Self-adaptive dual-microphone noise reduction method

Abstract

The invention discloses a self-adaptive dual-microphone noise reduction method, which comprises the following steps of S10, detecting a stable noise segment, and calculating a characteristic value of a current frame of a main microphone signal m1; stable noise segment detection is carried out based on a characteristic value mode, and a characteristic value baseline value and a characteristic value threshold value are dynamically updated; s20, performing noise tracking on the stable noise segment based on the minimum differential microphone array output power, and performing real-time updating on the delay d of the auxiliary microphone signal m2 to obtain a delay gradual change result Dsm; s30, performing primary noise reduction and secondary noise reduction on the basis of the delay gradual change result Dsm, and outputting the result; according to the adaptive dual-microphone noise reduction method provided by the invention, real-time computing resource consumption is reduced, and meanwhile, instant, efficient and deep noise reduction can be realized.

Description

Technical Field

[0001] The present invention belongs to the technical field of signal processing, and in particular relates to an adaptive dual-microphone noise reduction method. Background Art

[0002] For hearing-impaired people wearing hearing aids or cochlear implants, their speech recognition rate will drop significantly once they are in a noisy environment, while the call quality of normal people is also greatly affected by the surrounding noisy environment. An effective solution is to use speech enhancement technology to perform digital signal processing on the noisy speech signal collected by the device microphone to reduce noise, improve the speech signal-to-noise ratio, and thus improve the speech recognition rate.

[0003] According to the number of microphone channels, speech enhancement technology can be divided into single-microphone enhancement technology and microphone array enhancement technology. The latter can selectively enhance or suppress signals in a specific direction due to the introduction of spatial information, and can achieve good noise reduction effects when the signal source and the noise source are spatially separated. Generally, the more microphones there are, the stronger the directivity that can be achieved and the better the noise reduction effect, but the cost and the requirements for device size also increase accordingly. In order to take into account the applicability and directional noise reduction requirements, noise reduction technology based on dual microphones has received widespread attention in the application of small devices. Among them, the first-order differential microphone array (Differential Microphone Arrays, DMAs) technology array is compactly placed, has a frequency-invariant beam pattern, and can obtain the maximum directional gain when the number of array elements is given. Therefore, it is suitable for broadband speech signal processing of small devices.

[0004] Directional microphone noise reduction technology can be divided into two categories according to the noise reduction angle: directional noise reduction and adaptive directional noise reduction. The former can eliminate the noise from a preset direction (fixed beam), while the latter can calculate the sound source orientation of the target signal and noise based on the input signal, adaptively attenuate the azimuth signal with strong noise energy, and retain the target signal (adaptive beam). Obviously, the latter has a wider range of applications in daily use, and there is no need for users to manually switch the noise suppression angle according to the scene. However, this angle estimation (sound source localization) and adaptive control will obviously increase the amount of calculation, which is a technical problem that needs to be solved urgently for some edge devices with low power consumption requirements and small memory and computing power. Commonly used sound source localization algorithms such as generalized cross-correlation-steered-response power (Generalized Cross-Correlation-Steered-Response Power, GCC-SRP), multiple signal classification (MUltiple Signal Classification, MUSIC) and phase transform weighted controllable response power (Steered Response Power-PHAse Transform, SRP-PHAT) mostly involve grid search and have a large amount of calculation.

[0005] In summary, the existing adaptive directional noise reduction algorithms are based on the estimation of the direction of arrival (DOA) of the signal and noise sources, and most DOA estimation algorithms such as SRP-PHAT and MUSIC require grid search. In order to improve the accuracy of angle calculation, it is necessary to increase the spatial angle of the search, which will inevitably lead to an increase in the amount of computation and bring a burden to edge devices with low power requirements and low computing power. Summary of the invention

[0006] In order to solve the above problems, the technical solution of the present invention is as follows: an adaptive dual-microphone noise reduction method, comprising the following steps:

[0007] S10: Stable noise segment detection, calculate the characteristic value of the current frame of the main microphone signal m1; if the characteristic value of the current frame is lower than the characteristic value threshold, the current frame is determined to be a noise frame, and 1 is added to the frame count; when the frame count is greater than or equal to NoiseCounter, it is determined to enter the stable noise segment; if the characteristic value of the current frame is higher than the characteristic value threshold, NoiseCounter is set to zero, and it is determined not to belong to the stable noise segment;

[0008] Dynamically update the characteristic value baseline value and characteristic value threshold;

[0009] S20: performing noise tracking on the stable noise segment based on minimizing the differential microphone array output power, and updating the delay d of the secondary microphone signal m2 in real time to obtain a delay slow-changing result D_sm;

[0010] S30: Based on the delay slow-changing result D_sm, perform first-level noise reduction to obtain frequency domain DMA output;

[0011] S40: Output after secondary noise reduction.

[0012] Preferably, the eigenvalues ​​include but are not limited to energy entropy ratio estimation, energy zero ratio, autocorrelation function-based, sub-band variance-based, spectral distance-based, cepstrum distance, and spectrum entropy eigenvalues.

[0013] Preferably, in the stable noise segment detection of step S10, the specific steps of calculating the estimated value of the energy entropy ratio are as follows:

[0014] S101: Calculate the energy entropy ratio, perform windowing and framing processing and fast Fourier transform on the main microphone signal m1, and the calculation formula for the spectrum entropy of the i-th frame data is:

[0015]

[0016] Where i represents the i-th frame, N represents the frame length of the data interception, k represents the k-th frequency component, Pro i,k represents the normalized spectral probability density function of each frequency component, which is defined as:

[0017]

[0018] Among them, P i,K represents the energy of the kth frequency component in the i-th frame;

[0019] The short-time average energy is calculated, and the calculation formula of the short-time average energy is:

[0020] The energy entropy ratio EER i The calculation formula is:

[0021] S102: Calculate the estimated value of the energy entropy ratio, and use a decision-oriented method to estimate the calculation result of the current energy entropy ratio in combination with the estimated result of the energy entropy ratio at the previous moment, and the calculation formula is:

[0022]

[0023] Among them, β is the smoothing factor, which ranges from 0 to 1;

[0024] S103: Determination of noise frame and stable noise segment: Calculate the estimated value of the energy entropy ratio of the current frame of the main microphone signal m1; if the estimated value of the energy entropy ratio of the current frame is lower than the energy entropy ratio threshold, determine that the current frame is a noise frame, and add 1 to the frame count; when the frame count is greater than or equal to NoiseCounter, determine that it enters the stable noise segment; if the estimated value of the energy entropy ratio of the current frame is higher than the energy entropy ratio threshold, set NoiseCounter to zero, and determine that it does not belong to the stable noise segment;

[0025] The energy entropy ratio baseline value and the energy entropy ratio threshold are dynamically updated.

