CN105516876A - Spectrum entropy based howling detection method - Google Patents

Abstract

A howling detection method based on spectral entropy, comprising: performing frame division and windowing on a signal to be detected; performing spectrum analysis on the framed and windowed signal; dividing subbands and calculating the energy of each subband; calculating spectral entropy; For detection and judgment, the spectral entropy threshold is set. If the spectral entropy of the current frame signal is less than the spectral entropy threshold T0, it is judged as a howling frame, otherwise it is judged as a normal signal frame. The howling detection method based on spectral entropy of the present invention overcomes the shortcomings of existing howling detection algorithms that need to dynamically adjust thresholds and have poor robustness in different sound field environments. It has good detection effect in different sound field environments. The experimental results show that, compared with the howling detection algorithm based on PAPR, the detection algorithm based on spectral entropy can significantly reduce the false alarm rate while ensuring a high detection rate.

Description

A Howling Detection Method Based on Spectral Entropy

technical field

The invention relates to a howling detection method. In particular, it relates to a howling detection method based on spectral entropy.

Background technique

Acoustic feedback phenomenon means that in the sound reinforcement system, the microphone picks up sound and converts the acoustic signal into an electrical signal. After being amplified and output by the power amplifier, the sound returns to the microphone through the sound field and then amplified and output by the power amplifier. This repeated cycle forms positive feedback.

According to the Nyquist criterion, when the signal satisfies the phase and gain conditions at the same time, it will generate self-oscillation at the frequency point ω ₀ :

|G(w,t)F(w,t)|≥1

∠G(w,t)F(w,t)=n2π, n is an integer

G(w,t) is the transfer function of the forward path, and F(w,t) is the transfer function of the feedback path.

When the transfer function of the sound field satisfies the above phase and gain conditions, the amplitude of the output signal of the sound reinforcement system will continue to increase, resulting in harsh howling.

By detecting the howling frequency points that appear in the acoustic feedback, the notch processing is performed to reduce the gain at the howling frequency points and destroy the gain conditions for howling, so as to achieve the purpose of howling suppression. Howling detection is the key to the notch filter method. Only by timely and accurately detecting the frequency of howling components can a notch filter corresponding to the center frequency and notch depth be accurately designed. Filter through cascaded notch filters to suppress the occurrence of howling.

Figure 1 shows howling detection process. Since howling is essentially a sinusoidal signal with a single frequency, there are large frequency components in the frequency domain and the energy in the frequency domain is constantly increasing. Based on this, relevant scholars have proposed a series of corresponding howling detection algorithms. Mainly include: PAPR (Peak-to-AveragePowerRatio), PHPR (Peak-to-HarmonicPowerRatio), PNPR (Peak-to-NeighboringPowerRatio), IPMP (InterframePeakMagnitudePersistence), IMSD (InterframeMagnitudeSlopeDeviation).

After the howling component is detected, it is necessary to design a corresponding notch filter to reduce the gain at the howling frequency point, destroy the gain condition for howling, and achieve the purpose of suppressing howling. The most commonly used notch filter is the second-order IIR filter, because the IIR filter can obtain better selection characteristics with fewer orders, uses less storage units, and has fewer operations, which is more economical and efficient.

The system function of the 2nd order IIR filter:

$\mathrm{<span\; class="notranslate">\; h</span>}<span\; class="notranslate">\; (</span>\mathrm{<span\; class="notranslate">\; z</span>}<span\; class="notranslate">\; )</span><span\; class="notranslate">\; =</span>\frac{{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 0</span>}}{\mathrm{<span\; class="notranslate">\; z</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; z</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; b</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; z</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}}{\mathrm{<span\; class="notranslate">\; 1</span>}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 1</span>}}{\mathrm{<span\; class="notranslate">\; z</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}<span\; class="notranslate">\; +</span>{\mathrm{<span\; class="notranslate">\; a</span>}}_{\mathrm{<span\; class="notranslate">\; 2</span>}}{\mathrm{<span\; class="notranslate">\; z</span>}}^{<span\; class="notranslate">\; -</span>\mathrm{<span\; class="notranslate">\; 1</span>}}}<span\; class="notranslate">\; .</span>$

The existing howling detection algorithms need to dynamically adjust the threshold and have poor robustness in different sound field environments.

Contents of the invention

The technical problem to be solved by the present invention is to provide a howling detection method based on spectral entropy which has good detection effect under different sound field environments.

The technical solution adopted in the present invention is: a howling detection method based on spectral entropy, comprising the following steps:

1) Frame and window the signal to be detected;

2) Spectrum analysis is carried out on the signal after frame division and windowing;

3) divide the sub-bands, and calculate the energy of each sub-band;

4) Spectral entropy calculation

According to the subband energy calculated in step 3), the corresponding probability density function and spectral entropy are

$P_x(i,m)=S_x(i,m)/\underset{j=0}{\overset{m-1}{\&Sigma;}}{S}_x(i,j)$ m=0,1,2,...,M-1

$h_x\left(i\right)=\underset{m=0}{\overset{m-1}{\&Sigma;}}{P}_x(i,m)log\frac{1}{P_x(i,m)}$ m=0,1,2,...,M-1

Among them, P _x (i, m) represents the probability density function of the mth subband of the i-th frame signal, and H _x (i) represents the spectral entropy of the i-th frame signal;

5) Detection and judgment

Set the spectral entropy threshold T0. If the spectral entropy of the current frame signal is less than the spectral entropy threshold T0, it is judged as a howling frame, otherwise it is judged as a normal signal frame.

