place, or they can be distributed on a plurality of network units. Part or all of the units can be selected according to actual needs to achieve the purpose of the embodiment.

[0248] In addition, each functional unit in each embodiment of the present application can be integrated in one processing unit, or each unit can exist physically, or two or more units can be integrated in one unit. The integrated unit can be realized in the form of hardware or in the form of a software functional unit.

[0249] The integrated unit described above, if realized in the form of a software functional unit and sold or used as an independent product, can be stored in a computer readable memory. Based on this understanding, the technical solutions of the present application essentially or the part of the prior art that makes a contribution or the whole or part of the technical solutions can be embodied in the form of a software product, which is stored in a memory and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server or a network device, etc.) to execute all or part of the steps of the above-mentioned method of each embodiment of the present application. The aforementioned memory includes: a U disk, a read-only memory (ROM, Read-Only Memory), a random access memory (RAM, Random Access Memory), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store program codes.

[0250] Those skilled in the art can understand that all or part of the steps of various methods in the above embodiments can be completed by instructing the relevant hardware through a program, and the program can be stored in a computer readable memory, which can include a flash disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, etc.

[0251] The above has carried out the detailed introduction to the embodiments of the application, the principle and implementation mode of the application are described in this paper by applying specific examples, the above embodiment explanation is only for helping to understand the method of the application and its core idea; at the same time, for the general technical personnel in the art, according to the idea of the application, the specific implementation mode and application range will have the change, according to the above, the content of the specification should not be understood as the limitation of the application.

[0252] Although the present application is disclosed as above, the present application is not limited to this. Any person skilled in the art can easily think of changes or substitutions without departing from the spirit and scope of the present application, and can make various changes and modifications, including the combination of different functions and implementation steps, including the software and hardware implementation, which are all within the protection scope of the present application.

Claims

1. A noise reduction control system for a Bluetooth headset, the system comprising: Comprise: A server, a control processing module, a communication module; The control processing module is configured to: Obtain historical activity data of a first user of a Bluetooth headset and headset historical use data from the server through the communication module; Obtain time-based historical location data and historical behavior data from the historical activity data; Determine a plurality of first places according to the historical location data; Obtain first environmental noise data corresponding to different first places, and extract first spectral features, first time domain features, and first spatial domain features of each type of first environmental noise data; Train a first neural network using the first spectral features, the first time domain features, and the first spatial domain features to obtain a first environmental noise recognition model that can recognize different types of environmental noise; Obtain first head accompaniment data, and obtain a first noise cancellation model according to the first environmental noise data and the first head accompaniment data; wherein the first head accompaniment data includes acoustic transmission function data of the external auditory canal, acoustic structure data of the skull, and data on the influence of head direction on sound; Integrate the trained first noise recognition model and the first noise cancellation model into the Bluetooth headset; When the Bluetooth headset is working, determine the current environmental noise type according to the first noise recognition model; Select a matching noise cancellation model from the first noise cancellation model according to the current environmental noise type to generate optimal anti-noise for playback; In the step of obtaining first head accompaniment data and obtaining a first noise cancellation model according to the first environmental noise data and the first head accompaniment data, the control processing module is configured to: Construct an artificial ear canal simulator according to the ear canal data of the first user; Generate first simulated sound data according to the first environmental noise data, the headset historical use data, and the historical behavior data; Input the first simulated sound data into the artificial ear canal simulator to obtain the first head accompaniment data; Construct first scene sound data mixed with different types of first environmental noise data and first target sound data; Construct first anti-noise data using the first head accompaniment data; Train a second neural network using the first scene sound data and the first anti-noise data to obtain a first noise cancellation model.

2. The noise reduction control system for a Bluetooth headset of claim 1, wherein, In the step of constructing first scene sound data mixed with different types of first environmental noise data and first target sound data, the control processing module is configured to: Obtain the first spectral features, the first time domain features, and the first spatial domain features of the first environmental noise data; Collect speech samples, and determine a first vocabulary and a first language from the speech samples according to a first preset rule; Obtain the first target sound data from real human pronunciation or text-to-speech synthesis according to the first vocabulary and the first language, and extract second spectral features, second time domain features, and second spatial domain features from the first target sound data; According to different scene requirements, the first frequency spectrum feature, the first time domain feature and the first space domain feature, and the second frequency spectrum feature, the second time domain feature and the second space domain feature are used to determine different types of second environmental noise data and second target sound data corresponding to the different scene requirements respectively; The second environmental noise data of different types and the second target sound data are mixed in a corresponding proportion to obtain the first scene sound data corresponding to different scenes.

3. The noise reduction control system for a Bluetooth headset of claim 2, wherein, In the step of constructing the first anti-noise data by using the first head accompaniment sound data, the control processing module is configured to: sample the mixed first scene sound data, extract features and generate corresponding anti-noise sample data according to a preset anti-noise algorithm; organize the first scene sound data and the anti-noise sample data into a training data set for training a deep learning model to obtain an anti-noise generation rule; generate the first anti-noise data according to the anti-noise generation rule, the first scene sound data and the first head accompaniment sound data.

