CN114724574A - Double-microphone noise reduction method with adjustable expected sound source direction - Google Patents

Abstract

The invention discloses a double-microphone noise reduction method with adjustable expected sound source direction, which comprises the following steps: preprocessing the noisy signal x received by the dual microphones₁(t) and x₂(t) discrete sampling, pre-emphasis, framing and windowing, and performing short-time Fourier transform to obtain frequency domain signal X₁(omega) and X₂(ω); a beam forming process, introducing a virtual microphone at the middle point of the connection line of the two microphones, and performing frequency domain signal X according to a central difference format₁(omega) and X₂(omega) performing a differential transformation to construct a differential signal Y₁(omega) and Y₂(ω). Calculating a difference signal Y₁(omega) and Y₂(ω) and the ratio of the statistical averages is recorded as a directivity function Γ (ω, θ), the properties of the directivity function Γ (ω, θ) are analyzed,it is directly mapped to the noise masking value λ (ω) by the normalization function. Mixing X₁Multiplying (omega) by lambda (omega) to obtain a signal R with the competing direction noise eliminated₁(ω); post-wiener filtering process, for R₁Estimating signal energy and noise energy in (omega) to obtain channel signal-to-noise ratio and calculating gain function to further eliminate R₁Residual noise in (ω).

Description

A dual-microphone noise reduction method with adjustable desired sound source direction

technical field

The present invention relates to the technical field of noise reduction of speech signals, and in particular, to a method for noise reduction of dual microphones with adjustable desired sound source direction.

Background technique

Portable devices such as Bluetooth headsets have gradually become a good tool for improving efficiency in people's daily lives since their introduction. However, when users use them to make or receive calls, if they are disturbed by background noise and non-target voices, the call quality will drop sharply. In this case, it is an urgent need to retain the voice close to the user's speaking direction, and to suppress background noise and non-target voice as much as possible on the premise of ensuring that the voice is not distorted.

Existing generalized sidelobe cancellers (GSCs) and delayed beamformers use multiple microphone recording signals for spatial filtering. For portable devices such as Bluetooth headsets, GSCs are too complex and beyond the capabilities of tiny devices. Delayed beamforming techniques such as the first-order differential microphone (FDM) and adaptive null-forming (ANF) require only two microphones and are suitable for size-constrained and real-time processing settings. But this fixed beamformer has a maximum gain at 0° and a zero at 180°, which cannot eliminate noise in directions other than the zero. Based on the algorithm of the coherence function between the input signals, the properties of the real and imaginary parts of the coherence function are discussed to generate different means of masking noise. Methods based on coherence functions do not rely on noise statistics, but the target orientation is not adjustable. Competitive directional noise cancellation methods for hearing aids combine spectral estimation and array beamforming to suppress noise. Directivity coefficients are estimated in the pure noise interval and updated to accommodate moving noise. Similarly, this method can only set the desired direction to a limited range around 0°. Since the position of the sound source is sometimes not constant, it is important to design a noise reduction algorithm with adjustable sound source direction in practical applications.

In order to solve the problem that the non-target direction sound cannot be accurately eliminated when beamforming is directly applied to the short-range dual-microphone system and the desired sound source direction cannot be set according to the requirements, a two-step denoising method based on beamforming and Wiener filtering is proposed. The test results show that this method can effectively restore the energy distribution characteristics of the original signal, reduce the background noise and non-target direction speech, and significantly improve the signal-to-noise ratio in the case of low signal-to-noise ratio and the coexistence of various types of noise sources. , the present invention can preset the desired sound source direction to adapt to the scene where the sound source changes.

