CN109741759A - An acoustic automatic detection method for specific bird species - Google Patents

Abstract

The invention discloses a kind of acoustics automatic testing methods towards specific birds species.This method first uses the potential song segment of the specific birds species of acoustic events detection processing procedure extraction based on gauss hybrid models, in conjunction with based on candidate sound event energy, the duration that pipes, frequency distribution last handling process, complete robust detection and automatic segmentation；Then self-adapting signal noise reduction is carried out to each sound clip, corresponding mel cepstrum coefficients characteristic parameter is extracted to enhanced sound clip, obtains potential song segment characterizations collection；Potential song segment characterizations collection and support vector machines is finally combined to complete specific birds species Acoustic detection.The present invention realizes that process is convenient; furthermore the signal-to-noise ratio of sound event can be obviously improved by using the adaptive signal enhancement processing based on microphone array to improve particular species identification accuracy rate; the Acoustic detection for realizing specific birds species under the natural environment of field, is of great significance for the ecological protection of field Rare Birds species.

Description

An acoustic automatic detection method for specific bird species

technical field

The invention belongs to the field of ecological monitoring and acoustic signal technology identification, and particularly relates to an acoustic automatic detection method for specific bird species.

Background technique

Birds are widely distributed vocal animals. Compared with other animal communities, the living habits of birds are more easily observed by humans, and they can sensitively perceive small changes in the ecological environment. Therefore, birds are considered by most ecologists. It is an ideal species for monitoring ecological environment changes.

However, in recent years, due to the continuous expansion of human urbanization and the continuous destruction of the ecological environment by human activities, the number and species of birds have decreased on a large scale. Indicative species of woodland vegetation community, the sharp decline of birds will lead to the imbalance of ecological environment. Therefore, the protection of birds is the focus of attention in the ecological field, especially the supervision of rare birds is the top priority. Traditional pointcounts are costly and invasive of habitats when monitoring birds. In order to assess bird activities more quickly and non-invasively, the need to effectively implement automatic bird detection is increasingly urgent.

Among them, based on the song signals of bird species, the use of acoustic signal analysis methods to extract features from the field measured bird sound signals is the basis for subsequent large-scale data analysis and bird sound recognition model establishment. In a real and complex environment, considering that the bird species identification and classification method based on characteristic parameters is highly sensitive to environmental noise and other interfering sound source signals, many scholars are actively exploring methods for identifying bird calls in noise. Chinese patent CN102930870A discloses a bird sound recognition method using anti-noise power normalized cepstral coefficients. The method first uses multi-band spectral subtraction to perform noise reduction processing on the sound power spectrum, and then extracts the noise power spectrum after noise reduction. The anti-noise power normalized cepstral coefficients, and finally combined with Support Vector Machine (SVM) to identify 34 kinds of bird sounds under different environments and signal-to-noise ratios. Chinese patent CN103489446A discloses a bird song recognition method based on self-adaptive energy detection in a complex environment. In this method, the sound signal is first subjected to energy detection, and the sound signal frames selected by the detection are extracted from the wavelet packet decomposition sub-band inversion based on Mel scale. Spectral coefficient (Wavelet Packet decomposition Subband Cepstral Coefficient, WPSCC) anti-noise feature, combined with SVM to classify and identify 15 types of bird song in noisy environment. Liu Zhao et al. introduced a bird sound recognition algorithm in noisy environment based on random forest and large-scale acoustic feature extraction (Liu Zhao, Zhang Yuchen, Hu Hailong. Simulation of bird sound recognition in noisy environment with random forest and large-scale acoustic features [J] ]. System Simulation Technology, 2017(4):359-362.). Zhang Saihua et al. first used the Gaussian Mixture Model (GMM)-based acoustic event detection process to extract potential bird song segments, and then extracted the Mel subband-based parametric features of each segment, and used SVM to analyze the 11 categories in the wild environment. Classification and recognition of birdsong

Most of the above-mentioned existing studies only add one kind of noise or do not suppress the noise in the experiment. However, the field acoustic environment is very complex, with a variety of environmental interference sources. Only single-channel audio enhancement front-end processing cannot meet the actual field complex acoustics. Requirements for acoustic monitoring tasks in the environment.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide an acoustic automatic detection method for specific bird species.