[0026] Preferably, in step S10, the specific steps of dynamically updating the energy entropy ratio baseline value are as follows:

[0027] S104: Dynamically update the energy entropy ratio baseline value, and select the average of the energy entropy ratio estimation values ​​of several main microphone noisy speech signals as the energy entropy ratio baseline value; if the current frame is determined to be a frame containing a speech signal, the value of the energy entropy ratio baseline value will not be updated, and the last calculation result will be continued; if the current frame is determined to be a noise frame, different updates will be performed depending on whether it belongs to a stable noise segment.

[0028] Preferably, in step S10, the energy entropy ratio threshold is dynamically updated according to the energy entropy ratio calculation result, the energy entropy ratio estimation value and the energy entropy ratio baseline value.

[0029] Preferably, in the noise tracking of step S20, the specific steps of updating the secondary microphone signal delay d are as follows:

[0030] S201: Calculate d i-1 Delay to obtain DMA output signal strength:

[0031] S2011: Set the delay variable of the i-th frame to d i , with d i-1 Delay the auxiliary microphone signal m2 by a fractional delay value to obtain the delayed auxiliary microphone signal m2d;

[0032] S2012: Calculate the differential signal of the main microphone signals m1 and m2d of the current frame to obtain the time domain DMA output dmac;

[0033] S2013: Calculate the signal strength pc of dmac, the calculation formula is:

[0034]

[0035] S202 and S203: Delay the delayed secondary microphone signal m2d by step and advance by step sampling points respectively to obtain m2dr and m2dl, and calculate the time domain DMA output with m1 respectively, calculate the intensity of the output signal respectively, and obtain pr and pl. The calculation formula is:

[0036]

[0037]

[0038] S204: Compare the values ​​of pc, pr, and pl, and update the delay d of the secondary microphone based on the delay corresponding to the minimum value, as follows:

[0039]

[0040] S205: Calculate the delay slow change result D_sm using a decision-oriented method, the formula is as follows:

[0041] D_sm i =μ*D_sm i-1 +(1-μ)*d i

[0042] Among them, μ is the smoothing factor, and its value range is 0 to 1.

[0043] Preferably, in step S30, the specific steps of the first-level noise reduction are as follows:

[0044] The delay slow-changing result D_sm obtained in step S20 is used to perform a fractional delay on the secondary microphone signal m2, obtain the output dma1 of the first-order differential microphone array in the time domain, and perform a fast Fourier transform to the frequency domain, and use the filter H L Perform low-frequency energy compensation to obtain frequency domain DMA output as follows:

[0045] dma1=m1-m2d

[0046] out dma =fft(dma1)H L

[0047] Among them, the compensation filter of the first-order DMA is

[0048]

[0049] Angular frequency ω=2*pi*f,τ0=D / c,D is the distance between the main microphone and the auxiliary microphone,c is the speed of sound,α 1,1 =cosθ,θ is the zero point of the system.

[0050] Preferably, in step S40, a Wiener filtering algorithm based on an improved priori signal-to-noise ratio estimation method is combined with voice activity detection to perform secondary noise reduction, and the specific steps are as follows:

[0051] S401: Decision-oriented energy estimation and minimum energy estimation update;

[0052] S402: Decision-oriented noise spectrum estimation and voice activity detection update;

[0053] S403: 2DD algorithm estimates the prior signal-to-noise ratio ξ i,j ;

[0054] S404: Calculate the Wiener filter frequency gain coefficient g, the formula is as follows:

[0055]

[0056] S405: Adjusting Wiener filter coefficients based on voice activity detection;

[0057] S406: Calculate the Wiener filter noise reduction output and convert it to time domain output through inverse Fourier transform.

[0058] Preferably, the specific steps of step S401 are as follows:

[0059] S4011: Calculate the frequency energy after the first level of noise reduction, and update the speech detection and noise spectrum based on the minimum energy value tracking. Let the energy corresponding to the frequency point j of the i-th frame after the frequency domain compensation in step S30 be Pow i,j ;

[0060] S4012: Based on DD energy estimation and minimum energy estimation, obtain energy estimation value through DD method

[0061] The step S402 is specifically as follows:

[0062] S4021: Frequency point voice activity detection, set the threshold of voice activity detection at the jth frequency point in the i-th frame to TH i,j , if the instantaneous energy value Pow of a certain frequency point in the current frame i,j Higher than the threshold TH of its previous frame i-1,j , it is determined that there is voice activity at this frequency point, that is, VAD i,j =1, otherwise, VAD i,j =0;

[0063] S4022: Update the VAD decision threshold and noise spectrum based on DD.

[0064] Preferably, in S405, the noise suppression degree is controlled by setting a suppression factor R, and the formula is as follows:

[0065]

[0066] Among them, R takes values ​​from 0 to 1.

[0067] The beneficial effects of the present invention are:

[0068] 1. An adaptive dual-microphone noise reduction method of the present invention adopts a left-right search method similar to grid search, splits the calculation amount into different time windows, and gradually obtains fractional delays approaching the direction of the noise source, thereby reducing real-time computing resource consumption.

