Noise reduction method for helicopter voice processing system

Abstract

The invention provides a noise reduction method for a voice processing system of a helicopter, which comprises the following steps: firstly, inputting an audio signal, inputting the audio signal into window noise estimation for preset caching, setting the size of a window to be between 10ms and 30ms, determining that the voice with the time length is constant, and when the input signal is smaller than a threshold value theta, determining that the voice with the time length is not changed; when the input signal is larger than a threshold theta value, the input signal is considered as a noise signal, when the input signal is larger than the threshold theta value, the input signal is considered as a real-time voice signal, a useful signal can be obtained by subtracting a noise signal frequency spectrum from the real-time voice signal, and according to the logic, a prior pure noise lambda x (k) and a useful signal lambda d (k) are obtained; according to the noise reduction method for the voice processing system of the helicopter provided by the invention, the Wiener filtering algorithm is operated inside, so that the noise intensity in the collected voice is reduced, the vibration of a rotor wing, the vibration of a helicopter body, the noise of an engine and the like are suppressed, and the voice heard in an earphone is ensured to be clear and free of noise.

Description

Technical Field

[0001] The invention relates to the field of speech processing, and in particular to a noise reduction method for a helicopter speech processing system. Background Art

[0002] A large amount of noise is generated when a helicopter is driving or flying, including noise generated by the engine running, noise generated by the vibration of the fuselage, and specific noise generated by the rotor rotating and cutting the air. The intensity of these noises is as high as over 110dB and will enter the voice processing system through the microphone of the helmet.

[0003] At present, traditional speech noise reduction methods have certain limitations in helicopter speech processing systems. For example, simple filtering methods will lose part of the speech signal while removing noise, resulting in speech distortion. Some noise reduction methods based on statistical models often find it difficult to accurately estimate noise characteristics when dealing with complex helicopter noise environments, thus affecting the noise reduction effect.

[0004] In addition, the flight environment of helicopters is complex and changeable, and the intensity and frequency characteristics of noise will also change with the change of flight status.

[0005] Therefore, it is necessary to provide a noise reduction method for a helicopter voice processing system to solve the above technical problems. Summary of the invention

[0006] The present invention provides a noise reduction method for a helicopter voice processing system, which solves the problem that different noises are generated in the helicopter cabin during flight, and the noises enter the voice processing system from the microphone of the helmet and finally reach the earphone end, affecting the call effect.

[0007] In order to solve the above technical problems, the present invention provides a noise reduction method for a helicopter voice processing system, comprising the following steps:

[0008] S1. First, input the audio signal, and input the audio signal into the preset buffer in the window noise estimation. The window size should be set between 10ms and 30ms. The long voice is considered to be constant.

[0009] S2. When the input signal is less than the threshold θ value, the signal input at this time is considered to be a noise signal. When the input signal is greater than the threshold θ value, the input signal is considered to be a real-time voice signal. The useful signal can be obtained by subtracting the noise signal spectrum from the real-time voice signal.

[0010] S3. According to this logic, we get the prior pure noise λ x (k), and the useful signal λ d (k), and calculate the signal-to-noise ratio at this time

[0011] S4, replace λ in S3 x (k), λ d (k), ξ k Stored in the memory chip as initial parameters for subsequent MMSE-STSA algorithm and Wiener filter algorithm to call;

[0012] S5. At this time, according to the default function of the Wiener filter in S3, the Wiener filter formula Call out the window noise estimation function to feed ξ k For input signal X k Filtering is performed and the filtered data is defined as Send to MMSE-STSA algorithm;

[0013] S6, comparison If these two values ​​are consistent, data convergence is complete; if they are inconsistent, Substitute into the formula and calculate ξ k , and k Send it to the Wiener filter and update the filtering algorithm H(ω k );

[0014] S7. Finally, by analogy, the Wiener filter is continuously updated until convergence is achieved.

