A noise reduction method for remote audio system

Abstract

The present invention provides a remote audio system noise reduction method. The method aims at solving the contradiction between the performance of the noise reduction method adopted in the prior art and the resource requirements (mainly the amount of computation and storage space). The present invention uses a multi-threshold wavelet packet decomposition noise reduction algorithm to sort the leaf nodes of the wavelet packet tree. Different leaf nodes use different threshold estimation algorithms to achieve voice noise reduction of the remote conference system. The amount of computation and storage space are moderate and can be implemented in a DSP chip.

Description

Technical Field

[0001] The present invention relates to the technical field of audio processing, and more specifically, to a noise reduction method for a remote audio system. Background Art

[0002] The audio part of a teleconference system generally includes microphones, audio processing equipment, power amplifiers and speakers, which are mainly used to play human voices and hope to have higher speech clarity. The audio signal involves many links, and each link may bring some noise. In order to make the conference system have higher speech clarity, it is often necessary to perform noise reduction processing on the signal collected by the microphone.

[0003] As early as the 1960s, noise elimination technology was used for preprocessing of sound signals. With the maturity of sound signal processing theory and the increase in practical applications, a research climax was formed in the 1970s, and some mature theories and basic results were successively achieved. Noise elimination technology has gradually developed into an important branch of sound signal processing. Spectral subtraction is a classic method. The basic principle of this method is to first calculate the power spectrum of the noisy sound, then use the noise estimation algorithm to estimate the power spectrum of the noise, and finally subtract the power spectrum of the noise from the power spectrum of the noisy sound to obtain the expected power spectrum of the pure sound. The spectral subtraction algorithm is simple and easy to implement, but after spectral subtraction, the noise-reduced sound will have a large residual part of the power spectrum component, resulting in residual noise similar to underwater recording with a certain rhythmic fluctuation. This residual noise is also called "music noise". Lim and Oppenheim proposed a noise elimination algorithm based on the Wiener filter. The Wiener filter uses the minimum mean square error criterion to minimize the square difference between the output of the contaminated sound signal after the filter and the pure sound. Its optimal estimation needs to be performed under stable conditions, but the actual sound and noise signals are generally non-stationary, so the application of the Wiener filter method is not very extensive. Ephraim and Malah proposed the minimum mean square error method (Minimum Mean Square Error, MMSE) based on short-time spectral amplitude (Short-Time Spectral Amplitude, STSA) estimation. This method is insensitive to specific distortion criteria and posterior probability, and can effectively reduce the interference of music noise. Since the human ear's perception of spectral intensity is proportional to the logarithm of the spectral amplitude, it is more appropriate to use logarithmic spectral amplitude estimation. Ephraim then proposed the minimum mean square error method based on short-time logarithmic spectral amplitude (Logarithmic Spectrum Amplitude, LSA) estimation. Ephraim and Trees proposed a noise removal method based on signal subspace decomposition. The basic idea is to map the noisy sound signal to the signal subspace and the noise subspace; then remove the noise subspace to achieve noise removal, and use the optimal constraint estimator to estimate the sound features to achieve sound enhancement. This method can not only suppress noise well, but also cause little distortion, so it has been rapidly developed. However, this type of method has high computational complexity and requires a lot of computing power. Since the late 1990s, the rapid development of digital signal processing has brought a new round of research on noise removal technology, and various improved forms of the original algorithms and new algorithms have emerged one after another. For example: sound enhancement algorithms based on statistical models, algorithms that use neural networks to achieve noise removal, and auditory masking effects have gradually become a research hotspot in recent years. However, the performance of noise reduction algorithms and resource requirements (mainly computational complexity and storage space) are often contradictory. For example, the classical spectral subtraction method requires few resources, but the noise reduction effect is poor.The algorithm using neural network to realize noise elimination can achieve satisfactory noise reduction effect, but it requires more resources. Summary of the invention

[0004] The present invention provides a remote audio system noise reduction method, which uses a multi-threshold wavelet packet decomposition noise reduction algorithm to sort the leaf nodes of the wavelet packet tree, and adopts different threshold estimation algorithms for different leaf nodes to achieve voice noise reduction of the remote conference system.