[0069] 2. In an adaptive dual-microphone noise reduction method of the present invention, in order to prevent the front speech signal from interfering with the noise direction tracking, the noise source tracking is only performed in a stable noise segment.

[0070] 3. In the adaptive dual-microphone noise reduction method of the present invention, there is still some residual noise in the signal after being processed by the differential microphone array, especially when the noise source and the signal source are close in spatial orientation, the residual noise is more. By adopting the improved Wiener filtering method for secondary noise reduction, instant, efficient and deep noise reduction can be achieved.

[0071] 4. In an adaptive dual-microphone noise reduction method of the present invention, the gain value of the frequency point where the VAD detection median value is 0 is further reduced in the secondary noise reduction, and the degree of noise suppression is controlled by setting the suppression factor R, so as to further suppress background noise and improve the noise reduction effect. BRIEF DESCRIPTION OF THE DRAWINGS

[0072] Figure 1 This is an overall flow chart of an adaptive dual-microphone noise reduction method according to a specific embodiment of the present invention;

[0073] Figure 2 It is a schematic diagram of a process of stable noise detection of an adaptive dual-microphone noise reduction method according to a specific embodiment of the present invention;

[0074] Figure 3 A schematic diagram of a first-level noise reduction process of an adaptive dual-microphone noise reduction method according to a specific embodiment of the present invention;

[0075] Figure 4 A schematic diagram of a secondary noise reduction algorithm flow of an adaptive dual-microphone noise reduction method according to a specific embodiment of the present invention;

[0076] Figure 5 A schematic diagram of a recording environment of an adaptive dual-microphone noise reduction method according to a specific embodiment of the present invention;

[0077] Figure 6 This is an example diagram of a secondary microphone delay tracking of an adaptive dual-microphone noise reduction method according to a specific embodiment of the present invention;

[0078] Figure 7 A comparison diagram before and after noise reduction of an adaptive dual-microphone noise reduction method according to a specific embodiment of the present invention;

[0079] Figure 8 It is an enlarged schematic diagram of the comparison before and after noise reduction of an adaptive dual-microphone noise reduction method according to a specific embodiment of the present invention;

[0080] Fig. 9 This is a statistical diagram of the improvement values ​​of various noise reduction indicators under various signal-to-noise ratios of an adaptive dual-microphone noise reduction method according to a specific embodiment of the present invention. DETAILED DESCRIPTION

[0081] The preferred embodiments of the present invention will be described in detail below with reference to the accompanying drawings.

[0082] See also Figure 1 , is an adaptive dual-microphone noise reduction method of the present invention, comprising the following steps:

[0083] S10: Stable noise segment detection, calculate the characteristic value of the current frame of the main microphone signal m1; if the characteristic value of the current frame is lower than the characteristic value threshold, the current frame is determined to be a noise frame, and 1 is added to the frame count; when the frame count is greater than or equal to NoiseCounter (frame count), it is determined to enter the stable noise segment; if the characteristic value of the current frame is higher than the characteristic value threshold, NoiseCounter is set to zero, and it is determined not to belong to the stable noise segment;

[0084] Specifically, the stable noise segment refers to a continuous silent segment, that is, a period of time in which there is no main voice signal input for a period of time.

[0085] Dynamically update the characteristic value baseline value and characteristic value threshold;

[0086] Further, the characteristic value includes an energy entropy ratio estimation value, an energy zero ratio, a calculation method based on an autocorrelation function, based on a subband variance, based on a spectrum distance, a cepstrum distance, a spectrum entropy characteristic value, etc., which will not be described in detail here. In a specific embodiment of the present invention, the energy entropy ratio estimation value can be used as a characteristic value for stable noise segment detection, and the energy entropy ratio baseline value and the energy entropy ratio threshold value can be dynamically updated.

[0087] S20: Based on minimizing the output power of differential microphone arrays (DMA), the stable noise segment is tracked, and the delay d of the secondary microphone signal m2 is updated in real time to obtain a delay slow-changing result D_sm;

[0088] S30: Based on the delay slow-changing result D_sm, perform first-level noise reduction to obtain frequency domain DMA output;

[0089] S40: Output after secondary noise reduction.

[0090] This application adopts an energy-entropy ratio-based method to detect stable noise segments, and at the same time, noise source tracking is only detected in stable noise segments, which effectively improves the noise reduction effect; a left-right search method similar to grid search is adopted to split the calculation amount into different time windows, and gradually obtain decimal delays approaching the direction of the noise source, thereby reducing real-time computing resource consumption; by setting first-level noise reduction and second-level noise reduction, instant, efficient and deep noise reduction is achieved.

[0091] See also Figure 1 and Figure 2 , assuming that the two microphones that are closer to the signal source are the main microphone, and the other is the auxiliary microphone, and the distance between the two microphones is D.

[0092] The specific steps of S10 are as follows:

[0093] S101: Calculate the energy entropy ratio, perform windowing and framing processing and fast Fourier transform on the main microphone signal m1, that is, obtain the i-th frame of speech signal after windowing and framing processing, and then perform N-point length Fast Fourier Transform (Fast Fourier Transform, FFT) conversion, the calculation formula of the i-th frame data spectrum entropy is:

[0094]

[0095] Where i represents the i-th frame, N represents the frame length of the data interception, k represents the k-th frequency component, Pro i,k represents the normalized spectral probability density function of each frequency component, which is defined as:

[0096]

[0097] Among them, P i,K represents the energy of the kth frequency component in the i-th frame;

[0098] Specifically, the normalized spectral probability density function of noise is distributed more evenly, so the spectral entropy value is larger, while the speech signal has a resonance peak spectrum characteristic, so its normalized spectral probability density function is distributed unevenly. Therefore, the spectral entropy of speech is generally lower than that of noise.

