Intelligent noise reduction communication headset

Abstract

The present invention relates to an intelligent noise reduction communication headset, belonging to the field of noise reduction and electronic equipment. The environmental microphone of the present invention is located at the top of the headset to obtain the environmental noise signal; the call microphone is connected to one side of the headset through the call microphone rotating shaft, and the call microphone obtains the voice signal of personnel communication; the signal preprocessing module completes the amplification, filtering and inversion of the environmental noise signal, and completes the superposition operation with the amplified and filtered voice signal; the signal analysis and processing module includes a signal digital-to-analog conversion module and a microprocessor system, which completes the collection, storage, learning and modeling of the environmental noise inversion signal, the voice amplification and filtering signal and the superposition signal of the two, and the processing, mode selection and control functions of the digital voice signal after the model operation; the wireless transmission module completes the encoding and transmission of the voice signal after noise reduction, and receives the voice signal of the caller, and outputs it to the headset after decoding. The present invention is not interfered by environmental noise and can ensure the quality of voice communication.

Description

Technical Field

[0001] The invention belongs to the field of noise reduction and electronic equipment, and in particular relates to an intelligent noise reduction communication headset. Background Art

[0002] Voice communication is the most basic, convenient and common way for people to communicate. In the communication between tunnel construction workers, tank and armored vehicle crews, etc., voice signals will be interfered by various noises from the external environment, such as noise interference from the surrounding environment, noise interference introduced in the transmission medium, electrical noise interference from the communication equipment itself, etc. These noise interferences will pollute the useful voice at the receiving end. When the interference reaches a serious level, the voice can be completely buried in the noise. The strong noise environment affects the effect of calls and communication. In order to protect against noise, on the one hand, the personnel are equipped with noise-isolating earmuffs or helmets for passive noise reduction, and on the other hand, active noise reduction is adopted. The microphone collects the external environmental noise, and then the system transforms it into an inverted sound wave and adds it to the speaker end. The sound heard by the human ear is: environmental noise + inverted environmental noise. The superposition of the two noises can achieve sensory noise reduction. This method can achieve effective noise reduction to a certain extent. However, since the personnel are in different positions, the noise intensity and phase in the communication are not the same. Although it is offset, there is still a lot of noise. In addition, due to the different voice characteristics of each person, the impact of noise is also different. Summary of the invention

[0003] 1. Technical issues to be resolved

[0004] The technical problem to be solved by the present invention is how to provide an intelligent noise reduction communication headset to solve the problem that the noise reduction effect of existing headsets is not ideal.

[0005] (II) Technical solution

[0006] In order to solve the above technical problems, the present invention proposes an intelligent noise reduction communication headset, which includes an environmental microphone, a call microphone, a signal preprocessing module, a signal analysis and processing module and a wireless transmission module.

[0007] The ambient microphone is located on the top of the earphone to obtain ambient noise signals;

[0008] The call microphone is connected to one side of the headset through a call microphone rotating shaft. The rotating shaft can rotate downward to cover the lower part of the user's lips. The call microphone obtains the voice signal of the communication between people.

[0009] The signal preprocessing module completes the amplification, filtering and inversion of the environmental noise signal, and performs superposition operation with the amplified and filtered voice signal;

[0010] The signal analysis and processing module includes a signal digital-to-analog conversion module and a microprocessor system, which completes the acquisition, storage, learning and modeling of the environmental noise inversion signal, the voice amplification and filtering signal and the superposition signal of the two, as well as the processing, mode selection and control functions of the digital voice signal after model operation;

[0011] The wireless transmission module encodes and sends the voice signal after noise reduction, receives the voice signal of the caller, decodes it and outputs it to the headset.

[0012] Furthermore, the call microphone is connected to a microphone line, and the microphone line is protected by a metal hose and plastic and can be bent as needed. The call microphone is located at the end of the microphone line.

[0013] Furthermore, the signal analysis and processing module includes a sampling control module, a database, a machine learning module and a digital signal processing system. The superimposed signal, the environmental noise signal and the voice signal are sampled at different sampling frequencies under the sampling control module and stored in different databases; secondly, the signal in the database completes the modeling of the machine learning model in the machine learning module, and the sampled signal realizes signal separation and noise reduction under the corresponding model parameters; subsequently, the noise signal is further noise-reduced through the corresponding digital signal processing system.