The windowing described in step 1) is a windowing function.

Step 2) described frequency spectrum analysis utilizes FFT analysis to calculate energy spectrum:

$x(i,k)=\underset{\mathrm{no}=0}{\overset{N-1}{\&Sigma;}}x\left(\mathrm{no}\right)w\left(\mathrm{no}\right)e^{-j\frac{2\mathrm{\&pi;}\mathrm{no}k}{N}}$ k=0,1,2,...,N-1

R _x (i,k)=|X(i,k)| ² k=0,1,2,…,N-1

x(n) is the signal to be detected, w(n) is the added window function, N is the data length of FFT, e is the natural base, j represents the imaginary number, X(i,k) is the kth frame signal of i frequency spectrum, R _x (i,k) is the energy spectrum of the kth frequency point of the i-th frame signal.

Step 3) is to divide the entire frequency band into several subbands according to the energy spectrum obtained in step 2), and then calculate the energy of each subband respectively:

$S_x(i,m)=\underset{k\&Element;B_m}{\&Sigma;}{R}_x(i,k)$ m=0,1,2,...,M-1

S _x (i, m) represents the energy of the mth subband of the i-th frame signal, M represents the number of divided subbands, and B _m represents all frequency points corresponding to the mth subband.

The howling detection method based on spectral entropy of the present invention overcomes the shortcomings of existing howling detection algorithms that need to dynamically adjust thresholds and have poor robustness in different sound field environments. It has good detection effect in different sound field environments. The experimental results show that, compared with the howling detection algorithm based on PAPR, the detection algorithm based on spectral entropy can significantly reduce the false alarm rate while ensuring a high detection rate.

Description of drawings

Figure 1 is a schematic diagram of howling detection process;

Figure 2a is a detection effect diagram based on PAPR under the howling signal;

Figure 2b is a schematic diagram of the PAPR and the PAPR threshold of each frame under the howling signal;

Figure 2c is a detection effect diagram based on spectral entropy under howling signals;

Fig. 2d is a schematic diagram of the spectral entropy and the spectral entropy threshold of each frame under the howling signal;

Figure 3a is a detection effect diagram based on PAPR under normal music signals;

Fig. 3b is a schematic diagram of PAPR and PAPR threshold of each frame under normal music signal;

Figure 3c is a detection effect diagram based on spectral entropy under normal music signals;

Figure 3d is a schematic diagram of the spectral entropy and the spectral entropy threshold of each frame under a normal music signal;

Fig. 4 is a flowchart of a howling detection method based on spectral entropy of the present invention.

detailed description

A howling detection method based on spectral entropy of the present invention will be described in detail below with reference to the embodiments and the accompanying drawings.

A howling detection method based on spectral entropy of the present invention comprises the following steps:

1) Perform frame-by-frame windowing on the signal to be detected, and the windowing is a windowing function (such as Hanning window, Hamming window, rectangular window, etc.).

The time-domain discrete signal x(n) is infinitely long. Using FFT to estimate the energy spectrum must limit x(n) to a certain time range, that is, data truncation. Data truncation is equivalent to windowing and framing processing, the length of each frame is FrameLen, and the offset of each frame is ShiftLen. At the same time, in order to minimize the influence of spectrum leakage and inter-spectrum interference and prevent the data from being cut off suddenly, a slowly changing window (such as Hanning window, Hamming window, etc.) should be added to make the spectral sidelobe energy after windowing smaller , the leakage caused by convolution is smaller.

2) Spectrum analysis is carried out to the signal after the frame is added window, and described spectrum analysis is to utilize FFT analysis to calculate and obtain energy spectrum:

$x(i,k)=\underset{\mathrm{no}=0}{\overset{N-1}{\&Sigma;}}x\left(\mathrm{no}\right)w\left(\mathrm{no}\right)e^{-j\frac{2\mathrm{\&pi;}\mathrm{no}k}{N}}$ k=0,1,2,...,N-1

R _x (i,k)=|X(i,k)| ² k=0,1,2,…,N-1

x(n) is the signal to be detected, w(n) is the added window function, N is the data length of FFT, e is the natural base, j represents the imaginary number, X(i,k) is the kth frame signal of i frequency spectrum, R _x (i,k) is the energy spectrum of the kth frequency point of the i-th frame signal.

3) divide sub-band, and calculate the energy of each sub-band, be according to the energy spectrum that step 2) obtains, whole frequency band is divided into several sub-bands, then calculate the energy of each sub-band respectively:

$S_x(i,m)=\underset{k\&Element;B_m}{\&Sigma;}{R}_x(i,k)$ m=0,1,2,...,M-1

S _x (i, m) represents the energy of the mth subband of the i-th frame signal, M represents the number of divided subbands, and B _m represents all frequency points corresponding to the mth subband.