SUMMARY OF THE INVENTION

According to the problems existing in the prior art, the present invention discloses a dual-microphone noise reduction method with adjustable desired sound source direction, which specifically includes:

Preprocessing process: Discrete sampling, pre-emphasis, framing and windowing of the noisy signals x ₁ (t) and x ₂ (t) received by the dual microphones, and then short-time Fourier transform to obtain frequency domain signals X ₁ (ω) and X ₂ (ω);

Beamforming process: introduce a virtual microphone at the midpoint of the two-microphone connection, perform differential transformation on the frequency domain signals X ₁ (ω) and X ₂ (ω) according to the center differential format, and construct the differential signals Y ₁ (ω) and Y ₂ (ω), calculate the statistical mean of the power spectra of the differential signals Y ₁ (ω) and Y ₂ (ω), and record the ratio of the statistical mean as the directional function Γ(ω, θ), and analyze the directional function The properties of Γ(ω, θ) are directly mapped to the noise masking value λ(ω) through the normalization function, and the frequency domain signal X ₁ (ω) is multiplied by the noise masking value λ(ω) to eliminate competition. Signal R ₁ (ω) after directional noise;

Post-Wiener filtering process: Estimate the signal energy and noise energy in R ₁ (ω) to obtain the channel SNR and calculate the gain function, thereby eliminating the residual noise in the signal R ₁ (ω).

Further, the preprocessing process is:

The noisy signals x ₁ (t) and x ₂ (t) are discretely sampled, and then the high-frequency part of the speech is pre-emphasized;

Divide the sampled signals x ₁ (n) and x ₂ (n) into frames with a length of 10ms and then add an equal-length Hamming window w(n), and the windowed signals pass into the buffer area for processing. The Lie transform obtains the current frame frequency domain signal, and according to the conjugate symmetry of the Fourier transform of the real number sequence, the signal of the first 1/2 frequency points is output for beamforming processing.

Further, the beamforming process includes amplitude alignment, calculation of power spectrum, calculation of directivity function value, calculation of threshold value and normalization mapping;

The amplitude alignment method is to multiply the frequency domain signals X ₁ (ω) and X ₂ (ω) by the scale factor respectively to perform amplitude alignment;

When calculating the power spectrum: Assuming that the desired beam is S(ω) and its direction is preset to α, introduce a virtual microphone at the midpoint of the two-way microphone to receive the desired beam S(ω), according to the central difference format and the frequency domain signal X ₁ ( The spatial relationship between ω), X ₂ (ω) and the desired beam S(ω) constructs the differential signals Y ₁ (ω) and Y ₂ (ω), and calculates the power spectrum of the differential signals Y ₁ (ω) and Y ₂ (ω) ;

When calculating the value of the directional function: the ratio of the statistical average of the power spectrum of the differential signal Y ₁ (ω) and Y ₂ (ω) is the value of the directional function Γ(ω, θ). Properties, when the actual incident direction θ of the sound source is equal to the given desired sound source incident direction α, Γ(ω, θ) tends to infinity, and on both sides of the θ=α axis, the function value is monotonic and approximately symmetrical;

When calculating the threshold and normalized mapping: Since Γ(ω, θ) tends to infinity, a threshold Ω is calculated according to the preset main lobe width Θ, and the Γ(ω, θ is directly converted by the normalized mapping of the sigmoid function. ) is mapped to the noise masking value λ(ω) of the corresponding frequency point, and X ₁ (ω) is multiplied by λ(ω) to obtain the signal R ₁ (ω) after eliminating the noise in the competing direction.

Further, the post-Wiener filtering process includes calculating the signal-to-noise ratio index, calculating the logarithmic spectral deviation, modifying or resetting the noise flag and calculating the gain function value;

When calculating the signal-to-noise ratio index, the signal R ₁ (ω) is divided into several channels according to the critical bandwidth criterion, the energy of each channel is estimated, the channel noise energy estimation is initialized as the channel energy of the first four frames, and the channel signal is calculated accordingly. noise ratio index;

When calculating the logarithmic spectral deviation, a non-linear data table is designed as a speech index table, the signal-to-noise ratio index is mapped to a set of numbers to measure the speech quality, and the sum of the speech indexes in a certain frequency range is used as the speech quality evaluation result of the current channel. Take the logarithm of the signal energy of the current channel, and calculate the deviation of the long-term logarithmic spectral energy and the short-term logarithmic spectral energy;

Modify or reset the noise flag, judge whether the current frame is a speech frame or a noise frame, reset the noise update flag, and check the update flags of the previous frames according to the calculated sum of speech indicators, signal-to-noise ratio index, and logarithmic spectral deviation parameter information , if the noise cannot be updated for a long time and the result is considered unreliable, the SNR index will be forced to update;

When calculating the gain function value, use the channel signal-to-noise ratio index to calculate the channel gain value to remove the residual background noise, and update the noise energy estimate of the next frame according to the result of the noise update flag.

