An AI-based bird voiceprint analysis method and system

Abstract

The application discloses an AI-based bird soundprint analysis method and system, which comprises the following steps: collecting continuous environmental audio data, generating original audio clips, and constructing a multi-channel sound spectrum input tensor; inputting a double sound source structure coding network, screening a same sound emitting body candidate clip set; clustering and arranging the candidate clip set, constructing a tension debt backfolding chain, and generating a song tension evolution state; performing connection matching based on the tension evolution state, determining a tension foldback closed loop, and determining a song closed loop body; extracting a song structure evolution sequence, constructing a cross-transposition self-isomorphic seal spectrum, screening a stable bird soundprint body; generating a stable soundprint embedding representation, and outputting corresponding bird category and soundprint analysis results. The application realizes accurate determination of the continuous stable soundprint structure of the same sound emitting body in a complex natural sound field through double sound source structure coding, tension debt backfolding chain construction and cross-transposition self-isomorphic seal spectrum matching.

Description

Technical Field

[0001] This invention relates to the field of audio signal processing technology, and in particular to an AI-based method and system for bird voiceprint analysis. Background Technology

[0002] Bird acoustic monitoring is a crucial technique in ecological surveys, biodiversity assessments, habitat protection, and early warning of bird activity. With the development of audio acquisition equipment, edge computing devices, and artificial intelligence technologies, the automatic identification and analysis of bird calls using continuous environmental audio has become an important direction in current bird monitoring. Current technologies typically involve collecting continuous audio data from the field environment using fixed recording equipment or array microphones. This data is then preprocessed, framed, subjected to spectral transformation, and feature extraction. Subsequently, convolutional neural networks, recurrent neural networks, or Transformer-like models are used to classify and identify bird calls, thereby achieving bird species detection, call counting, and activity monitoring.

[0003] However, in complex natural sound fields, bird calls are often accompanied by wind noise, insect chirps, rain sounds, rustling of branches and leaves, and mixed calls from multiple birds in the same area. They are also affected by terrain obstruction, propagation attenuation, and reflected echoes, leading to frequency drift, duration stretching, local breaks, energy misalignment, and missing harmonic structures in multiple call segments produced by the same bird during the same vocalization process. Most existing technologies are based on classifying the entire audio segment or identifying and splicing based on the similarity between local segments. This approach typically only focuses on whether local acoustic features are similar, making it difficult to determine whether multiple disturbed segments truly originate from the same sound source, or whether these segments belong to the same continuous vocalization action chain. This results in problems such as misidentification, missegmentation, and pseudo-sponging in multi-interference scenarios.

[0004] Existing technologies for analyzing bird calls mostly remain at the level of single-segment recognition or short-time sequence matching. They lack effective means to characterize the stable structural skeleton of the same bird call under conditions of modulation, speed change, or local disturbance, making it difficult to identify continuous and stable bird vocal structures at a higher level.

[0005] Therefore, how to provide an AI-based method and system for bird voiceprint analysis is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose an AI-based method and system for bird vocalization analysis. This invention comprehensively employs methods such as multi-channel acoustic spectrum construction, EfficientNet-B1 backbone feature extraction, dual-source structure encoding, tension-debt loop construction, and transmodulation automorphic closure spectrum analysis. It details the implementation process of determining the same vocal source, continuous vocalization action chain, and stable song phrase skeleton for disturbed bird song segments under complex natural sound fields. It has the advantages of strong anti-interference ability, accurate homology determination, good continuity of structural analysis, and high reliability of results.

[0007] According to an embodiment of the present invention, an AI-based bird voiceprint analysis method includes:

[0008] Collect continuous environmental audio data within the target area, process the continuous environmental audio data to generate multiple raw audio segments, and construct a multi-channel acoustic spectrum input tensor for each raw audio segment;

[0009] Each multi-channel acoustic spectrum input tensor is input into a dual-source structure encoding network consisting of an EfficientNet-B1 backbone, a transharmonic borrow redemption chain generation layer, and a dual-source breathing synchronization rectification layer. Local time-frequency texture convolution extraction, transharmonic borrow redemption mapping, and breathing synchronization energy rectification are performed sequentially to generate the first potential sound source trajectory, the second potential sound source trajectory, the shared breathing envelope trajectory, and the dual-source alternating occupancy order. A set of candidate segments of the same sound source is obtained by filtering.

[0010] Cluster the candidate segments of the same vocal body in chronological order to generate multiple candidate vocal phrase segments. Construct a tension debt back-loop chain for each candidate vocal phrase segment group to generate the vocal phrase tension evolution state of each candidate vocal phrase segment group.

[0011] The tension evolution state of the call phrase is sequentially matched according to time sequence to identify the migration relationship between the initial tension state, holding tension state, unloading tension state and retraction state. It is determined whether multiple candidate call phrase segments form a complete tension return loop, and the candidate call phrase segment group that meets the requirements is identified as the call phrase loop body.

[0012] For multiple song closures, the corresponding song structure evolution sequences are extracted, and cross-modal automorphic relationship matching is performed between each song closure to construct a cross-modal automorphic closure spectrum. Stable bird voiceprints are obtained from the cross-modal automorphic closure spectrum.

[0013] A voiceprint embedding representation is generated for a stable bird voiceprint body, and the corresponding bird voiceprint analysis results are output based on the voiceprint embedding representation.

[0014] Optionally, the continuous environmental audio data includes bird call components, wind noise components, insect chirping components, mixed calls of multiple birds in the same domain components, and propagation, reflection, and masking sound components.

[0015] Optionally, the construction of a multi-channel acoustic spectrum input tensor for each original audio segment includes...

[0016] A short-time Fourier transform is performed on each original audio segment to obtain the corresponding time-frequency spectrum matrix. The time-frequency spectrum matrix is ​​mapped to a Mel filter bank, and the logarithm is taken to generate a log-Mel spectrum. Based on the fundamental frequency position, integer harmonic position, and continuous distribution relationship of adjacent harmonics of each frequency component in the time-frequency spectrum matrix, the harmonic components are enhanced to generate a harmonic enhancement map. The energy envelope is extracted based on the time-domain amplitude sequence of each original audio segment, and the envelope change between adjacent time points is calculated according to the time sequence to generate an envelope change map. The log-Mel spectrum, harmonic enhancement map, and envelope change map are overlaid according to the channel dimension to generate the corresponding multi-channel acoustic spectrum input tensor.

[0017] Optionally, the generation of the first potential sound source trajectory, the second potential sound source trajectory, the shared respiratory envelope trajectory, and the alternating occupancy order of the two sound sources, and the selection of a set of candidate segments from the same sound source, includes:

[0018] Each multi-channel acoustic spectrum input tensor is input into a dual-source structure coding network consisting of the EfficientNet-B1 backbone, the cross-harmonic borrow redemption processing layer, and the dual-source breathing synchronous rectification layer. The EfficientNet-B1 backbone extracts the local time-frequency texture features, harmonic continuous distribution features, envelope fluctuation coupling features, and cross-band cooperative variation features of the corresponding original audio segments to generate a basic acoustic text structure feature map.

[0019] The basic acoustic signature feature map is input into the cross-harmonic borrowing and redemption processing layer. Based on the dominant harmonic cluster at each time position, the adjacent harmonic clusters and sideband accompanying clusters are searched along the adjacent harmonic orders. When the dominant harmonic cluster is interrupted, the borrowed harmonic position is determined in the ectopic frequency band according to the frequency migration direction before and after the interruption, the harmonic spacing relationship and the sideband envelope continuity relationship. A redemption mapping with the interrupted dominant harmonic position is established. The redemption mappings in continuous time are sequentially connected to generate a cross-harmonic borrowing and redemption chain feature map.

[0020] The basic acoustic signature structure feature map and the transharmonic borrowing redemption chain feature map are input into the dual-source breathing synchronous rectification layer. The shared breathing initiation boundary, shared breathing expansion boundary, shared breathing release boundary and shared breathing convergence boundary are identified in chronological order. Based on the shared breathing boundary, synchronous rectification is performed on the energy fluctuation process in different frequency bands to generate the shared breathing envelope trajectory.