[0099] Next, calculate the short-time average energy. Short-time average energy is a commonly used feature for speech endpoint detection. When the signal-to-noise ratio is not very low, the short-time average energy of a segment with speech is generally higher than that of a segment without speech. The calculation formula for short-time average energy is:

[0100] Energy Entropy Ratio EER i The calculation formula is:

[0101] Furthermore, in order to reduce the interference of abnormal extreme values ​​on the characteristic values, the present invention may also adopt the square root of the energy entropy ratio as the detection characteristic value of the stable noise segment. The energy entropy ratio or some parameters related to the energy entropy ratio mentioned in the following scheme may also be based on the square root of the energy entropy ratio, which will not be elaborated here.

[0102] Furthermore, based on the short-term stability of speech, the frame length can be set to 10 to 30 ms, which is more reasonable; in the embodiment of the present invention, the length of each frame is set to 16 ms, the sampling rate is 16 kHz, the FFT point number is set to 256, and the frame overlap is 8 ms.

[0103] S102: Calculate the estimated value of the energy entropy ratio, and use a decision-oriented method to estimate the calculation result of the current energy entropy ratio in combination with the estimated result of the energy entropy ratio at the previous moment, and the calculation formula is:

[0104]

[0105] Among them, β is the smoothing factor, and its value is between 0 and 1; its value affects the smoothing performance. The larger the value, the smoother the result, but relatively, the instantaneousness of the calculation result will be reduced. It is recommended to use a value above 0.95 here.

[0106] S103: Determination of noise frame and stable noise segment: If the estimated value of the energy entropy ratio of the current frame is lower than the energy entropy ratio threshold, the current frame is determined to be a noise frame, and 1 is added to the frame count; when the frame count is greater than or equal to NoiseCounter, it is determined to enter the stable noise segment; if the estimated value of the energy entropy ratio of the current frame is higher than the energy entropy ratio threshold, NoiseCounter is set to zero, and it is determined not to belong to the stable noise segment.

[0107] Furthermore, the value of NoiseCounter can be 125, and other reasonable values ​​are also possible.

[0108] S104: Dynamically update the energy entropy ratio baseline value, select the average of several (may be 50) energy entropy ratio estimation values ​​of the main microphone noisy speech signals as the energy entropy ratio baseline value; if there are less than 50, take the average of all elements in the vector.

[0109] If the current frame is determined to be a frame containing a speech signal, the value of the energy entropy ratio baseline value is not updated, and the last calculation result is continued; if the current frame is determined to be a noise frame, different updates are performed depending on whether it belongs to a stable noise segment. If it belongs to a stable noise segment, when the number of frames is less than 50, the estimated value of the energy entropy ratio of the current frame is directly added to the result, otherwise, 80% of the estimated value of the energy entropy ratio is added. When the number of frames is higher than 50, the first element is removed, and according to the result of whether it belongs to a stable noise segment, the estimated value of the energy entropy ratio of the current frame or 80% thereof is added to the end. In this way, the elements of the vector are maintained at most 50. Of course, 50 is the value in this embodiment, and other reasonable values ​​are also acceptable.

[0110] S105: Dynamically update the energy entropy ratio threshold value according to the energy entropy ratio calculation result, the energy entropy ratio estimation value and the energy entropy ratio baseline value.

[0111] Specifically, the energy entropy ratio threshold is dynamically updated in a similar way to the energy entropy ratio estimate, but the size of the smoothing factor is regulated by the current instant energy entropy ratio, the energy entropy ratio estimate, and the energy entropy ratio baseline. If the current energy entropy ratio is higher than the energy entropy ratio threshold of the previous frame, the smoothing factor is set higher, and it is recommended to be above 0.99; if the current energy entropy ratio is lower than the energy entropy ratio threshold of the previous frame, that is, when the current frame is determined to be a noise frame, it is further updated to varying degrees based on the energy entropy ratio estimate and the energy entropy ratio baseline. When the energy entropy ratio estimate is less than 1.2 times the energy entropy ratio baseline, a faster update is performed, that is, a smaller smoothing factor is set, and it is recommended to be 0.85 to 0.9. Otherwise, a slower update is performed, and it is recommended that the smoothing factor be set to be above 0.99.

[0112] Furthermore, in order to reduce the weakening of the speech signal and prevent the unstable background noise energy due to the presence of speech pauses during speech segments, the noise tracking step is performed only in the stable noise segment, and the delay d of the secondary microphone signal m2 is updated in real time based on minimizing the output power of the differential microphone array. In order to accurately achieve the delay, d is initialized to 0, and its update step step is set to a small decimal less than 1. In this embodiment, 0.1 is used, and the unit is sampling point.

[0113] See also Figure 2 and Figure 3 In the noise tracking of step S20, the specific steps of updating the delay d of the secondary microphone signal are as follows:

[0114] S201: Calculate d i-1 Delay to obtain DMA output signal strength:

[0115] S2011: Set the delay variable of the i-th frame to d i , with d i-1The secondary microphone signal m2 is subjected to a decimal delay according to the delay value to obtain the delayed secondary microphone signal m2d; specifically, the decimal delay method can be implemented by constructing a decimal delay filter or the like.

[0116] S2012: Calculate the differential signal of the main microphone signals m1 and m2d of the current frame to obtain the time domain DMA output dmac;

[0117] S2013: Calculate the signal strength pc of dmac, the calculation formula is:

[0118]

[0119] S202 and S203: Delay the delayed secondary microphone signal m2d by step and advance by step sampling points respectively to obtain m2dr and m2dl, and calculate the time domain DMA output with m1 respectively, calculate the intensity of the output signal respectively, and obtain pr and pl. The calculation formula is:

[0120]

[0121]

[0122] S204: Compare the values ​​of pc, pr, and pl, and update the delay d of the secondary microphone based on the delay corresponding to the minimum value, as follows:

[0123]

[0124] S205: Calculate the delay slow change result D_sm using a decision-oriented method, the formula is as follows:

[0125] D_sm i =μ*D_sm i-1 +(1-μ)*d i

[0126] Among them, μ is the smoothing factor, and its value range is 0 to 1.