Due to the adoption of the above technical solution, the present invention provides a dual-microphone noise reduction method with adjustable desired sound source direction. After preprocessing the dual-microphone signals, the method first calculates the statistical average value of the power spectrum of the constructed differential signal. The ratio is used as the value of the directional function, and then the masking value of the noise is obtained by the mapping of the normalized function. At the same time, a Wiener filter is installed in the next step to reduce residual noise by estimating the signal-to-noise ratio index and calculating the gain function; the algorithm proposed by the present invention is simple and efficient. The signal-to-noise ratio and quality are significantly improved.

Description of drawings

In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the following briefly introduces the accompanying drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only These are some embodiments described in this application. For those of ordinary skill in the art, other drawings can also be obtained based on these drawings without any creative effort.

Fig. 1 is the overall schematic diagram of the method of the present invention;

Fig. 2 is the schematic diagram of pretreatment process in the present invention;

3 is a schematic diagram of a beamforming process in the present invention;

4 is a schematic diagram of sound source propagation in the present invention;

5 is a schematic diagram of the directional function in the present invention;

6 is a schematic diagram of a post-Wiener filtering process in the present invention;

7 is a graph showing the comparison results of the present invention and other noise reduction methods PESQ when a single noise source has different signal-to-noise ratios;

8 is a graph showing the comparison results of the present invention and other noise reduction methods PESQ when there are multiple noise sources with different signal-to-noise ratios;

FIG. 9 is a graph showing the comparison result of SegSNR between the present invention and other noise reduction methods when a single noise source has different signal-to-noise ratios;

FIG. 10 is a graph showing the comparison result of SegSNR between the present invention and other noise reduction methods when there are multiple noise sources and different signal-to-noise ratios;

FIG. 11 is a schematic diagram of the result when the desired sound source directions are different in the present invention.

Detailed ways

In order to make the technical solutions and advantages of the present invention clearer, the technical solutions in the embodiments of the present invention will be described clearly and completely below with reference to the accompanying drawings in the embodiments of the present invention:

As shown in FIG. 1 , a dual-microphone noise reduction method with adjustable sound source directions is included in the implementation process: a preprocessing process, a beamforming process, and a post-Wiener filtering process. The specific steps of the method disclosed in the present invention are as follows:

S1: The preprocessing process is shown in Figure 2. After discrete sampling, pre-emphasis, framing and windowing of the noisy signals x ₁ (t) and x ₂ (t) received by the dual microphones, they are obtained by short-time Fourier transform. The frequency domain signals X ₁ (ω) and X ₂ (ω) are specifically adopted as follows:

S11: First perform discrete sampling on the noisy continuous signals x ₁ (t) and x ₂ (t), and the sampling frequency is 16 kHz. The pre-emphasis on the high-frequency part of the speech is realized by a first-order FIR high-pass digital filter, where EMP_FAC is the pre-emphasis coefficient. Assuming that the sample value of the speech signal at time n is x(n), the result after pre-emphasis processing is:

z(n)=x(n)-EMP_FAC*x(n-1) (1)

Taking EMP_FAC=0.8, after the noise reduction process, the short-time Fourier transform is used to synthesize the time-domain signal, and a de-emphasis operation is needed to restore the high-frequency part.

S12: Divide the sampled signals x ₁ (n) and x ₂ (n) into frames, the frame length is 10ms, and add a Hamming window of equal length to the framed signals. The window function formula is as formula (2), after adding the window The signal passes into the buffer area, and the length of the buffer area is 5 times the number of FFT points.