[0014] Furthermore, the digital signal processing system includes two major modules, namely a filtering module and a weight adaptive adjustment module; the filtering module determines the structure of the digital signal processing system, and the weight adaptive adjustment module adjusts the weight of each input signal vector, which is determined by the machine learning model parameters.

[0015] Furthermore, the machine learning module provides input signals to the digital signal processing system, as well as the expected response and error signals generated during the adjustment. The weighted delay operators are superimposed through the adder, and the size of the weights is adjusted according to the model parameters to obtain the optimal response signal.

[0016] Furthermore, the machine learning module uses the preprocessed noise and speech signals as inputs of the machine learning model and the speech signals obtained in a low-noise environment as outputs. Based on the maximum signal-to-noise ratio as the convergence criterion of the model, a training model is established through machine learning, and the model parameters, expected responses, and error signals generated during the adjustment are output to the digital signal processing system.

[0017] Furthermore, the preprocessed noise is environmental noise or white noise.

[0018] Furthermore, the machine learning module first obtains in advance a large number of voice signals received by the call microphone in a low-noise environment and superimposed with noise signals in different environments as a noisy signal as an input signal for machine learning, and uses the low-noise environment signal as an output signal to train the noise reduction model; predicting and controlling the signal from the noisy voice signal requires training to obtain two models: one is a noise signal model, and the other is a voice signal model; after performing the steps of pre-emphasis, framing, windowing, fast Fourier transform (FFT), filter bank and discrete cosine transform (DCT) on the signal, the characteristic coefficient of the signal is obtained; after extracting the signal features, the naive Bayes algorithm in machine learning is used to determine whether each frame signal is a voice signal through the features, and a large amount of data is used for training to distinguish noise signals from voice signals and remove noise signals, and the noise signals are compared with the low-noise environment voice signal to determine whether the voice signal after the noise is removed is qualified, specifically, the signal-to-noise ratio is pre-set, and when the preset signal-to-noise ratio is reached, the next step is continued, and if it is not reached, the step of removing noise is continued until the requirement is met, and a noise reduction model that meets the best signal-to-noise ratio is obtained.

[0019] Furthermore, the machine learning module reads in the optimal model parameters obtained by machine learning based on the noise reduction model that satisfies the best signal-to-noise ratio, inputs the collected signal to be denoised, performs machine learning processing, and sends it to each end of the call through the transmission module.

[0020] Furthermore, each party in communication generates its own noise reduction model and outputs a speech signal with an optimal signal-to-noise ratio.

[0021] (III) Beneficial effects

[0022] The present invention proposes an intelligent noise reduction communication headset. The method can learn according to the noise characteristics and personal voice characteristics, and automatically generate a noise reduction model related to the noise and personal voice characteristics, so as to achieve optimized noise reduction and send the call voice to the other party more clearly; the voice signal received by itself is not interfered by the environmental noise, thereby ensuring the quality of voice communication.

[0023] The present invention adopts the principle of artificial intelligence to jointly model the voice signals of people in different environments and the environmental noise signals, and can achieve the best signal-to-noise ratio of voice communication signals in different noise environments. Specific advantages:

[0024] Self-learning function of communication headset;

[0025] Speech signal output and transmission under the optimal signal-to-noise ratio criterion based on artificial intelligence methods;

[0026] Modeling, debugging and verification in laboratory environment and intelligent information in noisy working environment. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] Figure 1 It is a block diagram of the system composition of the present invention;

[0028] Figure 2 is a machine learning flow chart of the present invention;

[0029] Figure 3 It is a structural diagram of a digital signal processing system of the present invention;

[0030] Figure 4 is a flow chart of machine training of the present invention;

[0031] Figure 5 This is a flowchart of machine learning in an actual noise environment of the present invention;

[0032] Figure 6 It is a flow chart of the noise reduction working process of the present invention;

[0033] Figure 7 This is a white noise-based machine learning model for the present invention. DETAILED DESCRIPTION

[0034] In order to make the purpose, content and advantages of the present invention more clear, the specific implementation methods of the present invention are further described in detail below in conjunction with the drawings and examples.