[0005] In order to achieve the above technical effects, the technical solution of the present invention is as follows:

[0006] A method for reducing noise in a remote audio system comprises the following steps:

[0007] S1: Find the upper and lower cutoff frequencies of the speaker frequency response curve, design low-pass and high-pass filters according to the upper and lower cutoff frequencies, limit the signal frequency between the upper and lower cutoff frequencies, and achieve frequency response curve matching between the speaker and the power amplifier;

[0008] S2: Flatten the acoustic frequency response curve of the remote audio system, increase the frequency band of 900-1100Hz by 1.414 times, and attenuate the frequency band below 200Hz by 6Db per octave, so as to achieve timbre balance of the sound signal;

[0009] S3: Add high-frequency harmonic components to the signal to enhance the sound brightness.

[0010] Furthermore, the specific process of step S1 is:

[0011] S11: Input the sound signal f(k) received by the microphone, assuming the length is N;

[0012] S12: Select wavelet basis function: use db4 wavelet with 4th order vanishing moment;

[0013] S13: Decompose the signal f(k) into several layers:

[0014]

[0015] S14: Perform J-layer wavelet decomposition on the signal f(k), extract J wavelet coefficients, and agree on the symbol w j,k Represents the kth wavelet coefficient of the jth layer, symbol wp j,k represents the kth wavelet packet coefficient of the jth layer;

[0016] S15: Calculate the standard deviation σ using the wavelet coefficients of each layer, where median() means taking the median of a given value:

[0017]

[0018] S16: Unified Threshold Estimation

[0019] Furthermore, the specific process of step S2 is:

[0020] S21: Maximum and minimum threshold estimation:

[0021] S22: SURE threshold estimation, assuming that the order of the sequence square of the wavelet packet coefficients is:

[0022] Define a risk vector, remember:

[0023] R=[r1,r2,...,r i ], i = 1, 2, ..., N;

[0024] in Let its minimum value be r min , and the corresponding threshold is obtained as:

[0025] S23: Threshold function adaptively adjusted according to the size of wavelet coefficients:

[0026]

[0027] The threshold function gives a method for thresholding each wavelet coefficient of the signal or each point of the wavelet packet coefficient sequence. After the threshold of the sequence is obtained using the threshold estimation algorithm, each point of the sequence is judged. When it is less than the threshold λ, it is regarded as a noise coefficient and directly cleared to zero. When it is greater than or equal to the threshold λ, the value of the sequence after thresholding can be calculated by the function.

[0028] S24: performing 5-layer wavelet packet decomposition on the signal f(k) of length N, and obtaining the coefficients of each frequency band by obtaining the leaf node coefficients of the wavelet packet tree;

[0029] S25: Solve the energy of the coefficient sequence of each wavelet packet tree leaf node. The formula is as follows: E j,k =∑|wp j,k | 2 ;

[0030] S26: Sort the nodes according to their energy, and select the k nodes with the lowest energy as noise nodes noise , find the sum of their energies, denoted as E noise .

[0031] Furthermore, the specific process of step S3 is:

[0032] S31: Classify nodes according to energy, greater than E noiseThe node is recorded as the signal node Node signal , less than E noise The transition node is recorded as Node transit , except that 1 / 4 of the nodes with lower node energy are regarded as noise nodes, 1 / 2 are classified as transition nodes, and 1 / 4 are classified as signal nodes;

[0033] S32: Calculate the standard deviation σ for the noise nodes using the following formula:

[0034]

[0035] S33: performing threshold estimation according to step S16;

[0036] S34: performing threshold estimation on the transition node according to step S21;

[0037] S35: performing threshold estimation on the signal node according to step S22;

[0038] S36: Perform threshold processing on each node using the estimated threshold value to obtain processed wavelet packet coefficients and index values;

[0039] S37: Reconstruct the wavelet packet using the index value to obtain a processed signal.

[0040] Compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0041] The present invention aims at the contradiction between the performance of the noise reduction method adopted by the existing technical solution and the resource requirements (mainly the amount of calculation and storage space). The present invention will use a multi-threshold wavelet packet decomposition noise reduction algorithm to sort the leaf nodes of the wavelet packet tree. Different leaf nodes use different threshold estimation algorithms to achieve voice noise reduction in the remote conferencing system. The amount of calculation and storage space are moderate and can be implemented in a DSP chip. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The figure is a flow chart of the method of the present invention. DETAILED DESCRIPTION

[0043] The drawings are for illustrative purposes only and should not be construed as limiting the present patent;

[0044] In order to better illustrate the present embodiment, some parts in the drawings may be omitted, enlarged or reduced, and do not represent the size of the actual product;

[0045] It is understandable to those skilled in the art that some well-known structures and their descriptions may be omitted in the drawings.