S13: Obtain the current frame frequency domain signal through fast Fourier transform, and output the signal of the first 1/2 frequency points for subsequent algorithm processing according to the conjugate symmetry of the fast Fourier transform of the real number sequence.

S2: The beamforming process is shown in Figure 3. A virtual microphone is introduced at the midpoint of the two-microphone connection, and the obtained frequency domain signals X ₁ (ω) and X ₂ (ω) are differentially transformed according to the central difference format to construct a differential Signals Y ₁ (ω) and Y ₂ (ω), calculate the statistical mean of the power spectra of Y ₁ (ω) and Y ₂ (ω), and record the ratio of the statistical mean as the directional function Γ(ω,θ ). Analyzing the properties of Γ(ω,θ), it can be directly mapped to the noise masking value through the normalization function. The specific method is as follows:

S21: Amplitude alignment, although the sound field is assumed to be far field, the received signals of the dual microphones have subtle differences in amplitude. In order to further meet the assumption, firstly, the two frequency domain signals X ₁ (ω) and X ₂ (ω) are multiplied by scale factors to align their amplitudes.

S22: Assuming that the desired beam is S(ω), its direction is preset to α, and a virtual microphone is introduced at the midpoint of the two-way microphone to receive the signal. The schematic diagram of sound source propagation is shown in Figure 4. The spatial relationship between X ₁ (ω), X ₂ (ω) and S(ω) is:

Among them, d is the distance between the microphones, v is the speed of sound, and θ is the actual incident direction of the sound source; according to the central difference format and the spatial relationship between X ₁ (ω), X ₂ (ω) and S (ω), the differential signal Y ₁ is constructed (ω) and Y ₂ (ω), and calculate the power spectrum of Y ₁ (ω) and Y ₂ (ω).

S23: The ratio of the statistical mean values of the power spectra of Y ₁ (ω) and Y ₂ (ω) is the value of the directional function Γ(ω, θ).

The image of Γ(ω, θ) is shown in Figure 5. It is found that when the actual incident direction θ of the sound source is equal to the given desired sound source incident direction α, Γ(ω, θ) tends to infinity; and on both sides of the θ=α axis , the function value is monotonic and approximately symmetrical;

S24: Since when θ=α, the infinity that Γ(ω, θ) tends to cannot be reached in actual calculation, a threshold Ω needs to be calculated according to the preset main lobe width Θ.

Since the sigmoid function is a sigmoid function with a value range of (0, 1), through a certain deformation, Γ(ω, θ) can be directly normalized and mapped to the noise mask value λ(ω) of the corresponding frequency point.

S25: The sound after masking the noise of the competing direction is:

R ₁ (ω)=λ(ω)X ₁ (ω) (8)

S3: The post-Wiener filtering process is shown in Figure 6. By estimating the signal energy and noise energy in R ₁ (ω), the channel SNR is obtained and the gain function is calculated to further eliminate the residual noise in R ₁ (ω), Specifically, the following methods are used:

S31: Divide R ₁ (ω) into NUM_CHAN channels according to the critical bandwidth criterion. Since the speech energy is mainly concentrated in 0.3-3.4kHz, the corresponding channel is relatively narrow at low frequency; the corresponding channel is relatively wide at high frequency. For each channel, estimate its energy, where β is the smoothing factor, M is the number of intermediate frequency points in the current channel, m is the label of the current frame channel, i is the label of the current frame, and k is the current channel. The label of the intermediate frequency point.

S32: Initialize the channel noise estimate to the channel energy of the first four frames, and the signal-to-noise ratio can be calculated by (10):

S33: Design a nonlinear data table as a speech index table, and map the SNR index (quantized SNR value) to a set of numbers that measure speech quality. When the signal-to-noise ratio is high, the voice quality is considered to be high, and the sum of the voice indicators in the frequency range of 0.3 to 3.4 kHz is calculated.

S34: Calculate the total noise energy estimation (tne) and the total energy estimation (tce) of the first HI_CHAN channels, that is, the total noise energy and the channel energy in the frequency range of 0.3-3.4 kHz.