[0015] Preferably, S4 comprises the following steps:

[0016] S41. Run the MMSE-STSA algorithm and obtain the corresponding formula based on the principle of least mean square estimation:

[0017]

[0018] S42, where the prior noise signal-to-noise ratio is the ξ stored in the chip in the early stage k , Y k For the voice signal collected at this time, the a posteriori signal-to-noise ratio It is known that according to the formula We can calculate v k , I0, I1 are Bessel functions, from which we can calculate The value of .

[0019] Preferably, the value in S42 is the value of the useful signal estimated according to the least mean square estimation principle, which is defined as

[0020] Preferably, the window noise estimation in S1 includes the following method:

[0021] S11, noise estimation algorithm for statistical models;

[0022] S12, noise estimation algorithm of signal subspace;

[0023] S13, deep learning noise estimation algorithm;

[0024] S14. A hybrid noise estimation algorithm combining multiple algorithms.

[0025] Preferably, the noise estimation algorithm of the statistical model in S11 comprises the following steps:

[0026] S111, minimum statistics tracking algorithm;

[0027] S112, recursive averaging algorithm.

[0028] Preferably, the noise estimation algorithm of the signal subspace in S12 comprises the following steps:

[0029] S121, eigenvalue decomposition algorithm;

[0030] S122. Singular value decomposition algorithm.

[0031] Preferably, the deep learning noise estimation algorithm in S13 comprises the following steps:

[0032] S131, deep neural network;

[0033] S132. Convolutional neural network.

[0034] Preferably, the hybrid noise estimation algorithm combining multiple algorithms in S14 comprises the following steps:

[0035] S141, Combining statistical models with deep learning;

[0036] S142. Multi-microphone noise estimation.

[0037] Preferably, the statistical model in S141 is combined with deep learning to combine the algorithm based on the statistical model with the deep learning algorithm to give full play to the advantages of both.

[0038] Preferably, the multi-microphone noise estimation in S142 can utilize multiple microphones on the helicopter to collect signals, and fuse the information of the multiple microphones through a signal processing algorithm to improve the accuracy of noise estimation.

[0039] Compared with the related art, the noise reduction method for a helicopter voice processing system provided by the present invention has the following beneficial effects:

[0040] The present invention provides a noise reduction method for a helicopter voice processing system, which reduces the noise intensity in the collected voice by running the Wiener filtering algorithm internally, suppresses the noise of the rotor, body vibration, and engine, and ensures that the voice heard in the earphone is clear and free of noise;

[0041] By running the MMSE-STSA algorithm, the short-time spectrum amplitude estimation value of the noise is obtained, and this parameter is used as the basic parameter of the Wiener algorithm to improve the convergence speed of the Wiener filter algorithm, ensure that the Wiener filter algorithm can converge quickly, and improve the response speed of the algorithm. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 A structural schematic diagram of a first embodiment of a noise reduction method for a helicopter speech processing system provided by the present invention;

[0043] Figure 2 This is a schematic diagram of the hardware principle;

[0044] Figure 3 This is a schematic diagram of the hardware layout;

[0045] Figure 4 This is a hardware schematic diagram with CS320AIC3106 as the core device;

[0046] Figure 5 This is a hardware schematic diagram with JTLV320AC23 as the core device;

[0047] Figure 6 It is a schematic diagram of the software processing process;

[0048] Figure 7 It is a schematic diagram of the hardware circuit;

[0049] Figure 8 This is a structural schematic diagram of a second embodiment of a noise reduction method for a helicopter speech processing system provided by the present invention. DETAILED DESCRIPTION

[0050] The present invention will be further described below in conjunction with the accompanying drawings and implementation modes.