[0035] In view of the above problems, the present invention proposes an intelligent noise reduction method based on artificial intelligence. The method can learn according to the noise characteristics and personal voice characteristics, automatically generate a noise reduction model related to the noise and personal voice characteristics, and achieve optimized noise reduction. The call voice is sent to the other party more clearly; the voice signal received by oneself is not interfered by environmental noise, ensuring the effect of voice communication quality.

[0036] The present invention uses speech clarity evaluation as a criterion. First, the signal obtained by the ambient noise receiving microphone and the signal received by the call microphone are used as the input signal of the system, and the voice of the caller in a low-noise environment is used as the output of the system. The machine learning method is used to establish a machine learning model. When in use, the callers train the system according to their respective call content so that the voices of the callers are clear.

[0037] After the training is completed, it can be applied. The present invention includes the following aspects:

[0038] 1. A noise-cancelling headset based on artificial intelligence.

[0039] Figure 1 The noise reduction headset includes an ambient microphone, a call microphone, a signal preprocessing module, a signal analysis and processing module, and a wireless transmission module.

[0040] The ambient microphone is located on the top of the earphone to pick up ambient noise signals.

[0041] The call microphone is connected to one side of the headset through the call microphone rotating shaft. The rotating shaft can rotate downward to cover the bottom of the user's lips. The call microphone is connected to the microphone cable. The microphone cable is protected by a metal hose and plastic and can be bent at will as needed. The call microphone is located at the end of the microphone cable and is used to obtain voice signals for communication between people.

[0042] The signal preprocessing module completes the amplification, filtering and inversion of the environmental noise signal, and performs superposition operation with the amplified and filtered voice signal;

[0043] The signal analysis and processing module is composed of a signal digital-to-analog conversion module and a microprocessor system, which completes the acquisition, storage, learning and modeling of the environmental noise inversion signal, the voice amplification and filtering signal and the superposition signal of the two, and the processing of the digital voice signal after model operation, mode selection and control, etc.

[0044] The wireless transmission module encodes and sends the voice signal after noise reduction, receives the voice signal of the caller, decodes it and outputs it to the headset.

[0045] Figure 2 The following is the system workflow. First, the system obtains the ambient noise signal and voice signal from the ambient microphone and the call microphone respectively. The two signals are superimposed after different preprocessing, and part of the noise in the superimposed signal is canceled. The basic principle of cancellation: let the call microphone be A and the ambient microphone be B. A and B are two microphones with the same performance. During a normal voice call, the mouth is close to microphone A, which produces a larger audio signal Va. At the same time, microphone B will also get some voice signal Vb, but it is much smaller than A. The two signals are input into the microphone processor, and the two signals are amplified and subtracted, so the resulting signal is Vm. a =Va a -Vb a If there is background noise in the usage environment, because the noise source is far away from the headset, the intensity of the sound waves reaching the two microphones of the headset is almost the same, that is, Va n ≈Vb n , so for background noise, although both microphones pick up, Vm n =Va n -Vb n ≈0, such a design can effectively resist the interference of surrounding environmental noise and greatly improve the clarity of normal calls. Under ideal conditions, the environmental noise of the call is completely eliminated, but under strong environmental noise conditions, some noise still remains in the voice signal, which requires further noise reduction.

[0046] like Figure 2As shown, the signal analysis and processing module includes a sampling control module, a database, a machine learning module and a digital signal processing system. The superimposed signal, the environmental noise signal and the voice signal are sampled at different sampling frequencies under the sampling control module and stored in different databases. Secondly, the signal in the database is modeled by the machine learning model in the machine learning module, and the sampled signal is separated and denoised under the corresponding model parameters; then, the noise signal is further denoised through the corresponding digital signal processing system.

[0047] Figure 3 This is the structure diagram of the digital signal processing system. It includes two modules, namely the filtering module and the weight adaptive adjustment module. The filtering module determines the structure of the digital signal processing system, and the weight adaptive adjustment module adjusts the weight of each input signal vector, which is determined by the machine learning model parameters. In addition to the input signal, the machine learning model is also required to provide the expected response and the error signal generated during the adjustment. The weighted delay operator is superimposed through the adder, and the size of the weight is adjusted according to the model parameters to obtain the optimal response signal.