[0127] S206: If the current frame belongs to a non-stable noise segment, then the average value of the delay slow-changing results within a short period of time (eg, 100-1000 ms) is taken as the current frame delay value.

[0128] In step S30, the specific steps of the first-level noise reduction are as follows:

[0129] S301: Using the delay slow variation result D_sm obtained in step S20, perform a decimal delay on the secondary microphone signal m2 to obtain m2d;

[0130] S302: Obtain the output dma1 of the time domain first-order differential microphone array (DMA), and perform a fast Fourier transform (FFT) to the frequency domain, using a filter H L Perform low-frequency energy compensation to obtain frequency domain DMA output as follows:

[0131] dma1=m1-m2d

[0132] ouot dma =fft(dma1)H L

[0133] Among them, the compensation filter of the first-order DMA is

[0134]

[0135] Angular frequency ω = 2*pi*f, τ0 = D / c, D is the distance between the main microphone and the auxiliary microphone, c is the speed of sound, α 1,1 =cosθ,θ is the zero point of the system.

[0136] After the first-level noise reduction using DMA, there will still be a certain amount of noise residual in the output. To further reduce the noise, this embodiment uses the Wiener filtering algorithm based on the improved prior signal-to-noise ratio estimation method combined with voice activity detection (VAD) to perform secondary noise reduction. Of course, other methods can also be used for secondary noise reduction. The Wiener filtering algorithm relies on accurate prior signal-to-noise ratio estimation, and one of the most commonly used methods for estimating the prior signal-to-noise ratio is to use the decision-directed (DD) method, which has a low algorithm complexity and can reduce the residual music noise after noise reduction. However, in the traditional DD method, the smoothing factor is a pre-set fixed value, which causes the algorithm to make a trade-off between speech distortion and music noise residual.

[0137] To address this problem, the present application adopts a convex combination of two DD methods with smoothing factors of different sizes (2DD algorithm), and calculates a hybrid parameter based on the principle of minimizing the root mean square error between the estimated value of the prior signal-to-noise ratio and the true value, so as to control the influence of the prior signal-to-noise ratio estimated by the 2DD method on the final prior signal-to-noise ratio estimation. The present application further combines the Wiener filter coefficient value calculated in this way with the speech detection algorithm to further suppress the coefficient value corresponding to the estimated noise frequency point.

[0138] See also Figure 4 The specific steps of the secondary noise reduction of S40 are as follows:

[0139] S401: Decision-oriented energy estimation and minimum energy estimation update;

[0140] The specific steps of step S401 are as follows:

[0141] S4011: Calculate the frequency energy after the first level of noise reduction, and the speech detection and noise spectrum update are based on the minimum energy value tracking; let the energy corresponding to the frequency point j of the i-th frame after the frequency domain compensation in S302 be Pow i,j ;

[0142] S4012: Based on DD energy estimation and minimum value estimation, obtain energy estimation value through DD method The calculation method is:

[0143]

[0144] Among them, β is the smoothing factor, and its value range is 0 to 1.

[0145] Then two DDs with different smoothing factors are used to update the minimum energy estimate Pow_min i,j , the method is as follows:

[0146]

[0147] Let β1 be smaller than β2, and the value of β2 is close to 1.

[0148] S402: Decision-oriented noise spectrum estimation and voice activity detection update;

[0149] Step S402 is specifically as follows:

[0150] S4021: Frequency point voice activity detection VAD; set the threshold of voice activity detection at the jth frequency point in the i-th frame to TH i,j , if the instantaneous energy value Pow of a certain frequency point in the current frame i,j Higher than the threshold TH of its previous frame i-1,j , it is determined that there is voice activity at this frequency point, that is, VAD i,j =1, otherwise, VAD i,j =0;

[0151] S4022: Update the VAD determination threshold and noise spectrum based on DD;

[0152] VAD detection threshold TH i,j And noise spectrum estimation noise i,j The update is dynamically updated through multiple decision-oriented methods with different smoothing factors to achieve different degrees of update in different situations. The method is as follows: the ratio of the energy estimate value of each frequency point to the minimum energy estimate value is calculated. If the ratio is higher than a certain set threshold, the probability of speech activity at the frequency point is high. In order to improve the instantaneousness of the decision, the instantaneous energy Pow of the current frequency point is compared. i,j With THi-1,j If the instantaneous energy is high, it will not be updated, that is, TH i,j =TH i-1,j , noise i,j =noise i-1,j Otherwise, the DD method with a higher smoothing factor (value close to 1) is used to update both. If the ratio of the energy estimation value of each frequency point to the minimum energy estimation value is lower than the set threshold, the probability of speech activity at the frequency point is low. At this time, the instantaneous energy Pow of the current frequency point is further compared. i,j With TH i-1,j According to the comparison results, different smoothing factor decision methods are used to determine the VAD detection threshold TH. i,j And noise spectrum estimation noise i,j In principle, when the instantaneous energy is high, a larger smoothing factor value is set.

[0153] S403: 2DD algorithm estimates the prior signal-to-noise ratio ξ i,j ;

[0154] Assume that the posterior signal-to-noise ratio and the prior signal-to-noise ratio of the jth frequency point in the i-th frame are η i,j and i,j , and its calculation formula is:

[0155]

[0156] Since the true priori signal-to-noise ratio cannot be obtained in practice, the estimated value is used here to obtain the estimated value λ of the mixing factor in the 2DD algorithm;

[0157] ξ i,j =max(η i,j -1,0)

[0158] set up and The DD method uses two smoothing factors of different sizes (a and b, assuming a>b, a is 0.99 and b is 0.6 in this embodiment) to calculate the prior signal-to-noise ratio ξ i,j The estimated result is

[0159]

[0160] The method of estimating the prior signal-to-noise ratio using 2DD is:

[0161]

[0162] Among them, λ is the mixing factor, and its value is obtained based on minimizing the cost loss function J:

[0163]

[0164] In the above formula, E{.} is the expected value. J is derived with respect to λ,

[0165]

[0166] make

[0167]

[0168] The value of λ is calculated as follows:

[0169]

[0170] S404: Calculate the Wiener filter frequency gain coefficient g, the formula is as follows:

[0171]

[0172] S405: Adjust the Wiener filter coefficient based on the VAD result. To further suppress the background noise, further reduce the gain value of the frequency point where the VAD detection value is 0. The noise suppression degree is controlled by setting the suppression factor R. The formula is as follows:

[0173]

[0174] Among them, R takes values ​​from 0 to 1.