[0051] First embodiment

[0052] Please refer to Figure 1 , Figure 2 , Figure 3 , Figure 4 , Figure 5 , Figure 6 and Figure 7 ,in, Figure 1 A structural schematic diagram of a first embodiment of a noise reduction method for a helicopter speech processing system provided by the present invention; Figure 2 This is a schematic diagram of the hardware principle; Figure 3 This is a schematic diagram of the hardware layout; Figure 4 This is a hardware schematic diagram with CS320AIC3106 as the core device; Figure 5 This is a hardware schematic diagram with JTLV320AC23 as the core device; Figure 6 It is a schematic diagram of the software processing process; Figure 7 A noise reduction method for a helicopter voice processing system comprises the following steps:

[0053] S1. First, input the audio signal, and input the audio signal into the preset buffer in the window noise estimation. The window size should be set between 10ms and 30ms. The long voice is considered to be constant.

[0054] S2. When the input signal is less than the threshold θ value, the signal input at this time is considered to be a noise signal. When the input signal is greater than the threshold θ value, the input signal is considered to be a real-time voice signal. The useful signal can be obtained by subtracting the noise signal spectrum from the real-time voice signal.

[0055] S3. According to this logic, we get the prior pure noise λ x (k), and the useful signal λ d (k), and calculate the signal-to-noise ratio at this time

[0056] S4, replace λ in S3 x (k), λ d (k), ξ k Stored in the memory chip as initial parameters for subsequent MMSE-STSA algorithm and Wiener filter algorithm to call;

[0057] S5. At this time, according to the default function of the Wiener filter in S3, the Wiener filter formula Call out the window noise estimation function and input ξ k For input signal X k Filtering is performed and the filtered data is defined as Send to MMSE-STSA algorithm;

[0058] S6, comparison If these two values ​​are consistent, data convergence is complete; if they are inconsistent, Substitute into the formula and calculate ξ k , and k Send it to the Wiener filter and update the filtering algorithm H(ω k );

[0059] S7. Finally, by analogy, the Wiener filter is continuously updated until convergence is achieved.

[0060] The S4 comprises the following steps:

[0061] S41. Run the MMSE-STSA algorithm and obtain the corresponding formula based on the principle of least mean square estimation:

[0062]

[0063] S42, where the prior noise signal-to-noise ratio is the ξ stored in the chip in the early stage k , Y k For the voice signal collected at this time, the a posteriori signal-to-noise ratio It is known that according to the formula We can calculate v k , I0, I1 are Bessel functions, from which we can calculate The value of .

[0064] This value in S42 is the value of the useful signal estimated according to the least mean square estimation principle, and is defined as

[0065] The digital voice automatic gain control method gets rid of the dependence on imported automatic gain control circuits, makes up for the problem that the range of domestic analog automatic gain circuits is not wide enough, simplifies the whole circuit, saves hardware costs, and can be flexibly configured in digital circuits, improving configuration efficiency. At the same time, it mainly has the following aspects:

[0066] reliability:

[0067] Digital voice automatic gain control reduces the hardware circuit of the whole machine and can be flexibly configured to each port, thus improving the reliability of the whole machine;

[0068] Analog automatic gain control circuit: increases the number of hardware circuits in the whole machine and reduces the reliability of the whole machine.

[0069] Economic benefits:

[0070] The digital voice automatic gain control method does not increase components, so the hardware cost of the whole machine is reduced;

[0071] Analog automatic gain control circuit: increases the number of hardware circuits in the whole machine and increases the cost of components.

[0072] Please refer to Figure 6 It is found that: considering the bandwidth of 300-3400Hz, the center frequency f of the speech signal is determined c =1850Hz. Since the purpose of this method is to realize automatic gain control of the speech signal amplitude, it is only necessary to detect the peak value of the signal. There is no need to judge the voiced and unvoiced sounds in the speech. The detected envelope signal does not need to be very precise. Therefore, the Hilbert transform method can be used. The formula is: The original signal is x(t), and the Hilbert transform of the original signal is The analytical signal z(t) is obtained, and the envelope signal of the speech signal can be obtained by modulo the analytical signal, wherein the baseband frequency parameter is set between 0.1 and 0.5 Hz and adjusted according to the required accuracy.