S35: Calculate the log spectrum of the current channel energy, and record the deviation between the long-term log spectrum energy and the short-term log spectrum energy as ch_enrg_dev.

ch_enrg_db(i,m)=10lg(ch_enrg(i,m)) (13)

S36: Calculate the long-term integral constant alpha, alpha is a function of the total channel energy (tce), that is, when the tce is high (-40dB), it is smoothed slowly (alpha=0.99); when the tce is low (-60dB), it is smoothed quickly (alpha = 0.50).

S37: Calculate and update the long-term logarithmic spectral energy.

S38 : Reset the noise update flag Update_flag through comparison according to the calculated speech index sum, signal-to-noise ratio, logarithmic spectral deviation and other parameters. "Update_flag=TRUE" indicates that the current frame is a noise frame, and "Update_flag=FALSE" indicates that the current frame is a speech frame. After that, it is necessary to check the noise update flags of the previous frames. If the noise cannot be updated for a long time, the results are considered unreliable, and the SNR index needs to be updated forcibly.

S39: Calculate the channel gain ftmp2 by using the obtained channel SNR index.

If the noise update flag Update_flag=TRUE, it is considered that the current frame is determined as a noise frame, and the energy estimation of the noise needs to be updated at this time.

To verify the effectiveness of the present invention, several tests were performed. It should be noted that in order to verify that the method is suitable for multiple types of sounds, the speech data used for evaluation comes from the TIMIT database, and the noise includes Babble noise and competing direction speech. The experimental results in the present invention are all averaged results obtained by processing 10 segments of speech data.

The present invention is compared with two methods, Coherence and SNR-coherence, first setting α=0°. Figures 7 and 8 show the PESQ scores of different methods after adding various noises (including competing speech and Babble noise). Obviously, the present invention is superior to the Coherence method, and comparable to SNR-coherence. In general, the PESQ results of the present invention are at least 0.5 higher than the unprocessed signal, and this effect is maintained under conditions of multiple noise sources.

Figures 9 and 10 show that when the interference is non-target speech and Babble noise, in the case of low SNR (-5dB and 0dB), the SegSNR value of the present invention is improved by at least 5dB compared to the unprocessed one. The SegSNR results of the present invention are both higher than the SNR-Coherence method and almost equal to the Coherence method. Furthermore, the present invention maintains optimum performance in the presence of multiple noise sources.

Meanwhile, the evaluation results when the desired direction is set at other angles are shown in FIG. 11 . It can be seen that, compared with the sound before processing, the present invention can still maintain a good noise suppression capability.

The above comparative tests all show the good noise reduction performance and good working stability of the present invention.

The above description is only a preferred embodiment of the present invention, but the protection scope of the present invention is not limited to this. The equivalent replacement or change of the inventive concept thereof shall be included within the protection scope of the present invention.

Claims (4)

1. A method for reducing noise of a dual microphone with adjustable expected sound source direction is characterized by comprising the following steps:

the pretreatment process comprises the following steps: for noisy signal x received by dual microphones₁(t) and x₂(t) discrete sampling, pre-emphasis, framing and windowing, and performing short-time Fourier transform to obtain frequency domain signal X₁(omega) and X₂(ω)；

And (3) beam forming process: introducing a virtual microphone at the midpoint of the two-microphone line, and applying the frequency domain signal X according to the central difference format₁(omega) and X₂(omega) performing differential conversion to construct a differential signal Y₁(omega) and Y₂(ω) calculating a difference signal Y₁(omega) and Y₂(ω) a statistical average of the power spectrum, and the ratio of the statistical average is recorded as a directivity function Γ (ω, θ), the properties of the directivity function Γ (ω, θ) are analyzed, it is directly mapped to a noise masking value λ (ω) by a normalization function, and the frequency domain signal X is converted into a frequency domain signal X₁Multiplying (omega) by the noise masking value lambda (omega) to obtain a signal R with the competing direction noise eliminated₁(ω)；

Post wiener filtering process: to R₁Estimating signal energy and noise energy in (omega) to obtain channel signal-to-noise ratio and calculating gain function, thereby eliminating signal R₁Residual noise in (ω).