[0021] Based on the rectified time-frequency structure, the first potential sound source trajectory and the second potential sound source trajectory are separated and generated under the constraint of the shared breathing envelope trajectory. The dual sound source alternation order is generated according to the dominant frequency band position, energy occupation order and alternation switching position of the first potential sound source trajectory and the second potential sound source trajectory in each time segment.

[0022] For the first potential sound source trajectory, the second potential sound source trajectory, the shared breathing envelope trajectory, and the alternating occupation order of the two sound sources among different original audio segments, the same source continuation check is performed. The original audio segments with continuous shared breathing boundaries, consistent alternating relationship between the two sources, and the ability to be connected sequentially across harmonic borrow redemption chains are retained. The retained original audio segments are then merged into the same set of candidate sound source segments.

[0023] Optionally, the step of constructing a tension debt backtracking chain for each candidate singularity fragment group and generating the singularity tension evolution state of each candidate singularity fragment group includes:

[0024] The candidate segments of the same sound source are sorted according to time order. Based on the shared respiratory envelope trajectory, the first potential sound source trajectory, the second potential sound source trajectory, and the alternating occupation order of the two sound sources, clustering and arrangement are performed according to the temporal adjacency relationship, respiratory boundary continuity relationship, dominant frequency band migration relationship, and alternating order relationship of the two sound sources to generate multiple candidate sound sentence segment groups.

[0025] For each candidate phonological segment group, the dominant frequency band bending pattern, harmonic tension pattern, envelope decompression pattern and side band recovery pattern are extracted in chronological order. The tension initiation leader segment, tension holding segment, tension unloading and folding segment and absorption bundle segment are identified to generate a tension segment sequence.

[0026] For the interruption position of adjacent original audio segments, identify the tension migration process that has not been closed. Based on the tension continuity pattern before and after the interruption, divide the missing segment into the pre-release debt segment, the post-absorption debt segment, and the borrowing foldback debt segment.

[0027] For each pre-released debt segment, post-absorption debt segment, and borrowing-back debt segment, cross-segment sequential connection is performed based on the consistent relationship of the dominant frequency band back direction before and after the interruption position, the relationship of harmonic expansion and contraction, the corresponding relationship of envelope release amplitude, the relationship of side band recovery sequence, and the relationship of alternating sequence of dual sound sources. Backward mapping is established for debt segments that can complete tension migration closure, generating tension debt backward chain.

[0028] Based on the initiation leading segment, holding and maintenance segment, unloading and return segment, return absorption bundle segment and tension-debt return chain corresponding to each candidate call phrase segment group, the explicit tension segment and debt return segment are jointly arranged in chronological order to generate the call phrase tension evolution state of each candidate call phrase segment group.

[0029] Optionally, determining whether multiple candidate sound fragments form a complete tension-reversal closed loop, and identifying the group of candidate sound fragments that satisfy the condition as the sound loop body, includes:

[0030] Read the tension evolution state of each candidate phrase fragment group, and arrange the initiation leading segment, holding segment, unloading return segment, return absorption bundle segment and tension debt return chain in chronological order to generate the candidate return sequence corresponding to each candidate phrase fragment group;

[0031] For each candidate turnaround sequence, the tension continuity relationship between the tension initiation leading segment and the tension holding segment, the pressure relief transition relationship between the tension holding segment and the tension release turnaround segment, and the recovery closure relationship between the tension release turnaround segment and the return absorption bundle segment are identified in chronological order. Insertion verification is performed on the tension debt turnaround chain in the candidate turnaround sequence to generate the corresponding tension migration chain.

[0032] For the interruption positions and cross-segment connection positions in each tension migration chain, the migration gaps that have not formed complete foldback closures are identified, and the corresponding dominant frequency band foldback direction, harmonic expansion and contraction sequence, envelope decompression and recovery sequence, shared breathing boundary continuation position, and dual sound source alternation sequence continuation position are extracted around each migration gap. Closed-loop filling verification is performed on each migration gap to generate the corresponding closed-loop verification results.

[0033] Based on the closed-loop verification results, a back-loop determination is performed on each tension migration chain. Tension migration chains in which the initiation pilot segment, tension holding segment, tension unloading back segment and back absorption bundle segment are closed in sequence and the tension debt back segment can complete the migration and replacement are retained. The retained tension migration chains are determined as the back-loop chain.

[0034] For each call phrase fold-back closed loop chain, the closed loop boundary is determined in chronological order. The corresponding starting position, holding and unfolding position, unloading and folding position, return to the absorption bundle position, and closed loop termination position are recorded. Candidate call phrase fragment groups with complete closed loop boundaries are determined as call phrase closed loop bodies.

[0035] Optionally, the construction of the transmodal autoisomorphic closure spectrum, and the selection of stable avian voiceprints from the transmodal autoisomorphic closure spectrum, includes:

[0036] Read the shared respiratory envelope trajectory, first potential sound source trajectory, second potential sound source trajectory, alternating occupation order of dual sound sources and sound tension evolution state corresponding to multiple sound sentence closed loops. Extract the dominant frequency band migration process, duration expansion and contraction process, alternating switching process of dual sound sources and tension return closed loop process corresponding to each sound sentence closed loop in chronological order to generate the sound sentence structure evolution sequence corresponding to each sound sentence closed loop.

[0037] For each call structure evolution sequence, the initial call segment, the upward migration segment, the stable maintenance segment, the folding and terminating segment, and the termination call segment are identified respectively. The sequential connection relationship, cross-segment continuity relationship, dual call source switching position relationship and shared respiratory boundary correspondence between each segment are recorded to generate the structural skeleton sequence corresponding to each call closed loop.

[0038] Cross-modulation alignment processing is performed on the structural skeleton sequences between different ringing closed loops. The overall pitch rise and pitch fall changes are uniformly mapped along the dominant frequency band migration direction. The overall acceleration and deceleration changes are uniformly mapped along the time axis. Based on the mapped correspondence, the skeleton co-position segments, skeleton offset segments and skeleton continuation segments between different ringing closed loops are identified, and candidate automorphic correspondences are generated.

[0039] For each candidate automorphic correspondence, the consistency of the dominant frequency band migration order, the consistency of the alternation of dual sound sources, the consistency of the tension reversal loop order, and the consistency of the shared breathing boundary distribution between the corresponding sound phrase closed loops are checked. Candidate automorphic correspondences that satisfy the consistency relationship are retained. The retained candidate automorphic correspondences are written into the connection relationship between the corresponding sound phrase closed loops to construct the transmodulation automorphic sealing spectrum.

[0040] The closed loops of each call phrase in the transmodal autosomal closed spectrum are closed and screened. The call phrase closed loops that have bidirectional connection, closed loop return and continuous connection in the transmodal autosomal closed spectrum are retained and the retained call phrase closed loops are merged into stable bird voiceprints.

[0041] Optionally, the output of the corresponding bird voiceprint analysis results based on voiceprint embedding representation includes:

[0042] The shared respiratory envelope trajectory, first potential song source trajectory, second potential song source trajectory, alternating occupancy order of dual song sources, song tension evolution state, and song structure evolution sequence corresponding to each stable bird voiceprint are jointly encoded to generate corresponding voiceprint embedding representations. Each voiceprint embedding representation is matched with the template embedding representations in the preset bird voiceprint template library to generate species matching results. Based on the continuity of shared respiratory boundaries, the integrity of tension return loop closure, the self-isomorphic closure of transmodulation, and the template matching, the corresponding analysis confidence level is generated. Based on the matching results, the corresponding bird category results, category confidence results, homologous song merging results, and structural integrity results are output to form the corresponding bird voiceprint analysis results.

[0043] According to an embodiment of the present invention, an AI-based bird voiceprint analysis system includes:

[0044] The audio segment generation module is used to collect continuous environmental audio data and perform processing to generate original audio segments and multi-channel acoustic spectrum input tensors;

[0045] The dual-source encoding module is used to perform structural encoding on the multi-channel acoustic spectrum input tensor, generate the first potential source trajectory, the second potential source trajectory, the shared respiratory envelope trajectory, and the alternating occupancy order of the dual sources, and filter to obtain a set of candidate segments of the same sound source.

[0046] The Singing Tension Construction Module is used to cluster and arrange a set of candidate segments from the same vocal body and construct a tension debt loop chain to generate the singing tension evolution state.

[0047] The tension closed-loop determination module is used to perform succession matching based on the tension evolution state of the syllable and determine the syllable closed-loop body.

[0048] The closure spectrum construction module is used to extract the evolutionary sequence of song structure based on the song closure body and construct the transmodal autosomal closure spectrum to generate a stable bird voiceprint body.

[0049] The voiceprint analysis module is used to generate voiceprint embedding representations of stable bird voiceprints and output bird voiceprint analysis results.

[0050] The beneficial effects of this invention are:

[0051] Compared with existing technologies, this invention combines multi-channel acoustic spectrum input tensors with a dual-source structure coding network to provide a more detailed structural representation of bird calls in continuous environmental audio. Based on this, it generates a first potential source trajectory, a second potential source trajectory, a shared breathing envelope trajectory, and a dual-source alternation order. It no longer relies solely on local spectral similarity to determine the segment affiliation. Instead, it can identify the intrinsic connections between multiple disturbed segments from the perspective of the same sound source and the same breathing constraint. Under complex natural sound field conditions such as wind noise, insect chirps, multiple bird calls in the same domain, and propagation masking, it can effectively reduce the erroneous merging of heterogeneous segments and the erroneous rejection of homogeneous segments, thereby improving the screening accuracy of candidate segment sets from the same sound source.

[0052] This invention constructs a tension debt foldback chain and generates a callus tension evolution state, elevating the relationship between multiple candidate callus fragments from a simple temporal adjacency relationship to a vocal action chain relationship. It can determine the continuity of bird callus fragments with local breaks, misalignments, or local energy deficiencies by focusing on the migration logic between the initial tension state, holding tension state, unloading tension state, and retraction state. Even if multiple callus fragments formed by the same bird in the same vocalization process are interrupted by interference, this invention can still identify fragments belonging to the same continuous callus based on the tension foldback closed-loop relationship, improving the completeness and continuity of callus segmentation results and reducing the problems of missegmentation and pseudo-splitting caused by local segment disturbances in existing technologies.

[0053] This invention extracts the evolutionary sequence of song phrase structures and constructs a transmodal autosomal closure spectrum. It can further identify structural skeletons that remain consistent under conditions of overall pitch rise, overall pitch fall, overall acceleration, and overall deceleration among multiple song phrase closures, obtaining stable bird voiceprints. Based on these stable voiceprints, it generates voiceprint embedding representations and bird voiceprint analysis results. Compared to existing technologies that mainly focus on single-segment classification or short-term matching, this invention can determine stable song phrase skeletons at a higher level. Therefore, it has advantages such as strong anti-interference ability, accurate homology determination, good structural analysis continuity, and high reliability of analysis results, making it more suitable for applications in ecological monitoring, bird identification, habitat surveys, and intelligent early warning systems. Attached Figure Description

[0054] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0055] Figure 1 This is a flowchart of an AI-based bird voiceprint analysis method proposed in this invention;

[0056] Figure 2 This is a schematic diagram of the structure of an AI-based bird voiceprint analysis system proposed in this invention. Detailed Implementation

[0057] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0058] refer to Figure 1 An AI-based method for bird voiceprint analysis includes:

[0059] Collect continuous environmental audio data within the target area, process the continuous environmental audio data to generate multiple raw audio segments, and construct a multi-channel acoustic spectrum input tensor for each raw audio segment;

[0060] Each multi-channel acoustic spectrum input tensor is input into a dual-source structure encoding network consisting of an EfficientNet-B1 backbone, a transharmonic borrow redemption chain generation layer, and a dual-source breathing synchronization rectification layer. Local time-frequency texture convolution extraction, transharmonic borrow redemption mapping, and breathing synchronization energy rectification are performed sequentially to generate the first potential sound source trajectory, the second potential sound source trajectory, the shared breathing envelope trajectory, and the dual-source alternating occupancy order. A set of candidate segments of the same sound source is obtained by filtering.

[0061] Cluster the candidate segments of the same vocal body in chronological order to generate multiple candidate vocal phrase segments. Construct a tension debt back-loop chain for each candidate vocal phrase segment group to generate the vocal phrase tension evolution state of each candidate vocal phrase segment group.

[0062] The tension evolution state of the call phrase is sequentially matched according to time sequence to identify the migration relationship between the initial tension state, holding tension state, unloading tension state and retraction state. It is determined whether multiple candidate call phrase segments form a complete tension return loop, and the candidate call phrase segment group that meets the requirements is identified as the call phrase loop body.

[0063] For multiple song closures, the corresponding song structure evolution sequences are extracted, and cross-modal automorphic relationship matching is performed between each song closure to construct a cross-modal automorphic closure spectrum. Stable bird voiceprints are obtained from the cross-modal automorphic closure spectrum.

[0064] A voiceprint embedding representation is generated for a stable bird voiceprint body, and the corresponding bird voiceprint analysis results are output based on the voiceprint embedding representation.

[0065] In this embodiment, the continuous environmental audio data includes bird call components, wind noise components, insect chirping components, mixed calls of multiple birds in the same area components, and propagation, reflection, and masking sound components.

[0066] In this embodiment, the process of processing continuous environmental audio data to generate multiple original audio segments includes sequentially performing pre-emphasis processing, bandpass filtering processing, and amplitude normalization processing on the continuous environmental audio data to obtain pre-processed audio data; performing frame segmentation processing on the pre-processed audio data according to a preset frame length and a preset frame shift, and performing windowing processing on each frame of audio data to obtain time-series frame data; identifying candidate call segments based on the short-time energy, spectral flux, spectral centroid, and harmonic salience of the time-series frame data, and segmenting the pre-processed audio data according to the start and end positions of the candidate call segments to generate multiple original audio segments.

[0067] In this embodiment, the construction of a multi-channel acoustic spectrum input tensor for each original audio segment includes...

[0068] A short-time Fourier transform is performed on each original audio segment to obtain the corresponding time-frequency spectrum matrix. The time-frequency spectrum matrix is ​​mapped to a Mel filter bank, and the logarithm is taken to generate a log-Mel spectrum. Based on the fundamental frequency position, integer harmonic position, and continuous distribution relationship of adjacent harmonics of each frequency component in the time-frequency spectrum matrix, the harmonic components are enhanced to generate a harmonic enhancement map. The energy envelope is extracted based on the time-domain amplitude sequence of each original audio segment, and the envelope change between adjacent time points is calculated according to the time sequence to generate an envelope change map. The log-Mel spectrum, harmonic enhancement map, and envelope change map are overlaid according to the channel dimension to generate the corresponding multi-channel acoustic spectrum input tensor.

[0069] In this embodiment, the generation of the first potential sound source trajectory, the second potential sound source trajectory, the shared respiratory envelope trajectory, and the alternating occupancy order of the two sound sources, and the selection of a set of candidate segments from the same sound source, includes:

[0070] The multi-channel acoustic spectrum input tensors are input into a dual-source structure coding network consisting of an EfficientNet-B1 backbone, a cross-harmonic borrow redemption processing layer, and a dual-source breathing synchronous rectification layer. The EfficientNet-B1 backbone extracts local time-frequency texture features, harmonic continuous distribution features, envelope fluctuation coupling features, and cross-band cooperative variation features of the corresponding original audio segments to generate a basic acoustic signature feature map. Specifically, the local time-frequency texture features, harmonic continuous distribution features, envelope fluctuation coupling features, and cross-band cooperative variation features extracted from the EfficientNet-B1 backbone for the corresponding original audio segments are as follows:

[0071] Each multi-channel acoustic spectrum input tensor is input into the initial convolutional layer of the EfficientNet-B1 backbone. Initial convolutional encoding is performed on adjacent regions along the time axis and frequency axis to generate a shallow time-frequency response map, and the local time-frequency texture features of the corresponding original audio segment are extracted. The shallow time-frequency response map is then sequentially input into multiple MBConv layers, and deep convolution processing and channel recalibration processing are performed on time-frequency regions of different scales to extract the continuous extension relationship between the dominant frequency band and adjacent frequency bands, the spacing preservation relationship between integer harmonic positions, and the continuous distribution relationship between adjacent harmonic clusters, generating a mid-level harmonic response map, and extracting the continuous harmonic distribution features of the corresponding original audio segment.

[0072] The middle harmonic response map is input into the high-level MBConv layer. The continuous change information of the envelope rising segment, envelope maintaining segment and envelope falling segment is aggregated along the time axis. The synchronous enhancement relationship, synchronous weakening relationship and alternating change relationship of different frequency bands in the same time segment are extracted to generate the high-level cooperative response map. The envelope fluctuation coupling characteristics and cross-frequency band cooperative change characteristics of the corresponding original audio segments are extracted.

[0073] The shallow time-frequency response map, the middle harmonic response map, and the high-level collaborative response map are subjected to uniform scale alignment and channel fusion to generate a basic acoustic text structure feature map.

[0074] The basic acoustic signature structure feature map is input into the cross-harmonic borrowing and redemption processing layer. Using the dominant harmonic cluster at each time position as a reference, adjacent harmonic clusters and sideband accompanying clusters are searched along adjacent harmonic orders. When the dominant harmonic cluster is interrupted, the borrowed harmonic position is determined in the ectopic frequency band based on the frequency migration direction before and after the interruption, the harmonic spacing relationship, and the sideband envelope continuity relationship. A redemption mapping with the interrupted dominant harmonic position is established. The redemption mappings in continuous time are sequentially connected to generate a cross-harmonic borrowing and redemption chain feature map. Specifically, the search for adjacent harmonic clusters and sideband accompanying clusters along adjacent harmonic orders is as follows:

[0075] Starting from the center frequency band of the dominant harmonic cluster at each time position, adjacent search bands are delineated along the high-frequency and low-frequency directions according to the preset harmonic order intervals. Within each adjacent search band, a continuous distribution segment of local energy peaks is extracted to generate corresponding candidate adjacent harmonic segments. Based on the frequency band spacing, frequency migration direction, and time coverage overlap relationship between each candidate adjacent harmonic segment and the dominant harmonic cluster, candidate adjacent harmonic segments that meet the conditions are retained and identified as adjacent harmonic clusters. Lateral search bands are delineated around the upper and lower boundaries of each adjacent harmonic cluster. Within the lateral search bands, accompanying energy distribution segments that fluctuate synchronously with the adjacent harmonic clusters are extracted. Based on the envelope phase change relationship, duration correspondence, and boundary following relationship between the accompanying energy distribution segments and the adjacent harmonic clusters, the corresponding lateral band accompanying clusters are determined.

[0076] Establish a redemption mapping with the location of the dominant harmonic interruption, specifically as follows:

[0077] Using the center frequency band position corresponding to the dominant harmonic position before the interruption and the center frequency band position corresponding to the position after the interruption as the two constraints, the target migration interval corresponding to the dominant harmonic position is determined. Within the target migration interval, the center frequency band position, harmonic spacing state, and envelope change state corresponding to each borrowed harmonic position are read. Each borrowed harmonic position is checked against the dominant harmonic position. Borrowed harmonic positions with consistent frequency migration direction, consistent harmonic spacing with adjacent dominant harmonic clusters, and consistent envelope continuation trend are retained and determined as target borrowed harmonic positions. Using the dominant harmonic position as the mapping start point and the target borrowed harmonic position as the mapping end point, the corresponding frequency band offset direction, frequency band offset length, start time position, and end time position are recorded to establish a redemption mapping with the dominant harmonic position.

[0078] The basic acoustic signature structure feature map and the transharmonic borrow-redemption chain feature map are input into the dual-source breathing synchronous rectification layer. Shared breathing initiation boundary, shared breathing expansion boundary, shared breathing release boundary, and shared breathing convergence boundary are identified sequentially over time. Based on the shared breathing boundaries, synchronous rectification is performed on the energy fluctuation process in different frequency bands to generate a shared breathing envelope trajectory. Specifically, the shared breathing initiation boundary, shared breathing expansion boundary, shared breathing release boundary, and shared breathing convergence boundary are identified sequentially over time as follows:

[0079] The basic acoustic signature structure feature map and the transharmonic borrow-redemption chain feature map are aligned and fused along the time axis to generate a joint energy response sequence at the corresponding time position. Based on the joint energy response sequence, the boundary position from a low-amplitude stable state to a continuously rising state is identified and determined as the shared breathing initiation boundary. The boundary position where the energy increase continues to expand and the coverage frequency range continues to expand along the time sequence is identified and determined as the shared breathing expansion boundary. After the shared breathing expansion boundary, the boundary position where the joint energy response changes from continuous expansion to peak decline and the dominant frequency band begins to contract is identified and determined as the shared breathing release boundary. After the shared breathing release boundary, the boundary position where the joint energy response falls back to a stable low-amplitude state and the frequency band coverage range completes convergence is identified and determined as the shared breathing convergence boundary.

[0080] Synchronous rectification is performed on energy fluctuation processes in different frequency bands based on a shared breathing boundary, specifically as follows:

[0081] The rectification time interval is defined by shared respiratory initiation boundary and shared respiratory termination boundary, and the expansion phase and release phase are divided by shared respiratory expansion boundary and shared respiratory release boundary. Within the rectification time interval, the energy fluctuation sequence corresponding to each frequency band is extracted. The energy rise position of each frequency band is aligned according to the shared respiratory initiation boundary, the energy expansion section of each frequency band is synchronously corrected according to the shared respiratory expansion boundary, the energy fall section of each frequency band is synchronously corrected according to the shared respiratory release boundary, and the energy termination position of each frequency band is aligned according to the shared respiratory termination boundary. The energy fluctuation sequence of each frequency band after alignment is jointly smoothed and amplitude consistent, and the frequency band energy sequence with consistent change direction and continuous fluctuation rhythm in the expansion phase and release phase is retained to generate a shared respiratory envelope trajectory.

[0082] Based on the rectified time-frequency structure, a first potential sound source trajectory and a second potential sound source trajectory are generated under the constraint of a shared respiratory envelope trajectory. The alternation sequence of the dual sound sources is generated according to the dominant frequency band position, energy occupancy order, and alternation switching position of the first and second potential sound source trajectories in each time segment. Specifically, the generation of the first and second potential sound source trajectories under the constraint of a shared respiratory envelope trajectory is as follows:

[0083] Using the time interval defined by the shared respiratory envelope trajectory as the candidate separation interval, the dominant energy distribution of each frequency band in the rectified time-frequency structure is read time by time within the candidate separation interval. According to the continuity of frequency band position, the stability of energy occupation and the consistency of frequency migration direction, dual-path tracking is performed on the dominant energy distribution at each time position to generate the first candidate sound source path and the second candidate sound source path. The duration, frequency band jump number and energy fall synchronization relationship of the first candidate sound source path and the second candidate sound source path in the shared respiratory expansion phase and the shared respiratory release phase are checked respectively. The paths that meet the conditions of duration, frequency band jump continuity and energy change synchronization with the shared respiratory envelope trajectory are retained and determined as the first potential sound source trajectory and the second potential sound source trajectory, respectively.

[0084] For the first potential sound source trajectory, the second potential sound source trajectory, the shared breathing envelope trajectory, and the alternating occupation order of the two sound sources among different original audio segments, the same source continuation check is performed. The original audio segments with continuous shared breathing boundaries, consistent alternating relationship between the two sources, and the ability to be connected sequentially across harmonic borrow redemption chains are retained. The retained original audio segments are then merged into the same set of candidate sound source segments.

[0085] In this embodiment, the step of constructing a tension debt backtracking chain for each candidate phonological fragment group and generating the phonological tension evolution state of each candidate phonological fragment group includes:

[0086] The candidate segments from the same sound source are sorted chronologically. Based on the shared respiratory envelope trajectory, the first potential sound source trajectory, the second potential sound source trajectory, and the alternating occupancy order of the two sound sources, clustering is performed according to temporal adjacency, respiratory boundary continuity, dominant frequency band migration, and alternating order of the two sound sources to generate multiple candidate sound phrase segments. Specifically, the clustering is performed based on temporal adjacency, respiratory boundary continuity, dominant frequency band migration, and alternating order of the two sound sources.

[0087] The starting and ending positions of each original audio segment are read in chronological order. Adjacent original audio segments with an interval not exceeding a preset segment continuation threshold are selected as candidate continuation segments. The continuation relationship between the shared respiratory convergence boundary of the previous original audio segment and the shared respiratory initiation boundary of the next original audio segment is checked for each candidate continuation segment. Original audio segment pairs with continuous or closable respiratory boundaries are retained. Based on the retained original audio segment pairs, the dominant frequency band migration direction, frequency band offset amplitude, and trajectory continuation position corresponding to the first and second potential sound source trajectories are read. Original audio segment pairs with consistent dominant frequency band migration directions and adjacent trajectory continuation positions are retained. The corresponding dual sound source alternation order is then read, and original audio segment pairs with consistent alternation order are retained. The original audio segments that meet the above conditions are merged in chronological order to generate multiple candidate sound phrase segment groups.

[0088] For each candidate phrase fragment group, the dominant frequency band bending pattern, harmonic tension pattern, envelope decompression pattern, and sideband recovery pattern are extracted sequentially along the time axis. The tension initiation leader segment, tension maintenance segment, tension unloading folding segment, and absorption bundle segment are identified, generating a tension segment sequence. Specifically, the tension initiation leader segment, tension maintenance segment, tension unloading folding segment, and absorption bundle segment are identified as follows:

[0089] The dominant frequency band center position, harmonic spacing change, envelope amplitude change, and sideband distribution range change corresponding to adjacent time positions in each candidate phonological fragment group are read sequentially over time. A continuous segment where the dominant frequency band continuously moves upward, the harmonic spacing continuously expands, the envelope amplitude continuously rises, and the sideband distribution range continuously expands is identified as the initial tensioning segment. A continuous segment where the dominant frequency band remains extended, the harmonic spacing remains stable, the envelope amplitude remains high, and the sideband distribution range remains expanded is identified as the tensioning maintenance segment. A continuous segment where the dominant frequency band reverses direction, the harmonic spacing changes from expansion to contraction, the envelope amplitude changes from high to low, and the sideband distribution range begins to shrink is identified as the tension unloading and reversal segment. A continuous segment where the dominant frequency band converges to a low amplitude position, the harmonic spacing continuously contracts, the envelope amplitude continuously falls, and the sideband distribution range is fully recovered is identified as the reabsorption segment.

[0090] For the interruption locations of adjacent original audio segments, the incomplete tension migration process is identified. Based on the tension continuity pattern before and after the interruption, the missing segment is divided into a pre-release debt segment, a post-absorption debt segment, and a borrowing-back debt segment. Specifically, the division of the missing segment into pre-release debt segment, post-absorption debt segment, and borrowing-back debt segment is as follows:

[0091] Read the dominant frequency band bending state, harmonic expansion state, envelope decompression state, and sideband recovery state corresponding to the end of the original audio segment before the interruption position; read the dominant frequency band bending state, harmonic expansion state, envelope decompression state, and sideband recovery state corresponding to the beginning of the original audio segment after the interruption position; when the original audio segment before the interruption position has entered the unwinding and folding stage and the envelope decompression process has not been completed, the missing segment is identified as the preceding decompression debt segment; when the original audio segment after the interruption position has entered the reabsorption beam stage and the preceding energy fallback process has not been closed, the missing segment is identified as the following reabsorption debt segment; when the dominant frequency bands before and after the interruption position are both in the folding migration process and the folding direction is consistent, but there is a missing dominant frequency band or a replacement frequency band in the middle, the missing segment is identified as the borrowed folding debt segment;

[0092] For each pre-release debt segment, post-absorption debt segment, and borrowing-back debt segment, cross-segment sequential connections are performed based on the consistency of the dominant frequency band back-back direction before and after the interruption position, the harmonic expansion and contraction continuity relationship, the envelope release amplitude correspondence, the side band recovery sequence relationship, and the alternating sequential continuation relationship of the dual sound sources. Back-back mappings are established for debt segments that can complete tension migration closure, generating tension debt back-back chains. Specifically, the back-back mapping for debt segments that can complete tension migration closure is as follows:

[0093] Using the interruption position corresponding to each debt segment as the mapping starting point, the dominant frequency band position, foldback direction, harmonic expansion and contraction state, envelope release amplitude, sideband recovery sequence, and dual-source alternation order corresponding to the end of the original audio segment before the interruption position and the beginning of the original audio segment after the interruption position are read; the debt segment with consistent foldback direction in the dominant frequency band, continuous harmonic expansion and contraction process, corresponding envelope release amplitude, connected sideband recovery sequence, and continuous dual-source alternation order is determined as the target debt segment; the front-end state corresponding to the target debt segment is used as the mapping starting state, and the back-end state corresponding to the target debt segment is used as the mapping ending state, and the corresponding foldback migration interval, state continuation direction, and continuation time range are recorded to establish the foldback mapping;

[0094] Based on the initiation leading segment, holding and maintenance segment, unloading and return segment, return absorption bundle segment and tension-debt return chain corresponding to each candidate call phrase segment group, the explicit tension segment and debt return segment are jointly arranged in chronological order to generate the call phrase tension evolution state of each candidate call phrase segment group.

[0095] In this embodiment, determining whether multiple candidate phrase fragments form a complete tension-reversal closed loop, and identifying the group of candidate phrase fragments that satisfy the condition as the phrase closed loop body, includes:

[0096] Read the tension evolution state of each candidate phrase fragment group, and arrange the initiation leading segment, holding segment, unloading return segment, return absorption bundle segment and tension debt return chain in chronological order to generate the candidate return sequence corresponding to each candidate phrase fragment group;

[0097] For each candidate turnaround sequence, the tension continuity relationship between the tension initiation leading segment and the tension holding segment, the pressure relief transition relationship between the tension holding segment and the tension release turnaround segment, and the recovery closure relationship between the tension release turnaround segment and the return absorption bundle segment are identified in chronological order. Insertion verification is performed on the tension debt turnaround chain in the candidate turnaround sequence to generate the corresponding tension migration chain.

[0098] For the interruption positions and cross-segment connection positions in each tension migration chain, the migration gaps that have not formed complete foldback closures are identified, and the corresponding dominant frequency band foldback direction, harmonic expansion and contraction sequence, envelope decompression and recovery sequence, shared breathing boundary continuation position, and dual sound source alternation sequence continuation position are extracted around each migration gap. Closed-loop filling verification is performed on each migration gap to generate the corresponding closed-loop verification results.

[0099] Based on the closed-loop verification results, a back-loop determination is performed on each tension migration chain. Tension migration chains in which the initiation pilot segment, tension holding segment, tension unloading back segment and back absorption bundle segment are closed in sequence and the tension debt back segment can complete the migration and replacement are retained. The retained tension migration chains are determined as the back-loop chain.

[0100] For each call phrase fold-back closed loop chain, the closed loop boundary is determined in chronological order. The corresponding starting position, holding and unfolding position, unloading and folding position, return to the absorption bundle position, and closed loop termination position are recorded. Candidate call phrase fragment groups with complete closed loop boundaries are determined as call phrase closed loop bodies.

[0101] In this embodiment, the construction of a transmodal autoisomorphic closure spectrum and the selection of stable avian voiceprints from the transmodal autoisomorphic closure spectrum include:

[0102] Read the shared respiratory envelope trajectory, first potential sound source trajectory, second potential sound source trajectory, alternating occupation order of dual sound sources and sound tension evolution state corresponding to multiple sound sentence closed loops. Extract the dominant frequency band migration process, duration expansion and contraction process, alternating switching process of dual sound sources and tension return closed loop process corresponding to each sound sentence closed loop in chronological order to generate the sound sentence structure evolution sequence corresponding to each sound sentence closed loop.

[0103] For each call structure evolution sequence, the initial call segment, the upward migration segment, the stable maintenance segment, the folding and terminating segment, and the termination call segment are identified respectively. The sequential connection relationship, cross-segment continuity relationship, dual call source switching position relationship and shared respiratory boundary correspondence between each segment are recorded to generate the structural skeleton sequence corresponding to each call closed loop.

[0104] Transmodulation alignment is performed on the structural skeleton sequences between different ringing phrases. A unified mapping is applied to overall pitch rise and pitch fall changes along the dominant frequency band migration direction, and a unified mapping is applied to overall acceleration and deceleration changes along the time axis. Based on the mapped correspondences, skeleton co-position segments, skeleton offset segments, and skeleton continuation segments between different ringing phrases are identified, generating candidate automorphic correspondences. Specifically, the generation of candidate automorphic correspondences is as follows:

[0105] After completing the transmodulation alignment process, the structural skeleton sequences corresponding to different sound phrase closed loops are read. The corresponding positional relationships between the starting sound segment, the upward migration segment, the stable maintenance segment, the foldback convergence segment, and the ending sound segment in each structural skeleton sequence are compared in chronological order. Segments with consistent segment order, consistent dominant frequency band migration topology, and corresponding dual sound source switching positions are identified as skeleton in-situ segments. Segments with consistent segment order but overall offset in frequency band position or duration are identified as skeleton offset segments. Segments that can form a sequential continuation between the ending sound segment of the previous sound phrase closed loop and the starting sound segment of the next sound phrase closed loop are identified as skeleton continuation segments. Sound phrase closed loops with corresponding skeleton in-situ segments, skeleton offset segments, and skeleton continuation segments are identified as candidate automorphic correspondences.

[0106] For each candidate automorphic correspondence, the consistency of the dominant frequency band migration order, the consistency of the alternation of dual sound sources, the consistency of the tension reversal loop order, and the consistency of the shared breathing boundary distribution between the corresponding sound phrase closed loops are checked. Candidate automorphic correspondences that satisfy the consistency relationship are retained. The retained candidate automorphic correspondences are written into the connection relationship between the corresponding sound phrase closed loops to construct the transmodulation automorphic sealing spectrum.

[0107] The closed loops of each call phrase in the transmodal autosomal closed spectrum are closed and screened. The call phrase closed loops that have bidirectional connection, closed loop return and continuous connection in the transmodal autosomal closed spectrum are retained and the retained call phrase closed loops are merged into stable bird voiceprints.

[0108] In this embodiment, the output of the corresponding bird voiceprint analysis results based on voiceprint embedding representation includes:

[0109] The shared respiratory envelope trajectory, first potential song source trajectory, second potential song source trajectory, alternating occupancy order of dual song sources, song tension evolution state, and song structure evolution sequence corresponding to each stable bird voiceprint are jointly encoded to generate corresponding voiceprint embedding representations. Each voiceprint embedding representation is matched with the template embedding representations in the preset bird voiceprint template library to generate species matching results. Based on the continuity of shared respiratory boundaries, the integrity of tension return loop closure, the self-isomorphic closure of transmodulation, and the template matching, the corresponding analysis confidence level is generated. Based on the matching results, the corresponding bird category results, category confidence results, homologous song merging results, and structural integrity results are output to form the corresponding bird voiceprint analysis results.

[0110] refer to Figure 2 An AI-based bird voiceprint analysis system includes:

[0111] The audio segment generation module is used to collect continuous environmental audio data and perform processing to generate original audio segments and multi-channel acoustic spectrum input tensors;

[0112] The dual-source encoding module is used to perform structural encoding on the multi-channel acoustic spectrum input tensor, generate the first potential source trajectory, the second potential source trajectory, the shared respiratory envelope trajectory, and the alternating occupancy order of the dual sources, and filter to obtain a set of candidate segments of the same sound source.

[0113] The Singing Tension Construction Module is used to cluster and arrange a set of candidate segments from the same vocal body and construct a tension debt loop chain to generate the singing tension evolution state.

[0114] The tension closed-loop determination module is used to perform succession matching based on the tension evolution state of the syllable and determine the syllable closed-loop body.

[0115] The closure spectrum construction module is used to extract the evolutionary sequence of song structure based on the song closure body and construct the transmodal autosomal closure spectrum to generate a stable bird voiceprint body.

[0116] The voiceprint analysis module is used to generate voiceprint embedding representations of stable bird voiceprints and output bird voiceprint analysis results.

[0117] Example 1: During a continuous bird acoustic monitoring cycle, the system received a batch of continuous environmental audio data collected by a fixed pickup terminal, with a total duration of 26 minutes and 14 seconds, a sampling rate of 24 kHz, and a quantization precision of 16 bits. This batch of audio included continuous wind noise, high-frequency insect calls, rustling sounds of branches and leaves, and alternating vocalizations from multiple birds at close range. After manual verification, 34 valid bird calls were identified in the audio, of which 11 were continuous calls from the target birds, with single-segment durations ranging from 1.8 seconds to 7.6 seconds. The average signal-to-noise ratio (SNR) of the original environmental audio was 8.4 dB, with a minimum SNR of 2.1 dB. Approximately 41% of the valid bird calls exhibited varying degrees of masking, interruption, or mixing.

[0118] After the audio enters the processing flow, the system first performs pre-emphasis, bandpass filtering, and amplitude normalization, with the bandpass range set from 1.2kHz to 8.6kHz. The average background noise energy before processing is 0.37, which is reduced to 0.24 after processing. Then, the system segments the audio into 25ms frames with a 10ms frame shift, and performs a short-time Fourier transform with a 512-point window. The system obtains 286 original audio segments from the continuous audio, with an average segment length of 0.73 seconds. Among these, there are 119 candidate bird song segments and 167 non-bird song segments. A multi-channel spectrogram input tensor is constructed for each original audio segment, where the log-Mel spectrogram, harmonic enhancement map, and envelope transformation map are all 128×172 in size. These three are superimposed by channel to form a 128×172×3 input tensor.

[0119] In the feature extraction stage, the system feeds the input tensor into the EfficientNet-B1 backbone network. After multi-scale convolutional responses, it outputs a basic acoustic signature feature map with a size of 16×22×320. For a target event lasting 2.94 seconds, the average response value of the dominant frequency band region in the feature map is 0.81, 0.26 for the background wind noise region, and 0.33 for the high-frequency band region of insect chirping. Subsequently, the system enters the cross-harmonic borrowing and redemption processing layer. Using the dominant harmonic cluster at each time position as a reference, the system searches for adjacent harmonic clusters and sideband accompanying clusters in the upper and lower adjacent harmonic orders. In this target event, the dominant harmonics that should have been continuous are interrupted from frame 37 to frame 42, resulting in a missing 6 frames. Based on the frequency migration direction before and after the interruption, the harmonic spacing preservation relationship, and the sideband envelope continuation relationship, the system determines 4 borrowed harmonic positions in the displaced frequency band and establishes 4 sets of redemption maps. The resulting transharmonic borrow-redemption chain has a length of 9 time nodes, which increases the dominant migration continuity before and after the interruption from 0.58 to 0.84.

[0120] During the dual-source breathing synchronization rectification stage, the system jointly processes the basic acoustic signature structure feature map and the redemption chain feature map to identify the shared breathing initiation boundary, expansion boundary, release boundary, and convergence boundary. In the aforementioned target event, a complete set of shared breathing boundaries was detected: the initiation boundary was located at frame 12, the expansion boundary at frame 29, the release boundary at frame 63, and the convergence boundary at frame 79. Based on the shared breathing boundaries, the system synchronously rectifies the energy fluctuations in different frequency bands to obtain the first and second potential sound source trajectories. In this event, the first potential sound source trajectory covers 53 time positions, and the second potential sound source trajectory covers 49 time positions, alternating 7 times, with the switching points concentrated near frames 18, 31, 45, and 67. After homogeneous continuation verification, 119 bird call candidate segments were merged into a set of 38 candidate segments from the same vocal body, among which the segments belonging to the continuous vocalization event of the target bird were merged from the original 19 segments into 8 groups.

[0121] After entering the resonance tension modeling stage, the system clusters and arranges the candidate segments of the same sound source in chronological order, resulting in 23 candidate resonance segment groups. Taking one target segment group as an example, this segment group consists of 3 original audio segments with a total duration of 4.12 seconds. The system extracts the dominant frequency band bending pattern, harmonic expansion pattern, envelope decompression pattern, and sideband recovery pattern in chronological order, identifying one initiation leading segment lasting 0.43 seconds; one holding segment lasting 1.12 seconds; one unloading and folding segment lasting 0.68 seconds; and one absorption segment lasting 0.39 seconds. There is a 0.21-second interruption between the second and third segments. The system identifies this interruption as corresponding to an incomplete tension migration process and divides it into one pre-decompression debt segment and one borrowed folding debt segment. Subsequently, based on the dominant frequency band folding direction before and after the interruption, the harmonic tension-contraction continuity relationship, the envelope decompression amplitude correspondence, and the alternating sequence continuity of the dual sound sources, cross-segment folding connections were completed, generating a tension debt folding chain with a length of 6 nodes, which improved the tension migration integrity of the segment group from 0.61 to 0.88.

[0122] In the tension return loop determination stage, the system performs sequential matching on each of the 23 candidate song phrase fragment groups to identify the migration relationships between the initial tension state, holding tension state, unloading tension state, and retraction state. The results show that 17 of the 23 candidate song phrase fragment groups meet the complete return loop closure conditions, while 6 are eliminated because they only have local frequency similarities and do not possess complete tension migration logic. Of the 8 candidate fragments for the target bird, 6 were identified as song phrase loop closures, and 2 were marked as low-confidence fragments. After manual review, 5 of the 6 song phrase loop closures were confirmed to be valid song phrases produced by the same bird during continuous vocalization, while 1 was an interference sound inserted by another bird at close range; this invention did not include it in the final stable results.

[0123] The system then extracted the song structure evolution sequences from 17 song closures and performed transmodulation automorphic relation matching. The system recorded the dominant frequency band migration process, duration scaling process, dual-source alternation switching process, and tension return process for each closure. Analysis of six target closures revealed that although the maximum absolute pitch shift reached 9.7% and the maximum duration deviation reached 13.4%, five of these closures maintained consistency in dominant frequency band migration topology, dual-source switching order, and tension return order. Based on this, the system constructed a transmodulation automorphic closure spectrum, obtaining 12 effective connecting edges and 3 closure loops. The closure loop corresponding to the target bird contained 5 closures, ultimately generating one stable bird voiceprint. The voiceprint embedding dimension of this stable bird voiceprint is 256-dimensional, with a cosine similarity of 0.91 with the target bird template in the template library, a similarity of 0.74 with the second-highest similarity template, a category confidence result of 0.89, and a structural integrity result of 0.86.

[0124] To verify the effectiveness of this invention, the system was trained using 9600 training samples, including 3200 complete bird calls, 2100 partially occluded bird calls, 1700 transposition and tempo shifting bird calls, 1400 mixed-call interference bird calls, and 1200 background noise artifacts. The test samples consisted of 1200 independent audio segments, including 700 valid bird calls and 500 interference bird calls. The method of this invention was compared with traditional Mel-spectrum mapping, ordinary convolutional classification, and temporal adjacency splicing methods on the same training and test sets. The test results show:

[0125] The traditional method achieves an accuracy rate of 81.6% in identifying bird call segments, while this invention achieves 91.8%.

[0126] The accuracy rate of grouping the same sound source using traditional methods is 72.4%, while that of this invention is 88.7%.

[0127] The continuous recovery rate of the syllables using traditional methods is 65.9%, while that of this invention is 85.6%.

[0128] The pseudo-splitting rate is 14.3% using traditional methods, while it is 4.8% using this invention.

[0129] The traditional method achieves a low signal-to-noise ratio sample recognition rate of 59.7%, while this invention achieves 80.9%.

[0130] In the 26 minutes and 14 seconds of live audio in this embodiment, the traditional method outputs 11 candidate target results, of which 6 are correct; the present invention outputs 7 candidate target results, of which 6 are correct, and the stable bird voiceprint is consistent with the results of manual verification.

[0131] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. An AI-based method for bird voiceprint analysis, characterized in that, include: Collect continuous environmental audio data within the target area, process the continuous environmental audio data to generate multiple raw audio segments, and construct a multi-channel acoustic spectrum input tensor for each raw audio segment; Each multi-channel acoustic spectrum input tensor is input into a dual-source structure encoding network consisting of an EfficientNet-B1 backbone, a transharmonic borrow redemption chain generation layer, and a dual-source breathing synchronization rectification layer. Local time-frequency texture convolution extraction, transharmonic borrow redemption mapping, and breathing synchronization energy rectification are performed sequentially to generate the first potential sound source trajectory, the second potential sound source trajectory, the shared breathing envelope trajectory, and the dual-source alternating occupancy order. A set of candidate segments of the same sound source is obtained by filtering. Cluster the candidate segments of the same vocal body in chronological order to generate multiple candidate vocal phrase segments. Construct a tension debt back-loop chain for each candidate vocal phrase segment group to generate the vocal phrase tension evolution state of each candidate vocal phrase segment group. The tension evolution state of the call phrase is sequentially matched according to time sequence to identify the migration relationship between the initial tension state, holding tension state, unloading tension state and retraction state. It is determined whether multiple candidate call phrase segments form a complete tension return loop, and the candidate call phrase segment group that meets the requirements is identified as the call phrase loop body. For multiple song closures, the corresponding song structure evolution sequences are extracted, and cross-modal automorphic relationship matching is performed between each song closure to construct a cross-modal automorphic closure spectrum. Stable bird voiceprints are obtained from the cross-modal automorphic closure spectrum. A voiceprint embedding representation is generated for a stable bird voiceprint body, and the corresponding bird voiceprint analysis results are output based on the voiceprint embedding representation.

2. The AI-based bird voiceprint analysis method according to claim 1, characterized in that, The continuous environmental audio data includes bird call components, wind noise components, insect chirping components, mixed calls of multiple birds in the same area components, and propagation, reflection, and masking sound components.

3. The AI-based bird voiceprint analysis method according to claim 1, characterized in that, The construction of multi-channel acoustic spectrum input tensors for each original audio segment includes... A short-time Fourier transform is performed on each original audio segment to obtain the corresponding time-frequency spectrum matrix. The time-frequency spectrum matrix is ​​mapped to a Mel filter bank, and the logarithm is taken to generate a log-Mel spectrum. Based on the fundamental frequency position, integer harmonic position, and continuous distribution relationship of adjacent harmonics of each frequency component in the time-frequency spectrum matrix, the harmonic components are enhanced to generate a harmonic enhancement map. The energy envelope is extracted based on the time-domain amplitude sequence of each original audio segment, and the envelope change between adjacent time points is calculated according to the time sequence to generate an envelope change map. The log-Mel spectrum, harmonic enhancement map, and envelope change map are overlaid according to the channel dimension to generate the corresponding multi-channel acoustic spectrum input tensor.

4. The AI-based bird voiceprint analysis method according to claim 1, characterized in that, The generation of the first potential sound source trajectory, the second potential sound source trajectory, the shared respiratory envelope trajectory, and the alternating occupancy order of the two sound sources, and the subsequent screening to obtain a set of candidate segments from the same sound-producing body, including: Each multi-channel acoustic spectrum input tensor is input into a dual-source structure coding network consisting of the EfficientNet-B1 backbone, the cross-harmonic borrow redemption processing layer, and the dual-source breathing synchronous rectification layer. The EfficientNet-B1 backbone extracts the local time-frequency texture features, harmonic continuous distribution features, envelope fluctuation coupling features, and cross-band cooperative variation features of the corresponding original audio segments to generate a basic acoustic text structure feature map. The basic acoustic signature feature map is input into the cross-harmonic borrowing and redemption processing layer. Based on the dominant harmonic cluster at each time position, the adjacent harmonic clusters and sideband accompanying clusters are searched along the adjacent harmonic orders. When the dominant harmonic cluster is interrupted, the borrowed harmonic position is determined in the ectopic frequency band according to the frequency migration direction before and after the interruption, the harmonic spacing relationship and the sideband envelope continuity relationship. A redemption mapping with the interrupted dominant harmonic position is established. The redemption mappings in continuous time are sequentially connected to generate a cross-harmonic borrowing and redemption chain feature map. The basic acoustic signature structure feature map and the transharmonic borrowing redemption chain feature map are input into the dual-source breathing synchronous rectification layer. The shared breathing initiation boundary, shared breathing expansion boundary, shared breathing release boundary and shared breathing convergence boundary are identified in chronological order. Based on the shared breathing boundary, synchronous rectification is performed on the energy fluctuation process in different frequency bands to generate the shared breathing envelope trajectory. Based on the rectified time-frequency structure, the first potential sound source trajectory and the second potential sound source trajectory are separated and generated under the constraint of the shared breathing envelope trajectory. The dual sound source alternation order is generated according to the dominant frequency band position, energy occupation order and alternation switching position of the first potential sound source trajectory and the second potential sound source trajectory in each time segment. For the first potential sound source trajectory, the second potential sound source trajectory, the shared breathing envelope trajectory, and the alternating occupation order of the two sound sources among different original audio segments, the same source continuation check is performed. The original audio segments with continuous shared breathing boundaries, consistent alternating relationship between the two sources, and the ability to be connected sequentially across harmonic borrow redemption chains are retained. The retained original audio segments are then merged into the same set of candidate sound source segments.

5. The AI-based bird voiceprint analysis method according to claim 1, characterized in that, The construction of tension debt backtracking chains for each candidate call sentence fragment group, generating the call sentence tension evolution state of each candidate call sentence fragment group, includes: The candidate segments of the same sound source are sorted according to time order. Based on the shared respiratory envelope trajectory, the first potential sound source trajectory, the second potential sound source trajectory, and the alternating occupation order of the two sound sources, clustering and arrangement are performed according to the temporal adjacency relationship, respiratory boundary continuity relationship, dominant frequency band migration relationship, and alternating order relationship of the two sound sources to generate multiple candidate sound sentence segment groups. For each candidate phonological segment group, the dominant frequency band bending pattern, harmonic tension pattern, envelope decompression pattern and side band recovery pattern are extracted in chronological order. The tension initiation leader segment, tension holding segment, tension unloading and folding segment and absorption bundle segment are identified to generate a tension segment sequence. For the interruption position of adjacent original audio segments, identify the tension migration process that has not been closed. Based on the tension continuity pattern before and after the interruption, divide the missing segment into the pre-release debt segment, the post-absorption debt segment, and the borrowing foldback debt segment. For each pre-released debt segment, post-absorption debt segment, and borrowing-back debt segment, cross-segment sequential connection is performed based on the consistent relationship of the dominant frequency band back direction before and after the interruption position, the relationship of harmonic expansion and contraction, the corresponding relationship of envelope release amplitude, the relationship of side band recovery sequence, and the relationship of alternating sequence of dual sound sources. Backward mapping is established for debt segments that can complete tension migration closure, generating tension debt backward chain. Based on the initiation leading segment, holding and maintenance segment, unloading and return segment, return absorption bundle segment and tension-debt return chain corresponding to each candidate call phrase segment group, the explicit tension segment and debt return segment are jointly arranged in chronological order to generate the call phrase tension evolution state of each candidate call phrase segment group.

6. The AI-based bird voiceprint analysis method according to claim 1, characterized in that, The step of determining whether multiple candidate phrase fragments form a complete tension-reversal closed loop, and identifying the group of candidate phrase fragments that satisfy the condition as the phrase closed loop body, includes: Read the tension evolution state of each candidate phrase fragment group, and arrange the initiation leading segment, holding segment, unloading return segment, return absorption bundle segment and tension debt return chain in chronological order to generate the candidate return sequence corresponding to each candidate phrase fragment group; For each candidate turnaround sequence, the tension continuity relationship between the tension initiation leading segment and the tension holding segment, the pressure relief transition relationship between the tension holding segment and the tension release turnaround segment, and the recovery closure relationship between the tension release turnaround segment and the return absorption bundle segment are identified in chronological order. Insertion verification is performed on the tension debt turnaround chain in the candidate turnaround sequence to generate the corresponding tension migration chain. For the interruption positions and cross-segment connection positions in each tension migration chain, the migration gaps that have not formed complete foldback closures are identified, and the corresponding dominant frequency band foldback direction, harmonic expansion and contraction sequence, envelope decompression and recovery sequence, shared breathing boundary continuation position, and dual sound source alternation sequence continuation position are extracted around each migration gap. Closed-loop filling verification is performed on each migration gap to generate the corresponding closed-loop verification results. Based on the closed-loop verification results, a back-loop determination is performed on each tension migration chain. Tension migration chains in which the initiation pilot segment, tension holding segment, tension unloading back segment and back absorption bundle segment are closed in sequence and the tension debt back segment can complete the migration and replacement are retained. The retained tension migration chains are determined as the back-loop chain. For each call phrase fold-back closed loop chain, the closed loop boundary is determined in chronological order. The corresponding starting position, holding and unfolding position, unloading and folding position, return to the absorption bundle position, and closed loop termination position are recorded. Candidate call phrase fragment groups with complete closed loop boundaries are determined as call phrase closed loop bodies.

7. The AI-based bird voiceprint analysis method according to claim 1, characterized in that, The construction of transmodal autoisomorphic closed spectra, and the selection of stable avian voiceprints from transmodal autoisomorphic closed spectra, includes: Read the shared respiratory envelope trajectory, first potential sound source trajectory, second potential sound source trajectory, alternating occupation order of dual sound sources and sound tension evolution state corresponding to multiple sound sentence closed loops. Extract the dominant frequency band migration process, duration expansion and contraction process, alternating switching process of dual sound sources and tension return closed loop process corresponding to each sound sentence closed loop in chronological order to generate the sound sentence structure evolution sequence corresponding to each sound sentence closed loop. For each call structure evolution sequence, the initial call segment, the upward migration segment, the stable maintenance segment, the folding and terminating segment, and the termination call segment are identified respectively. The sequential connection relationship, cross-segment continuity relationship, dual call source switching position relationship and shared respiratory boundary correspondence between each segment are recorded to generate the structural skeleton sequence corresponding to each call closed loop. Cross-modulation alignment processing is performed on the structural skeleton sequences between different ringing closed loops. The overall pitch rise and pitch fall changes are uniformly mapped along the dominant frequency band migration direction. The overall acceleration and deceleration changes are uniformly mapped along the time axis. Based on the mapped correspondence, the skeleton co-position segments, skeleton offset segments and skeleton continuation segments between different ringing closed loops are identified, and candidate automorphic correspondences are generated. For each candidate automorphic correspondence, the consistency of the dominant frequency band migration order, the consistency of the alternation of dual sound sources, the consistency of the tension reversal loop order, and the consistency of the shared breathing boundary distribution between the corresponding sound phrase closed loops are checked. Candidate automorphic correspondences that satisfy the consistency relationship are retained. The retained candidate automorphic correspondences are written into the connection relationship between the corresponding sound phrase closed loops to construct the transmodulation automorphic sealing spectrum. The closed loops of each call phrase in the transmodal autosomal closed spectrum are closed and screened. The call phrase closed loops that have bidirectional connection, closed loop return and continuous connection in the transmodal autosomal closed spectrum are retained and the retained call phrase closed loops are merged into stable bird voiceprints.

8. The AI-based bird voiceprint analysis method according to claim 1, characterized in that, The bird voiceprint analysis results output based on voiceprint embedding representation include: The shared respiratory envelope trajectory, first potential song source trajectory, second potential song source trajectory, alternating occupancy order of dual song sources, song tension evolution state, and song structure evolution sequence corresponding to each stable bird voiceprint are jointly encoded to generate corresponding voiceprint embedding representations. Each voiceprint embedding representation is matched with the template embedding representations in the preset bird voiceprint template library to generate species matching results. Based on the continuity of shared respiratory boundaries, the integrity of tension return loop closure, the self-isomorphic closure of transmodulation, and the template matching, the corresponding analysis confidence level is generated. Based on the matching results, the corresponding bird category results, category confidence results, homologous song merging results, and structural integrity results are output to form the corresponding bird voiceprint analysis results.

9. An AI-based bird voiceprint analysis system, executing the AI-based bird voiceprint analysis method according to any one of claims 1 to 8, characterized in that, include: The audio segment generation module is used to collect continuous environmental audio data and perform processing to generate original audio segments and multi-channel acoustic spectrum input tensors; The dual-source encoding module is used to perform structural encoding on the multi-channel acoustic spectrum input tensor, generate the first potential source trajectory, the second potential source trajectory, the shared respiratory envelope trajectory, and the alternating occupancy order of the dual sources, and filter to obtain a set of candidate segments of the same sound source. The Singing Tension Construction Module is used to cluster and arrange a set of candidate segments from the same vocal body and construct a tension debt loop chain to generate the singing tension evolution state. The tension closed-loop determination module is used to perform succession matching based on the tension evolution state of the syllable and determine the syllable closed-loop body. The closure spectrum construction module is used to extract the evolutionary sequence of song structure based on the song closure body and construct the transmodal autosomal closure spectrum to generate a stable bird voiceprint body. The voiceprint analysis module is used to generate voiceprint embedding representations of stable bird voiceprints and output bird voiceprint analysis results.

Images