4. The noise reduction control system for a Bluetooth headset of claim 3, wherein, The master control chip of the Bluetooth earphone is configured with a 3-axis or 6-axis digital gyroscope; two ear shells of the Bluetooth earphone are respectively configured with two or more microphones; a plurality of controllable small loudspeakers are integrated between the earplugs of the Bluetooth earphone; the control processing module is configured to: detect the motion trajectory data, rotation speed data and direction change data of the head of the first user through the digital gyroscope; calibrate the data of each microphone and detect the phase and amplitude difference of the current environmental noise; determine the head rotation direction and head rotation amplitude by using a preset head motion detection algorithm in combination with the motion trajectory data, the rotation speed data and the direction change data; determine the source direction of the current environmental noise by using a preset noise source positioning algorithm in combination with the phase and amplitude difference; generate directional first anti-noise to offset the current environmental noise by changing the phase and energy distribution of the anti-noise between the two ears according to the head rotation direction, the head rotation amplitude and the source direction by using a preset directional anti-noise generation model.

5. A noise reduction control method for a Bluetooth earphone, characterized by, The noise reduction control method for the Bluetooth earphone includes: obtain historical activity data and earphone historical use data of a first user of a Bluetooth earphone; obtain time-based historical location data and historical behavior data from the historical activity data; determine a plurality of first places according to the historical location data; obtain first environmental noise data corresponding to different first places, and extract first frequency spectrum features, first time domain features and first space domain features of various types of first environmental noise data; train a first neural network by using the first frequency spectrum features, the first time domain features and the first space domain features to obtain a first environmental noise recognition model capable of recognizing different types of environmental noise; acquire first head-voice data and obtain a first noise cancellation model according to the first environmental noise data and the first head-voice data; wherein the first head-voice data comprises acoustic transmission function data of an external ear canal, acoustic structure data of a skull, and data of the influence of head direction on sound; integrate the trained first noise identification model and the first noise cancellation model into the Bluetooth earphone; when the Bluetooth earphone is working, determine a current environmental noise type according to the first noise identification model; select a matched noise cancellation model from the first noise cancellation model according to the current environmental noise type to generate optimal anti-noise for playing; wherein the step of acquiring first head-voice data and obtaining a first noise cancellation model according to the first environmental noise data and the first head-voice data comprises: construct an artificial ear canal simulator according to the ear canal data of the first user; generate first simulated sound data according to the first environmental noise data, historical use data of the earphone, and historical behavior data; input the first simulated sound data into the artificial ear canal simulator to obtain the first head-voice data; construct first scene sound data mixed with different types of the first environmental noise data and first target sound data; construct first anti-noise data using the first head-voice data; train a second neural network using the first scene sound data and the first anti-noise data to obtain a first noise cancellation model.

6. The noise reduction control method for a Bluetooth headset according to claim 5, wherein, the step of constructing first scene sound data mixed with different types of the first environmental noise data and first target sound data comprises: acquire the first spectral feature, the first time-domain feature, and the first spatial feature of the first environmental noise data; collect speech samples, and determine a first vocabulary and a first language from the speech samples according to a first preset rule; obtain the first target sound data from real human pronunciation or text-to-speech according to the first vocabulary and the first language, and extract features from the first target sound data to obtain second spectral features, second time-domain features, and second spatial features; determine different types of second environmental noise data and second target sound data corresponding to different scene requirements using the first spectral feature, the first time-domain feature, and the first spatial feature, and the second spectral feature, the second time-domain feature, and the second spatial feature, respectively; mix different types of the second environmental noise data and the second target sound data in corresponding proportions to obtain corresponding first scene sound data under different scenes.

7. The noise reduction control method for a Bluetooth headset according to claim 6, wherein, the step of constructing first anti-noise data using the first head-voice data comprises: sample the mixed first scene sound data, extract features, and generate corresponding anti-noise sample data according to a preset anti-noise algorithm; organize the first scene sound data and the anti-noise sample data into a training data set for training a deep learning model to obtain an anti-noise generation rule; The first anti-noise data is generated according to the anti-noise generation rule, the first scene sound data and the first head accompaniment sound data.

8. The noise reduction control method for a Bluetooth headset according to claim 7, wherein, The master control chip of the Bluetooth earphone is configured with a 3-axis or 6-axis digital gyroscope; two ear shells of the Bluetooth earphone are respectively configured with two or more microphones; A plurality of controllable small loudspeakers are integrated between the earplugs of the Bluetooth earphone; the noise reduction control method for the Bluetooth earphone further comprises: The motion trajectory data, the rotation speed data and the direction change data of the head of the first user are detected by the digital gyroscope; The data of each microphone is calibrated, and the phase and amplitude difference of the current environmental noise is detected; The preset head motion detection algorithm is used to determine the head rotation direction and the head rotation amplitude in combination with the motion trajectory data, the rotation speed data and the direction change data; The preset noise source positioning algorithm is used to determine the source direction of the current environmental noise in combination with the phase and amplitude difference; The preset directional anti-noise generation model is used to change the phase and energy distribution of the anti-noise between the two ears according to the head rotation direction, the head rotation amplitude and the source direction, and to generate directional first anti-noise to offset the current environmental noise.

Images