Experiment The technical solution for the purpose of the present invention is: an acoustic automatic detection method for a specific bird species, comprising the following steps:

Step 1. Collect field continuous bird sound monitoring data signals and perform automatic segmentation, and then extract potential song fragments of specific bird species;

Step 2, performing adaptive signal noise reduction enhancement processing on each potential sound segment obtained in step 1;

Step 3, extracting feature parameters for each potential sound segment after noise reduction and enhancement in step 2, and constructing a feature set of potential sound segments;

Step 4. Acoustic detection of a specific bird species is completed by combining the feature set of potential song segments obtained in step 3 and the recognition algorithm in machine learning.

Compared with the prior art, the present invention has the following significant advantages: 1) In the present invention, a multi-dimensional stereo microphone array is used to collect continuous bird sound monitoring data. Monitoring of bird species; 2) The present invention can complete robust detection and detection by detecting potential song events of specific bird species based on a Gaussian mixture model, combined with a post-processing process based on the energy, song duration, and frequency distribution of candidate sound events. Automatic segmentation; 3) The present invention improves the signal-to-noise ratio of sound events by adopting the adaptive signal noise reduction enhancement processing based on the microphone array, thereby improving the identification accuracy rate of specific species; 4) The method of the present invention is convenient and easy to implement. implement.

The present invention will be described in further detail below with reference to the accompanying drawings.

Description of drawings

FIG. 1 is a flow chart of an acoustic automatic detection method for a specific bird species according to the present invention.

FIG. 2 is a flow chart of the detection of potential song events for a specific bird species according to the present invention.

FIG. 3 is a schematic diagram of the adaptive delay estimation of the P-element stereo microphone array.

FIG. 4 is a structural block diagram of a generalized sidelobe canceller of a P-element stereo microphone array.

FIG. 5 is a schematic diagram of adaptive delay estimation of a 4-element stereo microphone array.

FIG. 6 is a structural block diagram of a generalized sidelobe canceller of a 4-element stereo microphone array.

Detailed ways

1, the acoustic automatic detection method for specific bird species of the present invention includes the following steps:

Step 1. Collect field continuous bird sound monitoring data signals and perform automatic segmentation, and then extract potential song fragments of specific bird species.

Further, in conjunction with Fig. 2, step 1 is specifically:

Step 1-1. Use a multi-dimensional stereo microphone array to collect multi-channel field continuous bird sound monitoring data signals, and perform pre-emphasis processing on the collected field continuous bird sound monitoring data signals to compensate for the excessive attenuation of high-frequency signals and suppress low-frequency signals. wind noise;

Step 1-2, performing framing, windowing and fast Fourier transform on the continuous bird sound monitoring data signal processed in step 1-1 to obtain a power spectrogram;

Steps 1-3, set the lower and upper frequency limits as f _L and f _H respectively, and determine the short-term logarithmic energy le(l) of each frame. The formula used is:

le(l)=log ₁₀ (e(l))

in,

In the formula, l is the frame serial number, i is the frequency serial number, S(i,l) represents the short-time Fourier transform result at the time-frequency point (i,l), _{NL and NH represent f L and f} respectively The frequency point sequence number corresponding to _H , e(l) is the short-term energy of the lth frame;

Steps 1-4, using a Gaussian mixture model containing two Gaussian components to generate a frame logarithmic energy distribution, then the two Gaussian components respectively represent the probability density function of the potential sound event frame set and the environmental noise frame set;

Steps 1-5: For each frame, determine whether the frame belongs to a potential sound segment or an environmental noise segment by comparing the posterior probability, and obtain several potential sound segments, specifically:

Compare the posterior probability of each frame belonging to the set of potential whistling event frames with the posterior probability of belonging to the set of environmental noise frames, if the posterior probability of the frame belonging to the set of potential whistling event frames is greater than the posterior probability of belonging to the set of environmental noise frames , then the frame belongs to a potential sound segment, and other frames that are temporally continuous with the frame and also meet the above conditions also belong to the potential sound segment of the segment, thereby obtaining several potential sound segments, all potential sound segments. The sound segment constitutes a set D={AE ₁ ,AE ₂ ,...,AE _K }, where K is the number of potential sound segments;

Steps 1-6. Calculate the logarithmic energy of each potential sound segment obtained in steps 1-5. The formula used is:

and get the largest logarithmic energy ME among them:

For the k-th potential vocal segment, if ME-EAE _k ≥q, the potential vocal segment is considered to be a weak segment with less ecological research value, and it is eliminated, where q is a preset threshold according to the actual situation, The unit of q is dB;

Steps 1-7, based on the existing specific bird species song database data, obtain the upper and lower thresholds of the specific bird species song segment song duration through statistical analysis, that is, the longest song duration _tH and the shortest song duration _tL , and according to The signal sampling rate f _s translates t _H and t _L into a maximum tweet length n _H and a minimum tweet length n _L :

n _H = f _s ×t _H

n _L = f _s ×t _L

For each potential sound segment obtained in steps 1-6, obtain its length T as:

T = frame length × number of frames in the potential sound segment

Eliminate potential sound segments whose length T is less than n _L and greater than n _H ;

Steps 1-8, based on the existing sound database data of specific bird species, obtain the frequency distribution range of specific bird species song fragments through statistical analysis, and the potential song fragments obtained in steps 1-7 will exceed the frequency range data is set to zero.

Further, the q in steps 1-6 is taken as 20dB.

Step 2: Perform adaptive signal noise reduction enhancement processing on each potential sound segment obtained in Step 1.

Further, step 2 performs adaptive signal noise reduction enhancement processing on each potential sound segment obtained in step 1, specifically:

Assuming that the multi-element stereo microphone array is a P-element stereo microphone array, the channels of the P-element stereo microphone array are numbered 1, 2, 3...., P in a certain order;

Step 2-1. Combined with Figure 3, the adaptive filtering method is used to estimate the sound source direction for each potential sound segment, specifically:

Step 2-1-1. For the P channel signal data of one of the potential sound segments, it is assumed that the P channel signals are m ₁ (n), m ₂ (n), m ₃ (n), ..., m _P (n), n=1,2,3,...,L _m , L _m is the signal length, and the signal of channel 1 is used as the reference signal to construct the snapshot x _k of the signal of channel 2:

T

x _k =[m ₂ (k),m ₂ (k+1),...,m ₂ (k+L-1)];

In the formula, the subscript k=1,2,...,L _m -L+1 represents the k-th snapshot, L represents the filter length, and the superscript T represents the transposition;

Step 2-1-2, obtain the autocorrelation matrix R _xx , the formula used is:

In the formula, K=L _m -L+1 is the number of snapshots;

Step 2-1-3, to obtain the cross-correlation matrix r _xd , the formula used is:

In the formula, is the filter center point;

Step 2-1-4, to obtain the weight vector w ₁ , the formula used is:

w ₁ =R _xx ⁻¹ r _xd ;

Step 2-1-5, perform peak detection on the weight vector w ₁ obtained in step 2-1-4, and mark the abscissa of the peak value as z, then the number of delay points between the channel 1 signal and the channel 2 signal d ₁ =zD;

Step 2-1-6, repeat steps 2-1-1 to 2-1-5 to obtain the delay points d _c of the channel 1 signal and the c-th channel signal, c=1,2,...,P -1;

Step 2-2, with reference to Fig. 4, use a generalized sidelobe canceller to adaptively enhance the potential sound segment, specifically:

Step 2-2-1. Find the main channel signal d(k):

d(k)=w _c ^T m(k);

In the formula, w _c =[w _c1 ,w _c2 ,..,w _cP ] ^T is the static weight vector, w _c1 , w _c2 , .., w _cP are the corresponding weights of each channel and w _c1 +w _c2 + ...+w _cP =1; m(k)=[m ₁ (k),m ₂ (kd ₁ ),...,m _P (kd _P-1 )] ^T , k=1,2,. ..,L _m ;

Step 2-2-2, find the auxiliary channel signal e(k):

Among them, W _S is the blocking matrix of dimension P×(P-1);

Step 2-2-3. Find the pure potential sound segment signal y(k) after enhancement:

y(k)=d(k)-v ^T (k)e(k);

where v(k) represents the dynamic weight vector of the adaptive interference canceler.