4) Spectral entropy calculation

According to the subband energy calculated in step 3), the corresponding probability density function and spectral entropy are

$P_x(i,m)=S_x(i,m)/\underset{j=0}{\overset{m-1}{\&Sigma;}}{S}_x(i,j)$ m=0,1,2,...,M-1

$h_x\left(i\right)=\underset{m=0}{\overset{m-1}{\&Sigma;}}{P}_x(i,m)log\frac{1}{P_x(i,m)}$ m=0,1,2,...,M-1

Among them, P _x (i, m) represents the probability density function of the m-th subband of the i-th frame signal, and H _x (i) represents the spectral entropy of the i-th frame signal.

5) Detection and judgment

Howling is essentially a sinusoidal signal of a single frequency. When the howling component appears, a frequency component with higher energy will be generated in the frequency domain, and the spectral entropy of the signal is smaller at this time. Therefore, a spectral entropy threshold T0 can be set. If the spectral entropy of the current frame signal is smaller than the spectral entropy threshold T0, it is judged as a howling frame, otherwise it is judged as a normal signal frame.

Specific examples are given below:

Algorithm parameter setting: frame length FrameLen=1024, frame offset ShiftLen=512, number of FFT calculation points N=1024, window function w(n) is Hamming window, length 1024, number of divided subbands M=32, howling The decision threshold T0=0.05.

Select howling test voices in different sound field environments, and test and analyze the proposed howling detection algorithm based on spectral entropy. The specific implementation is as follows:

1. Read howling signal and normal music signal data, and perform frame-by-frame and window processing, each frame has 1024 sampling points, plus 1024-point Hamming window.

2. Perform 1024-point FFT on the windowed data of each frame, and calculate the energy spectrum R _x (i,k) of each frame of data.

3. Divide the entire frequency band into 32 sub-bands, and calculate the energy S _x (i,m) of each sub-band respectively.

4. Calculate the probability density function P _x (i,m) of each sub-band according to the energy of each sub-band, and calculate the spectral entropy H _x (i) of the current frame according to the probability density function.

5. Howling judgment is performed according to the calculated spectral entropy of the current frame. If H _x (i)<T0, it is determined as a howling frame, otherwise it is determined as a normal speech frame.

Claims (4)

1. A howling detection method based on spectral entropy, is characterized in that, comprises the steps: 1) Frame and window the signal to be detected; 2) Spectrum analysis is carried out on the signal after frame division and windowing; 3) divide the sub-bands, and calculate the energy of each sub-band; 4) Spectral entropy calculation According to the subband energy calculated in step 3), the corresponding probability density function and spectral entropy are $P_x(i,m)=S_x(i,m)/\underset{j=0}{\overset{m-1}{\&Sigma;}}{S}_x(i,j)$ m=0,1,2,...,M-1 $h_x\left(i\right)=\underset{m=0}{\overset{m-1}{\&Sigma;}}{P}_x(i,m)log\frac{1}{P_x(i,m)}$ m=0,1,2,...,M-1 Among them, P _x (i, m) represents the probability density function of the mth subband of the i-th frame signal, and H _x (i) represents the spectral entropy of the i-th frame signal; 5) Detection and judgment Set the spectral entropy threshold T0. If the spectral entropy of the current frame signal is less than the spectral entropy threshold T0, it is judged as a howling frame, otherwise it is judged as a normal signal frame. 2. A howling detection method based on spectral entropy according to claim 1, characterized in that the windowing described in step 1) is a windowing function. 3. A kind of howling detection method based on spectral entropy according to claim 1, is characterized in that, the spectrum analysis described in step 2) utilizes FFT analysis to calculate to obtain energy spectrum: $x(i,k)=\underset{\mathrm{no}=0}{\overset{N-1}{\&Sigma;}}x\left(\mathrm{no}\right)w\left(\mathrm{no}\right)e^{-j\frac{2\mathrm{\&pi;}\mathrm{no}k}{N}}$ k=0,1,2,...,N-1 R _x (i,k)=|X(i,k)| ² k=0,1,2,…,N-1 x(n) is the signal to be detected, w(n) is the added window function, N is the data length of FFT, e is the natural base, j represents the imaginary number, X(i,k) is the kth frame signal of i frequency spectrum, R _x (i,k) is the energy spectrum of the kth frequency point of the i-th frame signal. 4. a kind of howling detection method based on spectral entropy according to claim 1, is characterized in that, step 3) is According to the energy spectrum obtained in step 2), the entire frequency band is divided into several sub-bands, and then the energy of each sub-band is calculated separately: $S_x(i,m)=\underset{k\&Element;B_m}{\&Sigma;}{R}_x(i,k)$ m=0,1,2,...,M-1 S _x (i, m) represents the energy of the mth subband of the i-th frame signal, M represents the number of divided subbands, and B _m represents all frequency points corresponding to the mth subband.