[0048] Finally, the code is sent to the other callers. The voice signals of other callers enter the receiving system through the receiving antenna, are demodulated, and then decoded and output to the headset.

[0049] 2. Noise reduction method based on machine learning.

[0050] The machine learning module uses the pre-processed noise (environmental noise or white noise) and speech signal as the input of the machine learning model, and the speech signal obtained in a low-noise environment as the output. Based on the maximum signal-to-noise ratio as the convergence criterion of the model, a training model is established through machine learning, and the model parameters, expected response and error signals generated during the adjustment are output to the digital signal processing system. The specific steps are as follows: Figure 4 As shown, Figure 5 A flowchart for machine learning.

[0051] First, a large number of voice signals received by the microphone in a low-noise environment are obtained in advance, and the noise signals of different environments are superimposed to form a noisy signal as the input signal of machine learning. The low-noise environment signal is used as the output signal to train the noise reduction model. To predict and control the signal from the noisy speech signal, two models need to be trained: one is the noise signal model, and the other is the speech signal model. After the signal is pre-emphasized, framed, windowed, fast Fourier transform (FFT), filter bank and discrete cosine transform (DCT), the characteristic coefficient of the signal is obtained. After extracting the signal features, the naive Bayes algorithm in machine learning is used to determine whether each frame signal is a speech signal by features. A large amount of data is used for training to distinguish between noise signals and speech signals and remove the noise signal. The signal after noise removal is compared with the speech signal in the low-noise environment to determine whether the speech signal after noise removal is qualified. Specifically, the signal-to-noise ratio is pre-set. When the preset signal-to-noise ratio is reached, the next step is continued. If it is not reached, the step of removing noise is continued until the requirement is met, and a noise reduction model that meets the best signal-to-noise ratio is obtained.

[0052] Based on the noise reduction model that satisfies the best signal-to-noise ratio, the signal after digital-to-analog conversion is input into the model to obtain the voice signal output with the highest signal-to-noise ratio. All parties in communication can generate their own noise reduction model based on the above process and output the voice signal with the best signal-to-noise ratio. Figure 6 This is a flowchart of the noise reduction process. First, the optimal model parameters obtained by machine learning are read in, and the collected signal to be noise-reduced is input. After machine learning processing, it is sent to each end of the call through the transmission module.

[0053] 3. Signal transmission

[0054] The digital voice signal output by the system is encoded and transmitted by wire or wirelessly. After the other party receives the signal, it is decoded and converted to digital to analog and directly output to the receiver's headset, thus obtaining the other party's communication voice with the highest signal-to-noise ratio.

[0055] 4. Simulation Experiment Method

[0056] In a laboratory environment, in order to conduct machine learning, there is often a lack of an actual noise environment. In order to model the system for laboratory testing and debugging, the present invention proposes a system learning method based on white noise. This method uses white noise as a noise signal, superimposes white noise signals of different intensities with low-noise speech signals as the input of the machine learning system, and uses the low-noise speech signal as the output of the machine learning system for machine learning modeling. The method and process are the same as 2. The training process is as follows Figure 7As shown. First, read in the low-noise speech signal, input white noise of different intensities according to the requirements of the test and debugging to simulate the actual noise environment, compare the low-noise speech signal processed by machine learning with the read-in low-noise speech signal to determine whether the signal-to-noise ratio meets the requirements. After the model that meets the conditions is solidified and the parameters are stored, the modeling is completed.

[0057] Advantages of the present invention

[0058] The present invention adopts the principle of artificial intelligence to jointly model the voice signals of people in different environments and the environmental noise signals, and can achieve the best signal-to-noise ratio of voice communication signals in different noise environments. Specific advantages:

[0059] 1. Self-learning function of communication headset;

[0060] 2. Speech signal output and transmission under the optimal signal-to-noise ratio criterion based on artificial intelligence methods.

[0061] Modeling, debugging and verification in laboratory environment and intelligent information in noisy working environment.