[0175] S406: Calculate the Wiener filter output and convert it to the time domain output through Inverse Fast Fourier Transform (IFFT); let the value corresponding to the frequency point j of the i-th frame after the frequency domain compensation in S302 be Y i,j , the corresponding frequency point estimation after noise reduction is The number of FFT points is N, and the calculation formula for the final output output is:

[0176]

[0177] See also Figure 5 , the specific test method is as follows:

[0178] The distance between the two microphones of the device used in this embodiment is 66mm, and the ADC sampling rate is 16kHz. The device is used to collect dual-microphone radio signals as the input source of the algorithm to evaluate the noise reduction effect of the algorithm. The device is placed in the center of the soundproof room. The soundproof room is equipped with a circular sound field with 12 speakers. The distance between adjacent speakers is 30 degrees. The sound field control system software installed on a computer outside the soundproof room controls audio selection, speaker selection, signal-to-noise ratio setting, playback time, etc. The types of noise include white noise (white), spectral noise (ssn), multi-person conversation (babble), restaurant noise (canteen), and the voice is a Chinese female broadcast voice. The signal-to-noise ratio of the noisy speech is set to -5, 0, 5, and 10dB. When recording, it is set to play 10s of noise first, followed by 20s of noisy speech. The voice signal is fixed from the front of the device (0 degrees), and the noise comes from one of the four angles of 0, 90, 180, and 270 degrees. The dual-microphone radio signal of the device is cached and saved locally through another recording computer. See the schematic diagram of the recording environment. Figure 5 The audio data saved locally is spliced ​​in a manually set way to simulate the situation where the direction of the noise source changes in real life. The spliced ​​audio is used as the input of this algorithm to obtain the output after noise reduction. After that, the improvement of speech signal-to-noise ratio and speech intelligibility before and after noise reduction is compared as an indicator to measure the noise reduction performance of this algorithm.

[0179] Statistical test methods:

[0180] Matlab R2016a software was used for data analysis. Lilliefors was used to test the normal distribution of data, and Bartlett was used to test the homogeneity of variance of data. For comparisons between two groups, if normal distribution was satisfied, paired t-test or two-sample t-test was used to test the significance of the difference in the mean of data distribution. If normal distribution was not satisfied, Wilcoxon signed rank test (paired samples) or Mann-Whitney U test (independent samples) was used to test the significance of the difference in the median of data distribution. For the comparison of the improvement values ​​of various noise reduction indicators under different signal-to-noise ratios, except for the LSD indicator that met the requirements of normal distribution and homogeneity of variance of each group of data, one-way analysis of variance ANOVA was used, and the other indicators were all tested by Kruskal-Wallis non-parametric test, and the bonferroni test was used for post hoc multiple comparisons.

[0181] Noise reduction index description:

[0182] Four indicators are used to measure the objective noise reduction performance of the algorithm, including segmental signal-to-noise ratio segSNR (segmental Signal-to-Noise Ratio Measures), log-spectral distance LSD (Log-Spectral Distance), short-time objective intelligibility index STOI (short-time objective intelligibility) and speech quality perception evaluation index PESQ (Perceptual Evaluation of Speech Quality, using the value of widebandMOS LQO). segSNR first divides the input speech signal into frames, then calculates the signal-to-noise ratio for each frame of speech, and finally calculates the average. LSD is a distance measure between two spectra, also known as "log-spectral distortion". STOI is an objective evaluation indicator for measuring the improvement of speech intelligibility after noise reduction (Taal et al., 2011). The value range of this indicator is between 0 and 1. The larger the value, the higher the intelligibility and the clearer the speech. The PQESQ algorithm was one of the earliest objective speech quality evaluation algorithms proposed by the International Telecommunications Union (ITU). It is the algorithm with the highest correlation with subjective evaluation results among the objective speech quality evaluation algorithms published by ITU. The score range of the PESQ original model is between -0.5 and 4.5 (after mapping to MOS LQO, the value is between 1.0 and 4.6). The higher the score, the better the speech quality.

[0183] Example of noise reduction effect:

[0184] See the example of delayed tracking effect Figure 6 The noise type is babble, the signal-to-noise ratio of the noisy speech segment is 5dB, and the noise comes from the speakers at 0, 90, 180 and 270 degrees. The time domain waveform signal and time-frequency diagram of this file before and after noise reduction are shown in Figure 7 , and zoom in on some of its areas to show it more clearly, see Figure 8 .

[0185] Noise reduction effect analysis:

[0186] In order to simulate as many usage scenarios as possible, the recording files were cut and spliced ​​into 32 files by distinguishing the gender of the speaker, the noise category, the direction of noise source change and the signal-to-noise ratio. The noise reduction algorithm was used to denoise each file and four objective indicators were calculated before and after noise reduction, including segmented signal-to-noise ratio (segSNR), log spectral distance (LSD), short-term speech objective intelligibility (STOI) and speech quality index PESQ. Table 1 shows the improvement of various indicators in various noisy environments after using this noise reduction algorithm compared with before noise reduction. Statistical analysis of the differences between the indicators before and after noise reduction was performed, and the performance of each indicator after noise reduction was significantly improved (segSNR, p = 7.95e-7; LSD, p = 1.62e-22; STOI, p = 1.01e-4; PESQ, p = 1.71e-4).