Further, step 2 performs adaptive signal noise reduction enhancement processing on each potential sound segment obtained in step 1, specifically:

Assuming that the multi-element stereo microphone array is P=4-element stereo microphone array; the 4-element stereo microphone array is numbered 1, 2, 3, 4 in a certain order;

Step 2-1, with reference to Fig. 5, adopt the adaptive filtering method to estimate the sound source direction for each potential sound segment, specifically:

Step 2-1-1. For the 4-channel signal data of one of the potential sound segments, it is assumed that the 4-channel signals are m ₁ (n), m ₂ (n), m ₃ (n), and m ₄ (n) respectively. , n=1,2,3,...,L _m , L _m is the signal length, and the channel 4 signal is used as the reference signal to construct the snapshot x _k of the channel 1 signal:

x _k =[m ₁ (k),m ₁ (k+1),...,m ₁ (k+L-1)] ^T ;

In the formula, the subscript k=1,2,...,L _m -L+1 represents the k-th snapshot, L represents the filter length, and the superscript T represents the transposition;

Step 2-1-2, obtain the autocorrelation matrix R _xx , the formula used is:

In the formula, K=L _m -L+1 is the number of snapshots;

Step 2-1-3, to obtain the cross-correlation matrix r _xd , the formula used is:

In the formula, is the filter center point;

Step 2-1-4, to obtain the weight vector w ₁ , the formula used is:

w ₁ =R _xx ⁻¹ r _xd ;

Step 2-1-5, perform peak detection on the weight vector w ₁ obtained in step 2-1-4, and mark the abscissa of the peak value as z, then the number of delay points between the channel 4 signal and the channel 1 signal d ₁ =zD;

Step 2-1-6, repeating steps 2-1-1 to 2-1-5, to obtain the delay points of the channel 4 signal, the channel 2 signal, and the channel 3 signal, which are respectively d ₂ and d ₃ ;

Step 2-2, with reference to Fig. 6, use a generalized sidelobe canceller to adaptively enhance the potential sound segment, specifically:

Step 2-2-1. Find the main channel signal d(k):

d(k)=w _c ^T m(k);

In the formula, w _c =[w _c4 ,w _c1 ,w _c2 ,w _c3 ] ^T is the static weight vector, w _c1 ,...,w _c4 are the corresponding weights of each channel and w _c1 +w _c2 +w _c3 +w _c4 =1, m(k)=[m ₄ (k), m ₁ (kd ₁ ), m ₂ (kd ₂ ), m ₃ (kd ₃ )] ^T , k=1,2,... , L _m ;

Step 2-2-2, find the auxiliary channel signal e(k):

Among them, W _S is the blocking matrix of dimension 4 × 3;

Step 2-2-3. Find the pure potential sound segment signal y(k) after enhancement:

y(k)=d(k)-v ^T (k)e(k);

where v(k) represents the weight vector of the adaptive interference canceler.

Step 3: Extract feature parameters for each potential sound segment after noise reduction and enhancement in step 2, and construct a feature set of potential sound segments.

Further, step 3 is specifically:

Step 3-1, according to step 1-2, calculate the power spectrum of each potential sound segment after noise reduction enhancement;

Step 3-2, set up a Mel bandpass filter group within the frequency distribution range of the vocal segment of a specific bird species, and then pass the power spectrum of the potential vocal segment through the filter group to obtain the output of each filter;

Step 3-3, take the logarithm of the output result, and perform discrete cosine transform to obtain the characteristic parameters of Mel cepstral coefficients;

Step 3-4, combining the characteristic parameters of Mel cepstral coefficients corresponding to all potential singing segments to construct a feature set of potential singing segments.

