A method for speech presence judgment and adaptive noise reduction of array signals
By using the white noise gain parameters of adaptive MVDR in array signal processing, the problems of noise suppression and speech signal weakening in array signal processing are solved, and the intelligibility and fidelity of speech signals are improved. Spend.
Patent Information
- Application Number
- CN202210543879.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2022-05-19
- Publication Date
- 2024-12-03
- Estimated Expiration
- 2042-05-19
AI Technical Summary
In array signal processing, the prior art is difficult to effectively reduce noise and enhance voice signals in specific directions. The adaptive MVDR algorithm easily leads to weakening of voice signals during voice signal processing, affecting the intelligibility of the output signal.
The speech existence judgment is performed through the white noise gain (WNG) parameter of the adaptive MVDR, the weighting coefficient is updated according to the judgment results, the WNG value is calculated using the Fourier transform and the covariance matrix, and the speech existence is judged based on the fluctuation threshold and the weighting coefficient is updated.
It realizes accurate recognition of voice signals and reasonable update of weighting coefficients, improves the speech intelligibility and fidelity of the algorithm output, and ensures adaptive noise reduction performance.
Smart Images

Figure CN114927145B_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the technical field of signal processing, and in particular to a method for judging the presence of speech and adaptively reducing noise of an array signal. Background Art
[0002] In array signal processing, how to reduce noise and enhance speech in a specific direction has always been a difficult problem faced by front-end speech signal processing. In practical applications, due to factors such as temperature and device consistency differences, the array beam pointing is distorted. In addition, complex and changeable environmental noise also has a great impact on speech signals.
[0003] The main principle of beamforming is to add a complex weighting coefficient to the microphone signal of each channel, adjust the amplitude and phase of each channel, and make the weighted output result achieve the ideal effect. Take a one-dimensional M-element equidistant linear array as an example. Figure 1 shown.
[0004] Therefore, obtaining its weighting coefficient is the key to the entire algorithm. As a common beamforming algorithm, the MVDR algorithm often fails to achieve its theoretically optimal effect in practical applications. Adaptive noise reduction and speech distortion often compromise each other. When using adaptive MVDR, since there is speech signal in the signal, this measure will weaken the speech signal and affect the intelligibility of the output signal. Summary of the invention
[0005] In view of the above technical deficiencies, the present invention aims to provide a method for speech presence judgment and adaptive noise reduction of array signals, which uses the white noise gain (WNG) characteristic parameter of adaptive MVDR to realize speech presence judgment (VAD). Then, the update of weighting coefficients is determined according to the VAD results.
[0006] In order to solve the above technical problems, the present invention adopts the following technical solution: The present invention provides a method for determining the presence of speech and adaptively reducing noise of an array signal, comprising:
[0007] S1: Perform Fourier transform (FFT) on the input signal of each frame to obtain the frequency domain signal;
[0008] S2: Consider each N consecutive frames of signal as a whole and calculate according to the MVDR formula. If the current frame is not the Nth frame signal of the current segment, only the current frame data is retained and the subsequent algorithm processing is not continued. In the formula, R(f) represents the signal covariance matrix at frequency f, a(θ) represents the direction vector, and w opt represents the complex weighting coefficient of the microphone;
[0009] S3: According to the calculated w optObtain the WNG parameter dB value of the current segment. According to the existing conditions (that is, the upper half of the above formula is 1, and the lower half I is the unit matrix), the formula can be simplified to
[0010] S4: Accumulate the WNG values of the frequency bands with signal characteristics and calculate them as the WNG value of the current band;
[0011] S5: Count the WNG values of M consecutive signal segments, calculate the difference (delta) between the maximum and minimum values of the M WNG values, and compare it with the set fluctuation threshold (Th). If delta is greater than Th, it is determined that there is speech in the current segment, and VAD is 1. Otherwise, it is determined that there is no speech in the current segment, and VAD is 0;
[0012] S6: Determine the current speech weighting coefficient w according to the VAD value opt Whether to update the calculation. If VAD is 0, the weighted value will not be updated and the previous weighted coefficient will be used.
[0013] The beneficial effects of the present invention are:
[0014] The method of the present invention utilizes the changing characteristics of the adaptive WNG value when the speech signal appears, and accurately identifies the appearance time of the speech signal. The update of the weighting coefficient is determined according to the VAD value, which not only retains the performance of MVDR adaptive noise reduction, but also ensures that the speech signal will not be processed, greatly improving the speech intelligibility and fidelity of the algorithm output. BRIEF DESCRIPTION OF THE DRAWINGS
[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative work.
[0016] Figure 1 Schematic diagram of beamforming principle.
[0017] Figure 2 This is a graph showing the changes in the WNG value of a certain corpus.
[0018] Figure 3 It is a schematic diagram of the steps of the method of the present invention.