[0187]

[0188] Table 1 Improvement of various indicators after noise reduction (after noise reduction - before noise reduction)

[0189] Analysis of the difference in the improvement value of each noise reduction index under different noise backgrounds:

[0190] The mean (Mean±SD) improvement of segSNR of the noise reduction output under stable noise (white and ssn) background is 17.56±0.68dB, and the non-stable noise (including babble and canteen The mean segSNR improvement of the denoising output under the stable noise background is 13.85±0.85dB, which is significantly different from the two (p=1.54e-6); the mean LSD improvement of the denoising output under the stable noise background is -13.12±2.38, and the mean LSD improvement of the denoising output under the unstable noise background is -10.88±2.44, which is significantly different from the two (p=0.01); the mean STOI improvement of the denoising output under the stable noise background is 0.02±0.04, and the mean STOI improvement of the denoising output under the unstable noise background is 0.03±0.02, and there is no significant difference between the two (p=0.30); the mean PESQ improvement of the denoising output under the stable noise background is 0.40±0.19, and the mean PESQ improvement of the denoising output under the unstable noise background is 0.18±0.44, and there is no significant difference between the two (p=0.13).

[0191] Analysis of the difference in the improvement of each noise reduction index under different signal-to-noise ratios:

[0192] Fig. 9The figure shows the difference in the improvement of the four indicators after noise reduction compared with before noise reduction at each signal-to-noise ratio. The red horizontal line is the median value, and the upper and lower black short horizontal lines are the third and first quartiles respectively. For the STOI and PESQ graphs, the middle horizontal black dotted line shows the position of the 0 value. It can be seen that the improvement of segSNR is not much different at each signal-to-noise ratio. The improvement of LSD and PESQ increases with the increase of signal-to-noise ratio, and the improvement of STOI value decreases with the increase of signal-to-noise ratio. When the signal-to-noise ratio is as high as 10dB, the improvement value of STOI is less than 0, that is, the STOI value after noise reduction decreases. The segSNR improvement values ​​(Mean±SD) at -5, 0, 5, and 10dB were 15.22±2.32, 15.51±2.01, 15.80±2.30, and 16.30±1.66dB, respectively, with no significant difference between the groups (p=0.50); ​​the LSD improvement values ​​were -8.65±1.27, -11.27±1.37, -13.44±1.58, and -14.6 5±0.92, there were significant differences among the groups (p=2.87e-9), and post hoc multiple comparisons showed that there were significant differences between -5dB and 0dB, -5dB and 5dB, -5dB and 10dB, 0dB and 5dB, and 0dB and 10dB groups (p were 2.46e-3, 3.31e-7, 3.69e-9, 1.49e-2, and 1.04e-4, respectively); the STOI improvement values ​​were 0.056±0.013, 0.044±0.014, 0.012±0.013, -0.014±0.015, there were significant differences between the groups (p=1.42e-5), and post hoc multiple comparisons showed that there were significant differences between -5dB and 5dB, -5dB and 10dB, and 0dB and 10dB groups (p were 1.30e-2, 4.23e-5, and 1.73e-3, respectively); P The ESQ improvement values ​​were -0.12±0.48, 0.27±0.09, 0.44±0.11, and 0.56±0.09, respectively. There were significant differences between the groups (p=1.33e-5). Post hoc multiple comparisons showed that there were significant differences between the -5dB and 5dB, -5dB and 10dB, and 0dB and 10dB groups (p were 2.36e-3, 1.86e-5, and 1.86e-2, respectively).

[0193] The adaptive dual-microphone noise reduction method of the present invention detects stable noise segments based on the energy entropy ratio and tracks noise in stable noise segments; at the same time, the calculation amount is split into different time windows, and the fractional delay approaching the direction of the noise source is gradually obtained, thereby reducing the consumption of real-time computing resources. In addition, the first-level noise reduction and the second-level noise reduction are adopted to achieve instant, efficient and deep noise reduction.

[0194] Finally, it should be noted that the above preferred embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail through the above preferred embodiments, those skilled in the art should understand that various changes can be made in form and details without departing from the scope defined by the claims of the present invention.

Claims

1. An adaptive dual-microphone noise reduction method, characterized in that: The following steps are involved: S10: Stable noise segment detection, calculate the characteristic value of the current frame of the main microphone signal m1; if the characteristic value of the current frame is lower than the characteristic value threshold, the current frame is determined to be a noise frame, and 1 is added to the frame count; when the frame count is greater than or equal to NoiseCounter, it is determined to enter the stable noise segment; if the characteristic value of the current frame is higher than the characteristic value threshold, NoiseCounter is set to zero, and it is determined not to belong to the stable noise segment; Dynamically update the characteristic value baseline value and characteristic value threshold; S20: performing noise tracking on the stable noise segment based on minimizing the differential microphone array output power, and updating the delay d of the secondary microphone signal m2 in real time to obtain a delay slow-changing result D_sm; S30: Based on the delay slow-changing result D_sm, perform first-level noise reduction to obtain frequency domain DMA output; S40: Output after secondary noise reduction.

2. The adaptive dual-microphone noise reduction method according to claim 1, characterized in that: The eigenvalues ​​include but are not limited to energy entropy ratio estimation, energy zero ratio, autocorrelation function-based, sub-band variance-based, spectral distance-based, cepstrum distance, and spectrum entropy eigenvalues.