[0073] After obtaining the peak value of the speech signal, the peak value of the speech signal is compared with the corresponding peak value in the gain table. For example, when the first circuit is used, the suppression parameter in the gain table is set to 3000 when speech is input, and when receiving a digital signal, the suppression parameter in the gain table is set to 5000. When the second circuit is used, the suppression parameter in the gain table is set to 6000, and when receiving a digital signal, the suppression parameter in the gain table is set to 10000.

[0074] The volume of the signal is proportionally reduced according to the coefficient in the gain table, and this volume coefficient is assigned to the voice signal in the buffer. The signal output after the voice signal is multiplied by this coefficient is the voice signal after automatic gain adjustment.

[0075] Please refer to Figure 4 Among them, 3D1 is the domestic ARM chip HWD32F767, 3D24 is the domestic 422 bus chip HWD488, 3D2 is the domestic voice acquisition chip CS320AIC3106, 3C123, 3C124, 3C66, 3C67 are coupling capacitors, 3C305 is a bypass filter capacitor, 3R228 is a pull-up resistor, and 3R51, 3R52, 3R210 are matching resistors.

[0076] Please refer to Figure 5 3D30 is the domestic ARM chip FMQL20S484, D42 and D6 are the domestic 422 bus transceiver chips XZJ26C32 and XZJ26C31 respectively, 3D2 is the domestic voice acquisition chip CS320AIC3106, C471, C474, C477, C478 are coupling capacitors, C334, C335, C336, C337, C287, C96, C72, C69 , C75, C70 are bypass filter capacitors, R223, R227 are pull-up resistors, R204, R206, R225, R229, R415, R417 are pull-down resistors, R21, R30, R31, R34, R164, R165, R516, R166, R167 are matching resistors, V7, V8, V11, V13, V71, V74, V75, V78 are anti-static diodes.

[0077] CS320AIC3106 is the circuit of the acquisition chip, HWD488 is the 422 bus communication interface, HWD32F767 is the circuit of the processor. This circuit requires the input signal size to be 100~120mV. The second type uses JTLV320AC23 as the acquisition chip, XZJ26C32 and XZJ26C31 as the 422 bus communication interface, and FMQL20S484 as the circuit of the processor. This circuit requires the input signal size to be 200~240mV. The hardware circuit is shown in the figure. The I2S interface is used for communication between chips.

[0078] Compared with the related art, the noise reduction method for a helicopter voice processing system provided by the present invention has the following beneficial effects:

[0079] The present invention provides a noise reduction method for a helicopter voice processing system, which reduces the noise intensity in the collected voice by running the Wiener filter algorithm internally, suppresses the noise of the rotor, body vibration, engine, etc., and ensures that the voice heard in the earphone is clear and free of noise;

[0080] By running the MMSE-STSA algorithm, the short-time spectrum amplitude estimation value of the noise is obtained, and this parameter is used as the basic parameter of the Wiener algorithm to improve the convergence speed of the Wiener filter algorithm, ensure that the Wiener filter algorithm can converge quickly, and improve the response speed of the algorithm.

[0081] Second embodiment

[0082] Please refer to Figure 8 Based on the noise reduction method for a helicopter voice processing system provided in the first embodiment of the present application, the second embodiment of the present application proposes another noise reduction method for a helicopter voice processing system. The second embodiment is only a preferred embodiment of the first embodiment, and the implementation of the second embodiment will not affect the independent implementation of the first embodiment.

[0083] Specifically, the second embodiment of the present application provides a noise reduction method for a helicopter voice processing system. The difference is that, in a noise reduction method for a helicopter voice processing system, the window noise estimation in S1 includes the following method:

[0084] S11, noise estimation algorithm for statistical models;

[0085] S12, noise estimation algorithm of signal subspace;

[0086] S13, deep learning noise estimation algorithm;

[0087] S14. A hybrid noise estimation algorithm combining multiple algorithms.