[0046] The technical solution of the present invention is further described below in conjunction with the accompanying drawings and embodiments.

[0047] like Figure 1 As shown, a method for reducing noise in a remote audio system comprises the following steps:

[0048] S1: Find the upper and lower cutoff frequencies of the speaker frequency response curve, design low-pass and high-pass filters according to the upper and lower cutoff frequencies, limit the signal frequency between the upper and lower cutoff frequencies, and achieve frequency response curve matching between the speaker and the power amplifier (corresponding to steps 1-6);

[0049] S2: Flatten the acoustic frequency response curve of the conference system, increase the frequency band of 900-1100Hz by 1.414 times, and attenuate the frequency band below 200Hz by 6dB per octave, so as to achieve timbre balance of the sound signal (corresponding to steps 7-12);

[0050] S3: Add high-frequency harmonic components to the signal to enhance the sound brightness (corresponding to steps 13-19.

[0051] Specific process:

[0052] 1. Input the sound signal f(k) received by the microphone, and set its length to N.

[0053] 2. Select the wavelet basis function and use the db4 wavelet with 4th-order vanishing moment.

[0054] 3. Decompose the signal f(k) into layers:

[0055]

[0056] 4. Perform J-layer wavelet decomposition on the signal f(k), extract J wavelet coefficients, and agree on the symbol w j,k Represents the kth wavelet coefficient of the jth layer, symbol wp j,k represents the kth wavelet packet coefficient of the jth layer.

[0057] 5. Use the wavelet coefficients of each layer to calculate the standard deviation σ, where median() means taking the median of a given value:

[0058]

[0059] 6. Unified Threshold Estimation

[0060] 7. Minimax threshold estimation (MiniMaxi):

[0061]

[0062] 8.URE (Stein Unbiased Risk Estimator) Threshold Estimation

[0063] Assume that the order of the sequence square of the wavelet packet coefficients is:

[0064]

[0065] Define a risk vector, remember

[0066] R=[r1,r2,...,r i ],i=1,2,...,N

[0067] in

[0068] Assume that its minimum value is r min , and the corresponding threshold is obtained as:

[0069] 9. Threshold function adaptively adjusted according to the size of wavelet coefficients

[0070]

[0071] The threshold function is a method of performing threshold processing on each wavelet coefficient of a given signal or each point of a wavelet packet coefficient sequence. It can be understood that after the threshold of the sequence is obtained using the threshold estimation algorithm, each point of the sequence is judged. When it is less than the threshold λ, it is regarded as a noise coefficient and directly cleared to zero. When it is greater than or equal to the threshold λ, the value of the sequence after thresholding can be calculated through the function.

[0072] 10. The signal f(k) of length N is decomposed into 5 layers of wavelet packets, and the coefficients of each frequency band are obtained by obtaining the leaf node coefficients of the wavelet packet tree.

[0073] 11. Solve the energy of each wavelet packet tree leaf node coefficient sequence, the formula is as follows:

[0074] E j,k =∑|wp j,k | 2

[0075] 12. Sort the nodes according to their energy, and select the k nodes with the lowest energy as noise nodes noise , find the sum of their energies, denoted as E noise .

[0076] 13. Nodes are classified according to energy, greater than E noise The node is recorded as the signal node Node signal , less than E noise The transition node is recorded as Node transit , except that 1 / 4 of the nodes with lower node energy are regarded as noise nodes, 1 / 2 are classified as transition nodes, and 1 / 4 are classified as signal nodes.

[0077] 14. The standard deviation σ of the noise node is calculated using the following formula:

[0078]

[0079] 15. Estimate the threshold according to step 6.

[0080] 16. Estimate the threshold of the transition node according to step 7.

[0081] 17. Signal node step 8 performs threshold estimation

[0082] 18. Use the estimated threshold to perform threshold processing on each node to obtain the processed wavelet packet coefficients and index values.