Step 4. Acoustic detection of a specific bird species is completed by combining the feature set of potential song segments obtained in step 3 and the recognition algorithm in machine learning.

Further, the identification algorithm in step 4 specifically adopts the support vector machine identification algorithm.

Further, step 4 is specifically:

Taking the Mel cepstral coefficient feature in the existing vocal feature database of a specific bird species as the training sample, and using the feature set of potential vocal segments as the input sample of the support vector machine, through the decision of the support vector machine, it can automatically detect the specific bird species.

To sum up, an acoustic automatic detection method for a specific bird species of the present invention adopts the GMM detection of potential song events of a specific bird species combined with a post-processing process based on the energy, frame length, and frequency distribution of candidate sound events. , complete robust detection and automatic segmentation. In addition, the present invention can significantly improve the signal-to-noise ratio of sound events by adopting the adaptive noise reduction enhancement processing based on the microphone array, thereby improving the identification accuracy of specific species, and realizing the acoustic detection of specific bird species in the wild natural environment. The ecological protection of species and related ecological research are of great significance.

Claims (8)

1. an acoustic automatic detection method for specific bird species, is characterized in that, comprises the following steps: Step 1. Collect field continuous bird sound monitoring data signals and perform automatic segmentation, and then extract potential song fragments of specific bird species; Step 2, performing adaptive signal noise reduction enhancement processing on each potential sound segment obtained in step 1; Step 3, extracting feature parameters for each potential sound segment after noise reduction and enhancement in step 2, and constructing a feature set of potential sound segments; Step 4. Acoustic detection of a specific bird species is completed by combining the feature set of potential song segments obtained in step 3 and the recognition algorithm in machine learning. 2 . The acoustic automatic detection method for specific bird species according to claim 1 , characterized in that, according to step 1, the continuous bird sound monitoring data signal in the field is collected and automatically segmented, and then the potential song of the specific bird species is extracted. 3 . sound clip, which includes the following steps: Step 1-1. Use a multi-dimensional stereo microphone array to collect multi-channel continuous bird sound monitoring data signals in the field, and perform pre-emphasis processing on the collected field continuous bird sound monitoring data signals; Step 1-2, performing framing, windowing and fast Fourier transform on the continuous bird sound monitoring data signal processed in step 1-1 to obtain a power spectrogram; Steps 1-3, set the lower and upper frequency limits as f _L and f _H respectively, and determine the short-term logarithmic energy le(l) of each frame. The formula used is: le(l)=log ₁₀ (e(l)) in, In the formula, l is the frame serial number, i is the frequency serial number, S(i,l) represents the short-time Fourier transform result at the time-frequency point (i,l), _{NL and NH represent f L and f} respectively The frequency point sequence number corresponding to _H , e(l) is the short-term energy of the lth frame; Steps 1-4, using a Gaussian mixture model containing two Gaussian components to generate a frame logarithmic energy distribution, then the two Gaussian components respectively represent the probability density function of the potential sound event frame set and the environmental noise frame set; Steps 1-5: For each frame, determine whether the frame belongs to a potential sound segment or an environmental noise segment by comparing the posterior probability, and obtain several potential sound segments, specifically: Compare the posterior probability of each frame belonging to the set of potential whistling event frames with the posterior probability of belonging to the set of environmental noise frames, if the posterior probability of the frame belonging to the set of potential whistling event frames is greater than the posterior probability of belonging to the set of environmental