[0088] The noise estimation algorithm of the statistical model in S11 comprises the following steps:

[0089] S111, minimum statistics tracking algorithm;

[0090] S112, recursive averaging algorithm.

[0091] Minimum Statistics Tracking Algorithm: This algorithm estimates the noise power by tracking the minimum value of the signal power. In a helicopter environment, due to the non-stationary nature of the noise, the minimum statistics tracking algorithm can quickly adapt to noise changes and provide a more accurate noise estimate.

[0092] Recursive averaging algorithm: The input signal is recursively averaged to estimate the average power of the noise. This algorithm is simple to calculate and suitable for real-time processing, but it may not be accurate enough for rapidly changing noise.

[0093] The noise estimation algorithm of the signal subspace in S12 comprises the following steps:

[0094] S121, eigenvalue decomposition algorithm;

[0095] S122. Singular value decomposition algorithm.

[0096] Eigenvalue decomposition algorithm: Perform eigenvalue decomposition on the covariance matrix of the signal and divide the signal into signal subspace and noise subspace. By estimating the eigenvalue of the noise subspace, the noise power estimate can be obtained. In helicopter speech processing, this algorithm can effectively remove noise that is not related to the speech signal.

[0097] Singular value decomposition algorithm: Similar to eigenvalue decomposition, the singular value decomposition algorithm decomposes the signal matrix to obtain estimates of the signal and noise. This algorithm has good performance when processing high-dimensional data.

[0098] The deep learning noise estimation algorithm in S13 comprises the following steps:

[0099] S131, deep neural network;

[0100] S132. Convolutional neural network.

[0101] Deep neural network: Noise estimation is achieved by training DNN to learn the characteristics of noise. A large amount of noisy speech data and corresponding clean speech data can be used for training to improve the accuracy and robustness of the algorithm. In a helicopter environment, DNN can learn different types of noise characteristics and adapt to complex noise environments.

[0102] Convolutional Neural Network: CNN is used to process the time-frequency characteristics of speech signals and estimate noise power. CNN has powerful feature extraction capabilities and can effectively capture the local and global characteristics of noise.

[0103] The hybrid noise estimation algorithm combining multiple algorithms in S14 comprises the following steps:

[0104] S141, Combining statistical models with deep learning;

[0105] S142. Multi-microphone noise estimation.

[0106] The combination of the statistical model and deep learning in S141 can combine the algorithm based on the statistical model with the deep learning algorithm to give full play to the advantages of both.

[0107] A statistical model algorithm can be used to perform preliminary noise estimation, and then the estimation result can be used as input to the deep learning algorithm to further improve the accuracy of the noise estimation.

[0108] The multi-microphone noise estimation in S142 can collect signals using multiple microphones on the helicopter, and fuse the information of the multiple microphones through a signal processing algorithm to improve the accuracy of noise estimation.

[0109] Beamforming technology can be used to weight the signals from multiple microphones to enhance the speech signal and suppress noise.

[0110] Compared with the related art, the noise reduction method for a helicopter voice processing system provided by the present invention has the following beneficial effects:

[0111] The present invention provides a noise reduction method for a helicopter speech processing system. The use of a noise estimation algorithm of a statistical model, a noise estimation algorithm of a signal subspace, a noise estimation algorithm of a deep learning, and a hybrid noise estimation algorithm combining multiple algorithms in a window estimation noise algorithm can enhance the overall adaptability, simplify the calculation, have good stability, achieve significant denoising effect, have strong processing capability for complex noise, and have high robustness.

[0112] The above descriptions are merely embodiments of the present invention and are not intended to limit the patent scope of the present invention. Any equivalent structure or equivalent process transformation made using the contents of the present invention specification and drawings, or directly or indirectly applied in other related technical fields, are also included in the patent protection scope of the present invention.