[0062] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. An intelligent noise reduction communication headset, characterized in that: The noise reduction headset includes an ambient microphone, a call microphone, a signal preprocessing module, a signal analysis and processing module, and a wireless transmission module. The ambient microphone is located on the top of the earphone to obtain ambient noise signals; The call microphone is connected to one side of the headset through a call microphone rotating shaft. The rotating shaft can rotate downward to cover the lower part of the user's lips. The call microphone obtains the voice signal of the communication between people. The signal preprocessing module completes the amplification, filtering and inversion of the environmental noise signal, and performs superposition operation with the amplified and filtered voice signal; The signal analysis and processing module includes a signal digital-to-analog conversion module and a microprocessor system, which completes the acquisition, storage, learning and modeling of the environmental noise inversion signal, the voice amplification and filtering signal and the superposition signal of the two, as well as the processing, mode selection and control functions of the digital voice signal after model operation; The wireless transmission module encodes and sends the voice signal after noise reduction, and receives the voice signal of the caller, decodes it and outputs it to the headset; in, The signal analysis and processing module includes a sampling control module, a database, a machine learning module and a digital signal processing system. The superimposed signal, the environmental noise signal and the voice signal are sampled at different sampling frequencies under the sampling control module and stored in different databases; secondly, the signal of the database is modeled by a machine learning model in the machine learning module, and the sampled signal is separated and denoised under the corresponding model parameters; then, the noise signal is further denoised through the corresponding digital signal processing system; The machine learning module uses the pre-processed noise and speech signal as the input of the machine learning model, and the speech signal obtained in a low-noise environment as the output, and establishes a training model through machine learning based on the maximum signal-to-noise ratio as the convergence criterion of the model, and outputs the model parameters, expected response and error signal generated in the adjustment to the digital signal processing system; The machine learning module first obtains in advance a large number of voice signals received by the call microphone in a low-noise environment and superimposes the noise signals of different environments into a noisy signal as the input signal of the machine learning, and uses the low-noise environment signal as the output signal to train the noise reduction model; predicting and controlling the signal from the noisy voice signal requires training to obtain two models: one is a noise signal model, and the other is a voice signal model; after performing the steps of pre-emphasis, framing, windowing, fast Fourier transform (FFT), filter bank and discrete cosine transform (DCT) on the signal, the characteristic coefficient of the signal is obtained; after extracting the signal features, the naive Bayes algorithm in machine learning is used to determine whether each frame signal is a voice signal by features, and a large amount of data is used for training to distinguish the noise signal from the voice signal and remove the noise signal, and the noise signal is compared with the low-noise environment voice signal to determine whether the voice signal after the noise is removed is qualified, specifically, the signal-to-noise ratio is pre-set, and when the preset signal-to-noise ratio is reached, the next step is continued, and if it is not reached, the step of removing the noise is continued until the requirement is met, and the noise reduction model that meets the best signal-to-noise ratio is obtained; The machine learning module reads the optimal model parameters obtained by machine learning based on the noise reduction model that satisfies the best signal-to-noise ratio, inputs the collected signal to be denoised, performs machine learning processing, and sends it to each end of the call through the transmission module; Each party in the communication generates its own noise reduction model and outputs a speech signal with the best signal-to-noise ratio.

2. The intelligent noise reduction communication headset according to claim 1, characterized in that: The call microphone is connected to the microphone line, and the microphone line is protected by a metal hose and plastic and can be bent at will according to needs. The call microphone is located at the end of the microphone line.

3. The intelligent noise reduction communication headset according to claim 1, characterized in that: The digital signal processing system includes two major modules, namely a filtering module and a weight adaptive adjustment module; the filtering module determines the structure of the digital signal processing system, and the weight adaptive adjustment module adjusts the weight of each input signal vector, which is determined by the machine learning model parameters.

4. The intelligent noise reduction communication headset according to claim 3, characterized in that: The machine learning module provides input signals to the digital signal processing system, as well as the expected response and error signals generated during adjustment. The weighted delay operators are superimposed through the adder, and the weights are adjusted according to the model parameters to obtain the optimal response signal.

5. The intelligent noise reduction communication headset according to claim 1, characterized in that: The preprocessed noise is environmental noise or white noise.

Images