[0019] Figure 4 Time domain waveform of a corpus input
[0020] Figure 5 The spectrogram of a corpus input.
[0021] Figure 6 The WNG calculation result of a certain corpus.
[0022] Figure 7 This is a time domain waveform diagram of the output result of a certain corpus after the method of the present invention is applied
[0023] Figure 8 This is a spectrum diagram of the output result of a certain corpus after the method of the present invention is used. DETAILED DESCRIPTION
[0024] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0025] The beamforming algorithm is usually the first step in the algorithm module chain, so its input data usually reflects the signal captured by the microphone in the real scene. In the real collected signal, the weighting coefficient will change with the change of the environment. Since the weighting coefficient directly determines the WNG value, the change of the WNG value can reflect the change of the environment. This feature can be used to reasonably track the noise. Take a certain corpus as an example, Figure 2 The change of WNG value in each frequency band of the input signal during the adaptive process is shown, and the high and low dB values are represented by light and dark colors.
[0026] Based on the above analysis, the present invention proposes an adaptive noise reduction method that includes speech presence judgment. The method defines the WNG value of the characteristic signal and its fluctuation threshold. By comparing the WNG value and the fluctuation threshold of the characteristic signal, the weighting coefficient of the input signal can be reasonably updated.
[0027] Specific operation steps (see Figure 3 ) includes: first, performing FFT processing on the current frame input signal to obtain a frequency domain signal. Secondly, solving the covariance matrix of the frequency domain signals of N consecutive frames in sequence according to the frequency band, and the signals of N consecutive frames are regarded as one segment of signal. Using the covariance matrix and the direction vector of the known direction, the adaptive complex weighting coefficient w of MVDR is calculated opt The third step is to use the obtained w opt Calculate the WNG value. Step 4: Select the WNG value of the required feature frequency band and sum it. Step 5: Compare the difference between M consecutive WNGs (i.e., the extreme difference value) and the set fluctuation threshold to determine whether there is speech in the current segment. Step 6: Determine whether to update the weighting coefficient based on the VAD result and apply it to the pre-output signal.
[0028] The present invention is described in detail below with reference to specific embodiments.
[0029] Implementation Cases
[0030] use Figure 2 The time domain waveform and spectrum of the input signal are shown in Figure 4 and 5 As shown in the figure, there is no speech and background noise at the beginning. After a short period of silence, background noise begins to appear, mixed with some speech signals. It can be seen that the speech is mixed in the noise, which is difficult to judge in both time domain and frequency domain. The WNG and value of the characteristic frequency band of the corpus are calculated by adaptive MVDR. The calculation results are shown in the figure. Figure 6 As shown in the figure, it can be seen that no matter whether there is noise in the background environment, the change of WNG value tends to be stable, and only when there is speech will the WNG value change dramatically. Therefore, this change can reflect the change of VAD, which can be used to guide the update of weighting coefficients. Figure 7 and 8 The time domain waveform and spectrum diagram of the output result of the method of the present invention for the corpus are shown respectively. It can be seen that the technical method of the present invention accurately judges the existence of speech in the array input signal containing noise and effectively improves the signal-to-noise ratio.
[0031] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention fall within the scope of the claims of the present invention and their equivalents, the present invention is also intended to include these modifications and variations.
Claims
1. A method for determining the presence of speech and adaptively reducing noise in an array signal, characterized in that: include: Process the input signal of each frame to obtain a frequency domain signal; Take each N consecutive frames of signal as a whole and perform calculation. If the current frame is not the Nth frame signal of the current segment, retain the current frame data and do not perform calculation. After the calculation, a weighting coefficient w is obtained. opt ; According to the calculated weighting coefficient, the WNG value of the current segment is calculated; Accumulate the WNG values of the frequency bands with signal characteristics and calculate them as the WNG value of the current band; Count the WNG values of M consecutive signal segments, calculate the difference between the maximum and minimum values of the M WNG values, and compare it with the set fluctuation threshold; If the difference is greater than the fluctuation threshold, it is determined that there is speech in the current segment, and VAD is 1; otherwise, it is determined that there is no speech in the current segment, and VAD is 0; The VAD value is used to determine whether the weighting coefficient of the current speech segment is updated into the calculation. If the VAD is 1, it is updated; if the VAD is 0, it is not updated and the previous weighting coefficient is used.
2. The method for determining the presence of speech and adaptively reducing noise of an array signal according to claim 1, characterized in that: Perform Fourier transform on the input signal of each frame and obtain the frequency domain signal.
3. The method for determining the presence of speech and adaptively reducing noise of an array signal according to claim 1, characterized in that: The weighting coefficient w opt is the complex weighting coefficient of the microphone; Where R(f) represents the signal covariance matrix at frequency f, and a(θ) represents the direction vector.
4. The method for determining the presence of speech and adaptively reducing noise of an array signal as claimed in claim 3, characterized in that: The WNG value is calculated as follows:
