Birdsong species automatic identification method based on ECAPA-TDNN
By combining the ECAPA-TDNN model and the Gaussian mixture model, automatic identification of bird song species and silent detection are achieved, solving the problems of low identification accuracy and low efficiency of manual segmentation in existing technologies, and providing efficient automatic segmentation and a user-friendly identification interface.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- NANJING UNIV OF SCI & TECH
- Filing Date
- 2023-04-23
- Publication Date
- 2026-05-12
AI Technical Summary
Existing bird song species identification algorithms lack highly accurate neural network models, and manually extracting bird song fragments is labor-intensive and subject to observer bias, affecting in-depth analysis.
The ECAPA-TDNN model is used for automatic identification of bird song species, and a Gaussian mixture model is used for silence detection to achieve automatic segmentation and species identification of song audio. Mel frequency cepstral coefficients are used for feature extraction, and the identification results are displayed through a graphical user interface.
It improves the accuracy of bird song species identification, realizes automatic segmentation and efficient identification of song audio, and provides a convenient user interface for preliminary analysis and display.
Smart Images
Figure CN116524939B_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of acoustic monitoring and audio signal recognition technology, and in particular, it is an automatic bird song species identification method based on ECAPA-TDNN. Background Technology
[0002] Birds, as an important component of ecosystems, provide crucial data for ecologists to understand a region's biodiversity and climate change. Passive acoustic monitoring, with its low cost, wide range, and non-invasive nature, makes bird calls a vital data source for monitoring bird activity. Given the increasingly severe ecological degradation and numerous challenges in ecological restoration in recent years, species identification, behavioral analysis, and acoustic index research based on bird calls have immense application value.
[0003] Currently, the main bird song-based species identification algorithms are: 1) Template matching-based classification methods, such as the dynamic time programming template algorithm, which suffers from high computational cost and low efficiency; 2) Traditional machine learning algorithms, such as random forests, support vector machines, and hidden Markov models, which are greatly affected by noise and require a high signal-to-noise ratio for the dataset; 3) Species identification using deep learning, such as AlexNet and VGG16 models, which are currently popular, but rarely applied to bird song-based species identification in China.
[0004] The ECAPA-TDNN model, proposed in 2020, introduces an SE (Squeeze Excitation) module and a channel attention mechanism, enabling the model to learn more global information from audio data. It has become a mainstream voiceprint model. Baidu's Paddle Speech open-source voiceprint recognition system utilizes ECAPA-TDNN to extract voiceprint features, achieving an error rate as low as 0.95%.
[0005] Therefore, a major problem with existing technologies is the lack of validated, highly accurate mainstream neural network models in the field of bird song species identification. Furthermore, current bird song recognition algorithms require manual extraction of segments containing only bird sounds for prediction. When inputting long audio clips of bird songs, the manual workload is substantial and subject to observer bias, hindering in-depth analysis and subsequent research. Summary of the Invention
[0006] The purpose of this invention is to address the problems existing in the prior art by providing an automatic bird song species identification method based on ECAPA-TDNN, which improves the accuracy of identification while achieving automatic segmentation of song audio and species identification.
[0007] The technical solution to achieve the purpose of this invention is: an automatic bird song species identification method based on ECAPA-TDNN, the method comprising the following steps:
[0008] Step 1: Collect and preprocess bird song signals to construct a bird song dataset. Then, obtain Mel frequency cepstral coefficients through feature extraction.
[0009] Step 2: Input the Mel frequency cepstral coefficients into the ECAPA-TDNN network for training to obtain the ECAPA-TDNN bird song classification model;
[0010] Step 3: Input the bird song audio, and use the speech endpoint detection algorithm based on Gaussian mixture model to remove silent segments and extract the segments containing bird songs;
[0011] Step 4: Based on the segment containing bird calls, extract the Mel frequency cepstral coefficients in the manner described in Step 1, input them into the ECAPA-TDNN bird call classification model for recognition, and obtain the recognition result.
[0012] Furthermore, the preprocessing described in step 1 specifically includes:
[0013] Step 1-1-1: For bird call signals, eliminate the DC component and pre-emphasize the signal;
[0014] Step 1-1-2: Perform high-pass filtering;
[0015] Step 1-1-3: Perform frame segmentation processing;
[0016] Step 1-1-4: Add windows using Hanning windows.
[0017] Furthermore, the pre-emphasis in step 1-1-1 is specifically achieved through the transfer function H(z) = 1 - az. -1 The first-order FIR high-pass digital filter is implemented, where a is the pre-emphasis coefficient.
[0018] Furthermore, the step 1, which involves obtaining Mel frequency cepstral coefficients through feature extraction, specifically includes:
[0019] Step 1-2-1: Perform a short-time Fourier transform on the bird song signals in the bird song dataset, take the absolute value of the obtained value, and then square it to obtain the energy spectrum.
[0020] Step 1-2-2: Construct the Mel filter bank and perform a dot product operation with the energy spectrum to obtain the Mel spectrum.
[0021] Steps 1-2-3: Take the logarithm of the Mel spectrogram;
[0022] Steps 1-2-4 involve performing a discrete cosine transform and taking the first P data points to obtain the Mel frequency cepstral coefficients; where P is an integer.
[0023] Furthermore, the ECAPA-TDNN network includes convolutional neural layers, three SE-Res2Block layers, an Attentive Statistics Pooling layer, and a fully connected layer;
[0024] The Mel frequency cepstral coefficients are input into the ECAPA-TDNN network model for training. The specific process includes:
[0025] Mel frequency cepstral coefficients are input into the convolutional neural layer to obtain potential audio features;
[0026] The potential audio features are fused using the SE-Res2Block layer to extract global information.
[0027] The outputs of the three SE-Res2Block layers are concatenated according to the feature dimension;
[0028] The mean and standard deviation based on the attention mechanism are obtained through the Attentive Statistics Pooling layer, and then concatenated according to the feature dimensions to obtain a vector.
[0029] The vector is classified using softmax through a fully connected layer to obtain the classification result;
[0030] Based on the classification results, the network parameters are updated through backpropagation using the cross-entropy loss function to obtain the ECAPA-TDNN bird song classification model.
[0031] Furthermore, step 3 specifically includes the following processes:
[0032] Step 3-1: Preprocess and frame the bird song audio. The preprocessing process is the same as in Step 1. Create multiple classes and store the bird song audio segments into the classes.
[0033] Step 3-2, silence assessment, specifically includes:
[0034] Step 3-2-1: Divide the segment into subbands and calculate the logarithmic energy of the subbands;
[0035] Step 3-2-2: When the total frame energy is greater than the minimum energy required to trigger the audio signal, for each sub-band, calculate the probability P(X|H1) of speech based on the Gaussian mixture model and the probability P(X|H0) of noise based on the Gaussian mixture model.
[0036] Step 3-2-3: Calculate the likelihood ratio of the subband based on the two probabilities mentioned above: Likelihood ratio = log(P(X|H1) / P(X|H0));
[0037] Step 3-2-4: If the likelihood ratio of a certain sub-band meets the preset threshold, it is determined to be an audio segment.
[0038] The likelihood ratio of each sub-band is summed up to obtain the overall likelihood ratio. If the overall likelihood ratio meets the preset threshold, it is determined to be an audio segment.
[0039] Otherwise, it will be judged as a silent segment;
[0040] Step 3-3, audio clip collection, specifically includes:
[0041] Step 3-3-1: Create a data container with two ends, add an object, and get the number of classes with sound.
[0042] Step 3-3-2: When the number of audio clips exceeds 90% of the data container capacity, it is determined that birdsong has started, and the data in the current data container is written into the constructed empty list.
[0043] Step 3-3-3: Repeat steps 3-3-1 and 3-3-2. When the number of silent segments is greater than 90%, end the list writing.
[0044] Step 3-3-4 returns a list of data, which consists of only segments containing bird calls.
[0045] Further, step 3-2-1 specifically includes: based on the spectral characteristics and energy distribution of bird calls, dividing each frame of bird call signal into six sub-bands: 200-2000Hz, 2000-3000Hz, 3000-3500Hz, 3500-4500Hz, 4500-8000Hz, and 8000-24000Hz; calculating the sub-band energy and the total energy; and taking the logarithm.
[0046] Furthermore, the method also includes:
[0047] Step 5: Display the recognition results one by one on the graphical user interface, perform quantity statistics by category, draw a spectrum diagram, and display the recognition information in the exported table, including the start and end times of bird calls, recognition results, and similarity.
[0048] Furthermore, step 5 specifically includes the following processes:
[0049] Step 5-1: Build a graphical user interface to display the segmented fragment index, recognition results, and similarity scores one by one, and count the number of birdsongs by species.
[0050] Step 5-2: Plot the spectrogram of the input bird song audio.
[0051] Step 5-3: Create a new table to store the start and end times of bird calls, recognition results, and similarity information.
[0052] Compared with the prior art, the significant advantages of this invention are:
[0053] 1) Compared with traditional machine learning, deep learning can learn audio latent features quickly and accurately. The ECAPA-TDNN model emphasizes channel attention mechanism and multi-layer feature fusion. Experimental results show that the accuracy of bird song species classification is significantly improved.
[0054] 2) The silence detection algorithm designed based on the Gaussian mixture model in this invention realizes automatic segmentation of the sounding record, with high accuracy and good segmentation effect.
[0055] 3) The user graphical interface designed in this invention is convenient and intuitive. Users can independently select the audio to be identified. The statistical, graphing, and table functions realize the preliminary analysis and display of the audio to be identified.
[0056] The present invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description
[0057] Figure 1 This is a flowchart of the bird song species automatic identification method based on ECAPA-TDNN of the present invention.
[0058] Figure 2 This is a structural diagram of ECAPA-TDNN.
[0059] Figure 3 This is a diagram of the SE-Res2Block structure of ECAPA-TDNN.
[0060] Figure 4 This is a flowchart of the silence detection process based on the Gaussian mixture model. Detailed Implementation
[0061] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.
[0062] It should be noted that if the embodiments of the present invention involve directional indicators (such as up, down, left, right, front, back, etc.), the directional indicators are only used to explain the relative positional relationship and movement of the components in a certain specific posture (as shown in the figure). If the specific posture changes, the directional indicators will also change accordingly.
[0063] Furthermore, if the embodiments of this invention involve descriptions such as "first" or "second," these descriptions are for descriptive purposes only and should not be construed as indicating or implying their relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined with "first" or "second" may explicitly or implicitly include at least one of those features. Additionally, the technical solutions of the various embodiments can be combined with each other, but this must be based on the ability of those skilled in the art to implement them. If the combination of technical solutions is contradictory or impossible to implement, it should be considered that such a combination of technical solutions does not exist and is not within the scope of protection claimed by this invention.
[0064] In one embodiment, combined Figure 1 This paper provides an automatic bird song species identification method based on ECAPA-TDNN, the method comprising the following steps:
[0065] Step 1: Collect bird song signals and preprocess them to form a relatively pure bird song dataset. Then, obtain Mel frequency cepstral coefficients through feature extraction.
[0066] The preprocessing includes:
[0067] Step 1-1-1: For the bird song signal, eliminate the DC component and pre-emphasize it, specifically through the transfer function H(z) = 1 - az. -1 A first-order FIR high-pass digital filter is implemented, where a is the pre-emphasis coefficient, a = 0.97, to enhance the high-frequency components in the signal;
[0068] Step 1-1-2: Perform high-pass filtering; since speech noise and environmental noise are mainly concentrated below 350Hz, the signal is passed through an 8th-order Butterworth high-pass filter with a cutoff frequency of 350Hz to obtain a relatively pure bird sound signal.
[0069] Step 1-1-3: Perform frame splitting, with a frame length of 2048 and a frame shift of 512.
[0070] Step 1-1-4: Use Hanning windowing to add a window to eliminate the discontinuity between frames after framing.
[0071] The process of obtaining Mel frequency cepstral coefficients through feature extraction specifically includes:
[0072] Step 1-2-1: Perform a short-time Fourier transform on the bird song signals in the bird song dataset, take the absolute value of the obtained value, and then square it to obtain the energy spectrum.
[0073] Step 1-2-2: Construct a Mel filter bank with 128 filters, and perform a dot product operation with the energy spectrum to obtain the Mel spectrum.
[0074] Steps 1-2-3: Take the logarithm of the Mel spectrogram;
[0075] Steps 1-2-4 involve performing a discrete cosine transform and taking the first 80 data points to obtain the Mel frequency cepstral coefficients; where P is an integer.
[0076] Step 2: Input the Mel frequency cepstral coefficients into the ECAPA-TDNN network (a time-delay neural network emphasizing channel attention, propagation, and aggregation) for training to obtain the ECAPA-TDNN bird song classification model; combined with... Figure 2 The ECAPA-TDNN network comprises convolutional neural layers, three SE-Res2Block layers, an Attentive Statistics Pooling layer, and a fully connected layer. The specific process includes:
[0077] Step 2-1: Mel frequency cepstral coefficients are passed through a convolutional neural layer to obtain potential audio features;
[0078] Combination Figure 3 Step 2-2: The potential audio features are fused using the SE-Res2Block layer to extract global information.
[0079] Steps 2-3: Concatenate the outputs of the three SE-Res2Block layers according to the feature dimension;
[0080] Steps 2-4: Through the Attentive Statistics Pooling layer, the mean and standard deviation based on the attention mechanism are obtained, and concatenated according to the feature dimensions to obtain a 3072-dimensional vector;
[0081] Steps 2-5: Perform softmax classification on the vector using a fully connected layer to obtain the classification result;
[0082] Steps 2-6: Based on the classification results, the network parameters are updated through backpropagation using the cross-entropy loss function to obtain the ECAPA-TDNN bird song classification model.
[0083] Step 3: Input a 1-minute audio clip of birdsong. Using a Gaussian mixture model-based speech endpoint detection algorithm, remove silent segments and extract the segments containing birdsong. Combined with... Figure 4 Specifically, it includes:
[0084] Step 3-1: Preprocess and frame the bird song audio. The preprocessing process is the same as in Step 1. Create multiple classes and divide the bird song audio into segments every 100ms and store them in the classes.
[0085] Step 3-2, silence assessment, specifically includes:
[0086] Step 3-2-1: Divide the segment into sub-bands and calculate the logarithmic energy of each sub-band; Based on the spectral characteristics and energy distribution of bird calls, divide each frame of bird call signal into six sub-bands: 200-2000Hz, 2000-3000Hz, 3000-3500Hz, 3500-4500Hz, 4500-8000Hz, and 8000-24000Hz. Calculate the sub-band energy and the total energy, and take the logarithm.
[0087] Step 3-2-2: When the total frame energy is greater than the minimum energy required to trigger the audio signal, for each sub-band, calculate the probability P(X|H1) of speech based on the Gaussian mixture model and the probability P(X|H0) of noise based on the Gaussian mixture model.
[0088] Step 3-2-3: Calculate the likelihood ratio of the subband based on the two probabilities mentioned above: Likelihood ratio = log(P(X|H1) / P(X|H0));
[0089] Step 3-2-4: If the likelihood ratio of a certain sub-band meets the preset threshold, it is determined to be an audio segment (local).
[0090] Alternatively, the likelihood ratios of each sub-band can be summed to obtain the overall likelihood ratio. If the overall likelihood ratio meets a preset threshold, it is determined to be an audio segment (global). Here, as long as either the global or local likelihood ratio is met, it is considered an audio segment.
[0091] Otherwise, it will be judged as a silent segment;
[0092] Step 3-3, audio clip collection, specifically includes:
[0093] Step 3-3-1: Create a data container with two ends, add an object (an instance of the Frame class, the silence detection result), and get the number of classes with sound.
[0094] Step 3-3-2: When the number of audio clips exceeds 90% of the data container capacity, it is determined that birdsong has started, and the data in the current data container is written into the constructed empty list.
[0095] Step 3-3-3: Repeat steps 3-3-1 and 3-3-2. When the number of silent segments is greater than 90%, end the list writing.
[0096] Step 3-3-4 returns a list of data, which consists of only segments containing bird calls.
[0097] Step 4: Based on the segment containing bird calls, extract the Mel frequency cepstral coefficients in the manner described in Step 1, input them into the ECAPA-TDNN bird call classification model for recognition, and obtain the recognition result.
[0098] Step 5: Display the recognition results one by one on the graphical user interface, perform quantity statistics by category, plot a spectrogram, and display the recognition information in the exported table, including the start and end times of bird calls, recognition results, and similarity scores. Specifically, this includes:
[0099] Step 5-1: Build a graphical user interface to display the segmented fragment index, recognition results, and similarity scores one by one, and count the number of birdsongs by species.
[0100] Step 5-2: Plot the spectrogram of the input bird song audio.
[0101] Step 5-3: Create a new Excel spreadsheet to store the start and end times of bird calls, recognition results, similarity information, etc.
[0102] In one embodiment, an automatic bird song species identification system based on ECAPA-TDNN is provided, the system comprising:
[0103] The first module is used for preprocessing and feature extraction of bird song signals to obtain Mel frequency cepstral coefficients;
[0104] The second module is used to train the ECAPA-TDNN network based on Mel frequency cepstral coefficients to obtain the ECAPA-TDNN bird song classification model;
[0105] The third module is used to perform silence detection, remove silent segments, and generate a dataset containing only bird song segments;
[0106] The fourth module is used to preprocess and extract features from the segmented bird song fragments, and then identify them using the ECAPA-TDNN bird song classification model.
[0107] The fifth module is used to enable user interaction and display classification and analysis results.
[0108] Specific limitations regarding the ECAPA-TDNN-based automatic bird song species identification system can be found in the above section on the limitations of the ECAPA-TDNN-based automatic bird song species identification method, and will not be repeated here. Each module in the aforementioned ECAPA-TDNN-based automatic bird song species identification system can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in hardware or independent of the processor in a computer device, or stored in software in the memory of a computer device, so that the processor can call and execute the corresponding operations of each module.
[0109] In one embodiment, a computer device is provided, including a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the following steps:
[0110] Step 1: Collect and preprocess bird song signals to construct a bird song dataset. Then, obtain Mel frequency cepstral coefficients through feature extraction.
[0111] Step 2: Input the Mel frequency cepstral coefficients into the ECAPA-TDNN network for training to obtain the ECAPA-TDNN bird song classification model;
[0112] Step 3: Input the bird song audio, and use the speech endpoint detection algorithm based on Gaussian mixture model to remove silent segments and extract the segments containing bird songs;
[0113] Step 4: Based on the segment containing bird calls, extract the Mel frequency cepstral coefficients in the manner described in Step 1, input them into the ECAPA-TDNN bird call classification model for recognition, and obtain the recognition result;
[0114] Step 5: Display the recognition results one by one on the graphical user interface, perform quantity statistics by category, draw a spectrum diagram, and display the recognition information in the exported table, including the start and end times of bird calls, recognition results, and similarity.
[0115] For specific limitations on each step, please refer to the limitations of the ECAPA-TDNN-based automatic bird song species identification method above, which will not be repeated here.
[0116] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, the computer program performing the following steps when executed by a processor:
[0117] Step 1: Collect and preprocess bird song signals to construct a bird song dataset. Then, obtain Mel frequency cepstral coefficients through feature extraction.
[0118] Step 2: Input the Mel frequency cepstral coefficients into the ECAPA-TDNN network for training to obtain the ECAPA-TDNN bird song classification model;
[0119] Step 3: Input the bird song audio, and use the speech endpoint detection algorithm based on Gaussian mixture model to remove silent segments and extract the segments containing bird songs;
[0120] Step 4: Based on the segment containing bird calls, extract the Mel frequency cepstral coefficients in the manner described in Step 1, input them into the ECAPA-TDNN bird call classification model for recognition, and obtain the recognition result;
[0121] Step 5: Display the recognition results one by one on the graphical user interface, perform quantity statistics by category, draw a spectrum diagram, and display the recognition information in the exported table, including the start and end times of bird calls, recognition results, and similarity.
[0122] For specific limitations on each step, please refer to the limitations of the ECAPA-TDNN-based automatic bird song species identification method above, which will not be repeated here.
[0123] This invention is convenient, fast, practical, and highly accurate. It fully leverages the advantages of deep learning to achieve automatic segmentation and species identification of bird song segments, which is of great significance for studying the biodiversity of ecosystems and protecting endangered birds.
[0124] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely illustrative of the principles of the invention. Any modifications, equivalent substitutions, or improvements made within the spirit and principles of the present invention without departing from its spirit and scope should be included within the protection scope of the present invention.
Claims
1. An automatic bird song species identification method based on ECAPA-TDNN, characterized in that, The method includes the following steps: Step 1: Collect and preprocess bird song signals to construct a bird song dataset. Then, obtain Mel frequency cepstral coefficients through feature extraction. Step 2: Input the Mel frequency cepstral coefficients into the ECAPA-TDNN network for training to obtain the ECAPA-TDNN bird song classification model; Step 3: Input the bird song audio, and use the speech endpoint detection algorithm based on Gaussian mixture model to remove silent segments and extract the segments containing bird songs; Step 4: Based on the segment containing bird calls, extract the Mel frequency cepstral coefficients in the manner described in Step 1, input them into the ECAPA-TDNN bird call classification model for recognition, and obtain the recognition result; Step 3 specifically includes the following processes: Step 3-1: Preprocess and frame the bird song audio. The preprocessing process is the same as in Step 1. Create multiple classes and store the bird song audio segments into the classes. Step 3-2, silence assessment, specifically includes: Step 3-2-1: Divide the segment into subbands and calculate the logarithmic energy of the subbands; Step 3-2-2: When the total frame energy is greater than the minimum energy required to trigger the audio signal, for each sub-band, calculate the probability P(X|H1) of speech based on the Gaussian mixture model and the probability P(X|H0) of noise based on the Gaussian mixture model. Step 3-2-3: Calculate the likelihood ratio of the subband based on the two probabilities mentioned above: Likelihood ratio = log(P(X|H1) / P(X|H0)); Step 3-2-4: If the likelihood ratio of a certain sub-band meets the preset threshold, it is determined to be an audio segment. Alternatively, the likelihood ratios of each sub-band can be summed as the overall likelihood ratio. If the overall likelihood ratio meets a preset threshold, it is determined to be an audio segment. Otherwise, it will be judged as a silent segment; Step 3-3, audio clip collection, specifically includes: Step 3-3-1: Create a data container with two ends, add an object, and get the number of classes with sound. Step 3-3-2: When the number of audio clips exceeds 90% of the data container capacity, it is determined that birdsong has started, and the data in the current data container is written into the constructed empty list. Step 3-3-3: Repeat steps 3-3-1 and 3-3-2. When the number of silent segments is greater than 90%, end the list writing. Step 3-3-4 returns a list of data, which consists of only segments containing bird calls.
2. The method for automatic bird song species identification based on ECAPA-TDNN according to claim 1, characterized in that, The preprocessing described in step 1 specifically includes: Step 1-1-1: For bird call signals, eliminate the DC component and pre-emphasize the signal; Step 1-1-2: Perform high-pass filtering; Step 1-1-3: Perform frame segmentation processing; Step 1-1-4: Add windows using Hanning windows.
3. The method for automatic bird song species identification based on ECAPA-TDNN according to claim 2, characterized in that, In step 1-1-1, the pre-emphasis is specifically achieved through the transfer function as follows: The implementation of a first-order FIR high-pass digital filter, wherein This is the pre-emphasis coefficient.
4. The method for automatic bird song species identification based on ECAPA-TDNN according to claim 2, characterized in that, Step 1, which involves obtaining Mel frequency cepstral coefficients through feature extraction, specifically includes: Step 1-2-1: Perform a short-time Fourier transform on the bird song signals in the bird song dataset, take the absolute value of the obtained value, and then square it to obtain the energy spectrum. Step 1-2-2: Construct the Mel filter bank and perform a dot product operation with the energy spectrum to obtain the Mel spectrum. Steps 1-2-3: Take the logarithm of the Mel spectrogram; Steps 1-2-4 involve performing a discrete cosine transform and taking the first P data points to obtain the Mel frequency cepstral coefficients; where P is an integer.
5. The method for automatic bird song species identification based on ECAPA-TDNN according to claim 1, characterized in that, The ECAPA-TDNN network includes a convolutional neural layer, three SE-Res2Block layers, an Attentive Statistics Pooling layer, and a fully connected layer. The Mel frequency cepstral coefficients are input into the ECAPA-TDNN network model for training. The specific process includes: Mel frequency cepstral coefficients are input into the convolutional neural layer to obtain potential audio features; The potential audio features are fused using the SE-Res2Block layer to extract global information. The outputs of the three SE-Res2Block layers are concatenated according to the feature dimension; The Attentive Statistics Pooling layer is used to obtain the mean and standard deviation based on the attention mechanism, which are then concatenated according to the feature dimensions to obtain a vector. The vector is classified using softmax through a fully connected layer to obtain the classification result; Based on the classification results, the network parameters are updated through backpropagation using the cross-entropy loss function to obtain the ECAPA-TDNN bird song classification model.
6. The method for automatic bird song species identification based on ECAPA-TDNN according to claim 1, characterized in that, Step 3-2-1 specifically includes: based on the spectral characteristics and energy distribution of bird calls, dividing each frame of bird call signal into six sub-bands: 200-2000Hz, 2000-3000Hz, 3000-3500Hz, 3500-4500Hz, 4500-8000Hz, and 8000-24000Hz; calculating the sub-band energy and the total energy; and taking the logarithm.
7. The method for automatic bird song species identification based on ECAPA-TDNN according to claim 1, characterized in that, The method further includes: Step 5: Display the recognition results one by one on the graphical user interface, perform quantity statistics by category, draw a spectrum diagram, and display the recognition information in the exported table, including the start and end times of bird calls, recognition results, and similarity.
8. The method for automatic bird song species identification based on ECAPA-TDNN according to claim 7, characterized in that, Step 5 includes the following specific steps: Step 5-1: Build a graphical user interface to display the segmented fragment index, recognition results, and similarity scores one by one, and count the number of birdsongs by species. Step 5-2: Plot the spectrogram of the input bird song audio. Step 5-3: Create a new table to store the start and end times of bird calls, recognition results, and similarity information.
9. An automatic bird song species identification system based on ECAPA-TDNN, based on the method of any one of claims 1 to 8, characterized in that, The system includes: The first module is used for preprocessing and feature extraction of bird song signals to obtain Mel frequency cepstral coefficients; The second module is used to train the ECAPA-TDNN network based on Mel frequency cepstral coefficients to obtain the ECAPA-TDNN bird song classification model; The third module is used to perform silence detection, remove silent segments, and generate a dataset containing only bird song segments; The fourth module is used to preprocess and extract features from the segmented bird song fragments, and then identify them using the ECAPA-TDNN bird song classification model. The fifth module is used to enable user interaction and display classification and analysis results.
10. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method as described in any one of claims 1 to 8.