[0083] 19. Use the index value to reconstruct the wavelet packet and obtain the processed signal.

[0084] The same or similar reference numerals correspond to the same or similar components;

[0085] The positional relationships described in the drawings are only for illustrative purposes and should not be construed as limiting the present patent.

[0086] Obviously, the above embodiments of the present invention are merely examples for clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. For those skilled in the art, other different forms of changes or modifications can be made based on the above description. It is not necessary and impossible to list all the embodiments here. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included in the protection scope of the claims of the present invention.

Claims

1. A remote audio system noise reduction method, characterized in that: The following steps are involved: S1: Find the upper and lower cutoff frequencies of the speaker frequency response curve, design low-pass and high-pass filters according to the upper and lower cutoff frequencies, limit the signal frequency between the upper and lower cutoff frequencies, and achieve frequency response curve matching between the speaker and the power amplifier; S2: Flatten the acoustic frequency response curve of the remote audio system, increase the frequency band of 900-1100Hz by 1.414 times, and attenuate the frequency band below 200Hz by 6Db per octave, so as to achieve timbre balance of the sound signal; S3: Add high-frequency harmonic components to the signal to enhance the brightness of the sound; The specific process of step S1 is: S11: Input the sound signal f(k) received by the microphone, assuming the length is N; S12: Select wavelet basis function; S13: Decompose the signal f(k) into several layers: S14: Perform J-layer wavelet decomposition on the signal f(k), extract J wavelet coefficients, and agree on the symbol w j,k Represents the kth wavelet coefficient of the jth layer, symbol wp j,k represents the kth wavelet packet coefficient of the jth layer; S15: Calculate the standard deviation σ using the wavelet coefficients of each layer, where median() means taking the median of a given value: S16: Unified Threshold Estimation The db4 wavelet with 4th order vanishing moment is used; The specific process of step S2 is: S21: Maximum and minimum threshold estimation: S22: SURE threshold estimation, assuming that the order of the sequence square of the wavelet packet coefficients is: W n 2 =[W1 2 ,W2 2 ,...,W N 2 ],W1 2 ≤W2 2 ≤...≤W N 2 , define a risk vector, remember: R=[r1,r2,...,r i ],i=1,2,…,N; in Let its minimum value be r min , and the corresponding threshold is obtained as: S23: Threshold function adaptively adjusted according to the size of wavelet coefficients: S24: performing 5-layer wavelet packet decomposition on the signal f(k) of length N, and obtaining the coefficients of each frequency band by obtaining the leaf node coefficients of the wavelet packet tree; S25: Solve the energy of the coefficient sequence of each wavelet packet tree leaf node. The formula is as follows: E j,k =Σ|wp j,k | 2 ; S26: Sort the nodes according to their energy, and select the k nodes with the lowest energy as noise nodes noise , find the sum of their energies, denoted as E noise .

2. The remote audio system noise reduction method according to claim 1, characterized in that: The threshold function gives a method for thresholding each wavelet coefficient of the signal or each point of the wavelet packet coefficient sequence. After the threshold of the sequence is obtained using the threshold estimation algorithm, each point of the sequence is judged. When it is less than the threshold λ, it is regarded as a noise coefficient and directly cleared to zero. When it is greater than or equal to the threshold λ, the value of the sequence after thresholding can be calculated through the function.

3. The remote audio system noise reduction method according to claim 2, characterized in that: The specific process of step S3 is: S31: Classify nodes according to energy, greater than E noise The node is recorded as the signal node Node signal , less than E noise The transition node is recorded as Node transit ; S32: Calculate the standard deviation σ for the noise nodes using the following formula: S33: performing threshold estimation according to step S16; S34: performing threshold estimation on the transition node according to step S21; S35: performing threshold estimation on the signal node according to step S22; S36: Perform threshold processing on each node using the estimated threshold value to obtain processed wavelet packet coefficients and index values; S37: Reconstruct the wavelet packet using the index value to obtain a processed signal.

4. The remote audio system noise reduction method according to claim 3, characterized in that: In addition to the 1 / 4 nodes with lower node energy being treated as noise nodes, 1 / 2 are classified as transition nodes, and 1 / 4 are classified as signal nodes.

Images