noise frames , then the frame belongs to a potential sound segment, and other frames that are temporally continuous with the frame and also meet the above conditions also belong to the potential sound segment of the segment, thereby obtaining several potential sound segments, all potential sound segments. The sound segment constitutes a set D={AE ₁ ,AE ₂ ,...,AE _K }, where K is the number of potential sound segments; Steps 1-6. Calculate the logarithmic energy of each potential sound segment obtained in steps 1-5. The formula used is: and get the largest logarithmic energy ME among them: For the k-th potential sound segment, if ME-EAE _k ≥ q, the potential sound segment is eliminated, where q is a preset threshold according to the actual situation, and the unit of q is dB; Steps 1-7, based on the existing specific bird species song database data, obtain the upper and lower thresholds of the specific bird species song segment song duration through statistical analysis, that is, the longest song duration _tH and the shortest song duration _tL , and according to The signal sampling rate f _s translates t _H and t _L into a maximum tweet length n _H and a minimum tweet length n _L : n _H = f _s ×t _H n _L = f _s ×t _L For each potential sound segment obtained in steps 1-6, obtain its length T as: T = frame length × number of frames in the potential sound segment Eliminate potential sound segments whose length T is less than n _L and greater than n _H ; Steps 1-8, based on the existing sound database data of specific bird species, obtain the frequency distribution range of specific bird species song fragments through statistical analysis, and the potential song fragments obtained in steps 1-7 will exceed the frequency range data is set to zero. 3 . The acoustic automatic detection method for a specific bird species according to claim 2 , wherein the q in steps 1-6 is set to be 20 dB. 4 . 4. The acoustic automatic detection method for a specific bird species according to claim 1 or 2, wherein the step 2 performs adaptive signal noise reduction enhancement processing on each potential song segment obtained in step 1, specifically Include the following steps: Assuming that the multi-element stereo microphone array is a P-element stereo microphone array, the channels of the P-element stereo microphone array are numbered 1, 2, 3...., P in a certain order; Step 2-1. Use the adaptive filtering method to estimate the sound source direction for each potential singing segment, specifically: Step 2-1-1. For the P channel signal data of one of the potential sound segments, it is assumed that the P channel signals are m ₁ (n), m ₂ (n), m ₃ (n), ..., m _P (n), n=1,2,3,...,L _m , L _m is the signal length, and the signal of channel 1 is used as the reference signal to construct the snapshot x _k of the signal of channel 2: x _k =[m ₂ (k),m ₂ (k+1),...,m ₂ (k+L-1)] ^T ; In the formula, the subscript k=1,2,...,L _m -L+1 represents the k-th snapshot, L represents the filter length, and the superscript T represents the transposition; Step 2-1-2, obtain the autocorrelation matrix R _xx , the formula used is: In the formula, K=L _m -L+1 is the number of snapshots; Step 2-1-3, to obtain the cross-correlation matrix r _xd , the formula used is: In the formula, is the filter center point; Step 2-1-4, to obtain the weight vector w ₁ , the formula used is: w ₁ =R _xx ⁻¹ r _xd ; Step 2-1-5, perform peak detection on the weight vector w ₁ obtained in step 2-1-4, and mark the abscissa of the peak value as z, then the number of delay points between the channel 1 signal and the channel 2 signal d ₁ =zD; Step 2-1-6, repeat steps 2-1-1 to 2-1-5 to obtain the delay points d _c of the channel 1 signal and the c-th channel signal, c=1,2,...,P -1; Step 2-2, using a generalized sidelobe canceller to adaptively enhance the potential sound segment, specifically: Step 2-2-1. Find the main channel signal d(k): d(k)=w _c ^T m(k); In the formula, w _c =[w _c1 ,w _c2 ,..,w _cP ] ^T is the static weight vector, w _c1 , w _c2 , .., w _cP are the corresponding weights of each channel and w _c1 +w _c2 + ...+w _cP =1; m(k)=[m ₁ (k),m ₂ (kd ₁ ),...,m _P (kd _P-1 )] ^T , k=1,2,. ..,L _m ; Step 2-2-2, find the auxiliary channel signal e(k): Among them, W _S is the blocking matrix of dimension P×(P-1); Step 2-2-3. Find the pure potential sound segment signal y(k) after enhancement: y(k)=d(k)-v ^T (k)e(k); where v(k) represents the dynamic weight vector of the adaptive interference canceler. 