2. The method of claim 1, wherein: the pretreatment process comprises the following steps:

will carry the noise signal x₁(t) and x₂(t) discrete sampling is carried out, and then pre-emphasis processing is carried out on the high-frequency part of the voice;

sampling signal x₁(n) and x₂(n) dividing the signals into frames with the length of 10ms, adding equal-length Hamming windows w (n), introducing the windowed signals into a buffer area for processing, obtaining the frequency domain signals of the current frame through short-time Fourier transform, and outputting the signals of the first 1/2 frequency points for beam forming processing according to the conjugate symmetry of real number sequence Fourier transform.

3. The method of claim 1, wherein: the beam forming process comprises amplitude alignment, power spectrum calculation, directivity function value calculation, threshold calculation and normalization mapping;

the amplitude alignment mode is to the frequency domain signal X₁(omega) and X₂(ω) multiplying by a scaling factor respectively for amplitude alignment;

when calculating the power spectrum: assuming that the desired beam is S (co),the direction of the two-way microphone is preset to be alpha, a virtual microphone is introduced at the midpoint of the two-way microphone to receive a desired beam S (omega), and a central difference format and a frequency domain signal X are used₁(ω)、X₂The spatial relationship of (ω) to the desired beam S (ω) constructs a differential signal Y₁(omega) and Y₂(ω) calculating a difference signal Y₁(omega) and Y₂(ω) a power spectrum;

when calculating the directional function value: wherein the differential signal Y₁(omega) and Y₂(ω) the ratio of the statistical average of the power spectra is the value of the directivity function Γ (ω, θ), which tends to infinity when the actual sound source incidence direction θ is equal to the given desired sound source incidence direction α, and which functions monotonically and approximately symmetrically on both sides of the α -axis, discussing the nature of Γ (ω, θ);

when calculating the threshold and normalizing the mapping: as gamma (omega, theta) tends to be infinite, a threshold omega is calculated according to a preset main lobe width theta, the gamma (omega, theta) is directly mapped into a noise masking value lambda (omega) of a corresponding frequency point through normalized mapping of a sigmoid function, and X is used for mapping X to the noise masking value lambda (omega) of the corresponding frequency point₁Multiplying (omega) by lambda (omega) to obtain a signal R with the competing direction noise eliminated₁(ω)

4. The method of claim 1, wherein: the post-wiener filtering process comprises the steps of calculating a signal-to-noise ratio index, calculating a logarithmic spectrum deviation, modifying or resetting a noise mark and calculating a gain function value;

when calculating the SNR index, it willSignal R₁(omega) dividing the channel into a plurality of channels according to a critical bandwidth criterion, estimating the energy of each channel, initializing the channel noise energy estimation into the channel energy of the first four frames, and calculating a channel signal-to-noise ratio index according to the channel noise energy estimation;

when calculating the logarithmic spectrum deviation, designing a nonlinear data table as a voice index table, mapping the signal-to-noise ratio index into a group of numbers for measuring the voice quality, taking the sum of the voice indexes in a certain frequency range as the voice quality evaluation result of the current channel, taking the logarithm of the signal energy of the current channel, and calculating the deviation of the long-time logarithmic spectrum energy and the short-time logarithmic spectrum energy;

modifying or resetting the noise mark, judging whether the current frame is a voice frame or a noise frame according to the calculated voice index sum, the signal-to-noise ratio index and the log spectrum deviation parameter information, resetting the noise updating mark, checking the updating marks of the previous frames, and if the noise cannot be updated for a long time and the result is unreliable, forcibly updating the signal-to-noise ratio index;

when the gain function value is calculated, the channel signal-to-noise ratio index is used for calculating the channel gain value to remove residual background noise, and the noise energy estimation of the next frame is updated according to the result of the noise updating mark.

Images