3. The adaptive dual-microphone noise reduction method according to claim 2, characterized in that: When the characteristic value is the estimated value of the energy entropy ratio, in the stable noise segment detection of step S10, the specific steps of calculating the estimated value of the energy entropy ratio are as follows: S101: Calculate the energy entropy ratio, perform windowing and framing processing and fast Fourier transform on the main microphone signal m1, and the calculation formula for the spectrum entropy of the i-th frame data is: Where i represents the i-th frame, N represents the frame length of the data interception, k represents the k-th frequency component, Pro i,k represents the normalized spectral probability density function of each frequency component, which is defined as: Among them, P i,k represents the energy of the kth frequency component in the i-th frame; The short-time average energy is calculated, and the calculation formula of the short-time average energy is: The energy entropy ratio EER i The calculation formula is: S102: Calculate the estimated value of the energy entropy ratio, and use a decision-oriented method to estimate the calculation result of the current energy entropy ratio in combination with the estimated result of the energy entropy ratio at the previous moment, and the calculation formula is: Among them, β is the smoothing factor, which ranges from 0 to 1; S103: Determination of noise frame and stable noise segment: Calculate the estimated value of the energy entropy ratio of the current frame of the main microphone signal m1; if the estimated value of the energy entropy ratio of the current frame is lower than the energy entropy ratio threshold, determine that the current frame is a noise frame, and add 1 to the frame count; when the frame count is greater than or equal to NoiseCounter, determine that it enters the stable noise segment; if the estimated value of the energy entropy ratio of the current frame is higher than the energy entropy ratio threshold, set NoiseCounter to zero, and determine that it does not belong to the stable noise segment; The energy entropy ratio baseline value and the energy entropy ratio threshold are dynamically updated.

4. The adaptive dual-microphone noise reduction method according to claim 3, characterized in that: In step S10, the specific steps of dynamically updating the energy entropy ratio baseline value are as follows: S104: Dynamically update the energy entropy ratio baseline value, and select the average of the energy entropy ratio estimation values ​​of several main microphone noisy speech signals as the energy entropy ratio baseline value; if the current frame is determined to be a frame containing a speech signal, the value of the energy entropy ratio baseline value will not be updated, and the last calculation result will be continued; if the current frame is determined to be a noise frame, different updates will be performed depending on whether it belongs to a stable noise segment.

5. The adaptive dual-microphone noise reduction method according to claim 4, characterized in that: In the step S10, the energy entropy ratio threshold is dynamically updated according to the energy entropy ratio calculation result, the energy entropy ratio estimation value and the energy entropy ratio baseline value.

6. The adaptive dual-microphone noise reduction method according to claim 1, characterized in that: In the noise tracking of step S20, the specific steps of updating the delay d of the secondary microphone signal are as follows: S201: Calculate the value of d i-1 Delay to obtain DMA output signal strength: S2011: Set the delay variable of the i-th frame to d i , with d i-1 Delay the auxiliary microphone signal m2 by a fractional delay value to obtain the delayed auxiliary microphone signal m2d; S2012: Calculate the differential signal of the main microphone signals m1 and m2d of the current frame to obtain the time domain DMA output dmac; S2013: Calculate the signal strength pc of dmac, the calculation formula is: S202 and S203: Delay the delayed secondary microphone signal m2d by step and advance by step sampling points respectively to obtain m2dr and m2dl, and calculate the time domain DMA output with m1 respectively, calculate the intensity of the output signal respectively, and obtain pr and pl. The calculation formula is: S204: Compare the values ​​of pc, pr, and pl, and update the delay d of the secondary microphone based on the delay corresponding to the minimum value, as follows: S205: Calculate the delay slow change result D_sm using a decision-oriented method, the formula is as follows: D_sm i =μ*D_sm i-1 +(1-μ)*d i Among them, μ is the smoothing factor, and its value range is 0 to 1.

7. The adaptive dual-microphone noise reduction method according to claim 1, characterized in that: In step S30, the specific steps of the first-level noise reduction are as follows: The delay slow-changing result D_sm obtained in step S20 is used to perform a fractional delay on the secondary microphone signal m2, obtain the output dma1 of the first-order differential microphone array in the time domain, and perform a fast Fourier transform to the frequency domain, and use the filter H L Perform low-frequency energy compensation to obtain frequency domain DMA output as follows: dma1=m1-m2d out dma =fft(dma1)H L Among them, the compensation filter of the first-order DMA is Angular frequency ω = 2*pi*f, τ0 = D / c, D is the distance between the main microphone and the auxiliary microphone, c is the speed of sound, α 1,1 =cosθ, θ is the zero point of the system.

8. The adaptive dual-microphone noise reduction method according to claim 1, characterized in that: In step S40, a Wiener filter algorithm based on an improved priori signal-to-noise ratio estimation method is used in combination with voice activity detection to perform secondary noise reduction, and the specific steps are as follows: S401: Decision-oriented energy estimation and minimum energy estimation update; S402: Decision-oriented noise spectrum estimation and voice activity detection update; S403: 2DD algorithm estimates the prior signal-to-noise ratio ξ i，j ; S404: Calculate the Wiener filter frequency gain coefficient g, the formula is as follows: S405: Adjusting Wiener filter coefficients based on voice activity detection; S406: Calculate the Wiener filter noise reduction output and convert it to time domain output through inverse Fourier transform.

9. The adaptive dual-microphone noise reduction method according to claim 8, characterized in that: The specific steps of step S401 are as follows: S4011: Calculate the frequency energy after the first level of noise reduction. Speech detection and noise spectrum update are based on the minimum energy value tracking. Let the energy corresponding to the frequency point j of the i-th frame after the frequency domain compensation in step S30 be Pow i,j ; S4012: Energy estimation and minimum energy estimation based on DD, and energy estimation value obtained by DD method The step S402 is specifically as follows: S4021: Frequency point voice activity detection VAD, set the threshold of voice activity detection at the i-th frequency point in the i-th frame to TH i,j , if the instantaneous energy value Pow of a certain frequency point in the current frame i,j Higher than the threshold TH of its previous frame i-1，j , it is determined that there is voice activity at this frequency point, that is, VAD i,j =1, otherwise, VAD i,j =0; S4022: Update the VAD decision threshold and noise spectrum based on DD.

10. The adaptive dual-microphone noise reduction method according to claim 9, characterized in that: In S405, the noise suppression degree is controlled by setting the suppression factor R, and the formula is as follows: Among them, R takes values ​​from 0 to 1.

Images