5 . The automatic acoustic detection method for specific bird species according to claim 4 , wherein the step 2 performs adaptive signal noise reduction enhancement processing on each potential singing segment obtained in step 1, which specifically includes the following: 6 . step: Assuming that the multi-element stereo microphone array is P=4-element stereo microphone array; the 4-element stereo microphone array is numbered 1, 2, 3, 4 in a certain order; Step 2-1. Use the adaptive filtering method to estimate the sound source direction for each potential singing segment, specifically: Step 2-1-1. For the 4-channel signal data of one of the potential sound segments, it is assumed that the 4-channel signals are m ₁ (n), m ₂ (n), m ₃ (n), and m ₄ (n) respectively. , n=1,2,3,...,L _m , L _m is the signal length, and the channel 4 signal is used as the reference signal to construct the snapshot x _k of the channel 1 signal: x _k =[m ₁ (k),m ₁ (k+1),...,m ₁ (k+L-1)] ^T ; In the formula, the subscript k=1,2,...,L _m -L+1 represents the k-th snapshot, L represents the filter length, and the superscript T represents the transposition; Step 2-1-2, obtain the autocorrelation matrix R _xx , the formula used is: In the formula, K=L _m -L+1 is the number of snapshots; Step 2-1-3, to obtain the cross-correlation matrix r _xd , the formula used is: In the formula, is the filter center point; Step 2-1-4, to obtain the weight vector w ₁ , the formula used is: w ₁ =R _xx ⁻¹ r _xd ; Step 2-1-5, perform peak detection on the weight vector w ₁ obtained in step 2-1-4, and mark the abscissa of the peak value as z, then the number of delay points between the channel 4 signal and the channel 1 signal d ₁ =zD; Step 2-1-6, repeating steps 2-1-1 to 2-1-5, to obtain the delay points of the channel 4 signal, the channel 2 signal, and the channel 3 signal, which are respectively d ₂ and d ₃ ; Step 2-2, using a generalized sidelobe canceller to adaptively enhance the potential sound segment, specifically: Step 2-2-1. Find the main channel signal d(k): d(k)=w _c ^T m(k); In the formula, w _c =[w _c4 ,w _c1 ,w _c2 ,w _c3 ] ^T is the static weight vector, w _c1 ,...,w _c4 are the corresponding weights of each channel and w _c1 +w _c2 +w _c3 +w _c4 =1, m(k)=[m ₄ (k), m ₁ (kd ₁ ), m ₂ (kd ₂ ), m ₃ (kd ₃ )] ^T , k=1,2,... , L _m ; Step 2-2-2, find the auxiliary channel signal e(k): Among them, W _S is the blocking matrix of dimension 4 × 3; Step 2-2-3. Find the pure potential sound segment signal y(k) after enhancement: y(k)=d(k)-v ^T (k)e(k); where v(k) represents the weight vector of the adaptive interference canceler. 6. The acoustic automatic detection method for a specific bird species according to claim 4 or 5, characterized in that, in step 3, feature parameters are extracted from the potential song segment after noise reduction and enhancement in step 2, and a feature of the potential song segment is constructed. set, including the following steps: Step 3-1, according to step 1-2, calculate the power spectrum of each potential sound segment after noise reduction enhancement; Step 3-2, set up a Mel bandpass filter group within the frequency distribution range of the vocal segment of a specific bird species, and then pass the power spectrum of the potential vocal segment through the filter group to obtain the output of each filter; Step 3-3, take the logarithm of the output result, and perform discrete cosine transform to obtain the characteristic parameters of Mel cepstral coefficients; Step 3-4, combining the characteristic parameters of Mel cepstral coefficients corresponding to all potential singing segments to construct a feature set of potential singing segments. 7 . The acoustic automatic detection method for a specific bird species according to claim 6 , wherein the identification algorithm in step 4 specifically adopts a support vector machine identification algorithm. 8 . 8 . The acoustic automatic detection method for specific bird species according to claim 6 , wherein the specific bird is completed by combining the feature set of potential song segments obtained in step 3 and the identification algorithm in machine learning according to step 4 . Acoustic detection of species, specifically: Taking the Mel cepstral coefficient feature in the existing specific bird species song feature database as the training sample, and the potential song segment feature set as the input sample of the support vector machine, through the decision of the support vector machine, the specific bird species.