Claims

1. A noise reduction method for a helicopter speech processing system, characterized in that: The following steps are involved: S1. First, input the audio signal, and input the audio signal into the preset buffer in the window noise estimation. The window size should be set between 10ms and 30ms. The long voice is considered to be constant. S2. When the input signal is less than the threshold θ value, the signal input at this time is considered to be a noise signal. When the input signal is greater than the threshold θ value, the input signal is considered to be a real-time voice signal. The useful signal can be obtained by subtracting the noise signal spectrum from the real-time voice signal. S3. According to this logic, we get the prior pure noise λ x (k), and the useful signal λ d (k), and calculate the signal-to-noise ratio at this time S4, replace λ in S3 x (k), λ d (k), ξ k Stored in the memory chip as initial parameters for subsequent MMSE-STSA algorithm and Wiener filter algorithm to call; S5. At this time, according to the default function of the Wiener filter in S3, the Wiener filter formula Call out the window noise estimation function and input ξ k For input signal X k Filtering is performed and the filtered data is defined as Send to MMSE-STSA algorithm; S6, comparison If these two values ​​are consistent, data convergence is complete; if they are inconsistent, Substitute into the formula and calculate ξ k , and k Send it to the Wiener filter and update the filtering algorithm H(ω k ); S7. Finally, by analogy, the Wiener filter is continuously updated until convergence is achieved.

2. The noise reduction method for a helicopter speech processing system according to claim 1, characterized in that: The S4 comprises the following steps: S41. Run the MMSE-STSA algorithm and obtain the corresponding formula based on the principle of least mean square estimation: S42, where the prior noise signal-to-noise ratio is the ξ stored in the chip in the early stage k , Y k For the voice signal collected at this time, the a posteriori signal-to-noise ratio It is known that according to the formula We can calculate v k , I0, I1 are Bessel functions, from which we can calculate The value of .

3. The noise reduction method for a helicopter speech processing system according to claim 2, characterized in that: This value in S42 is the value of the useful signal estimated according to the least mean square estimation principle, and is defined as 4. The noise reduction method for a helicopter voice processing system according to claim 1, characterized in that: The window noise estimation in S1 includes the following methods: S11, noise estimation algorithm for statistical models; S12, noise estimation algorithm of signal subspace; S13, deep learning noise estimation algorithm; S14. A hybrid noise estimation algorithm combining multiple algorithms.

5. The noise reduction method for a helicopter voice processing system according to claim 4, characterized in that: The noise estimation algorithm of the statistical model in S11 comprises the following steps: S111, minimum statistics tracking algorithm; S112, recursive averaging algorithm.

6. The noise reduction method for a helicopter speech processing system according to claim 4, characterized in that: The noise estimation algorithm of the signal subspace in S12 comprises the following steps: S121, eigenvalue decomposition algorithm; S122. Singular value decomposition algorithm.

7. The noise reduction method for a helicopter speech processing system according to claim 4, characterized in that: The deep learning noise estimation algorithm in S13 comprises the following steps: S131, deep neural network; S132. Convolutional neural network.

8. The noise reduction method for a helicopter speech processing system according to claim 4, characterized in that: The hybrid noise estimation algorithm combining multiple algorithms in S14 comprises the following steps: S141, Combining statistical models with deep learning; S142. Multi-microphone noise estimation.

9. The noise reduction method for a helicopter speech processing system according to claim 1, characterized in that: The combination of the statistical model and deep learning in S141 can combine the algorithm based on the statistical model with the deep learning algorithm to give full play to the advantages of both.

10. The noise reduction method for a helicopter speech processing system according to claim 1, characterized in that: The multi-microphone noise estimation in S142 can collect signals using multiple microphones on the helicopter, and fuse the information of the multiple microphones through a signal processing algorithm to improve the accuracy of noise estimation.

Images