The BirdCLEF2025 task aims to train a multi-label classifier to infer the presence probabilities of multiple species from the audio signals. In this work, we introduce a dual branch architecture to build an end-to-end passive acoustic monitoring predictor. Specifically, two diferent types of acoustic features (i.e., Mel-spectrogram and Mel-Frequency Cepstral Coeficients) are first extracted from the raw signals. ResNet and ConvNeXt are then used to learn two-branch features. Afterwards, the features are concatenated and fed into a fully connected layer to output the prediction probabilities. Finally, we conduct comparative experiments on the competition test data. Experimental results show that the proposed model achieves 0.751 and 0.771 macro-mean ROC-AUC on the 34% and 66% test dataset, respectively.
The identification of under-studied species via passive acoustic monitoring is less afected by weather
and more detectable in enhancing biodiversity monitoring[
Although previous methods have achieved promising results, they overlook the usage of multichannel features and the impact of pretrained weights on the feature mapping backbones. To this end, we in this work propose a dual-branch neural network that uses two types of acoustic features to better learn the latent meaningful features. The main contributions of this work are outlines are follows.
(1) A dual-branch neural network is proposed to build an end-to-end passive acoustic monitoring predictor to identify species. Two types of features, including Mel-spectrogram (Mel) and Mel-Frequency Cepstral Coeficients (MFCC) are extracted from raw audio signals, which are then fed into two typical pretrained feature representation networks (i.e., ResNet and ConvNeXt).
(2) We conduct comparative experiments to evaluate the efectiveness of the proposed model. Particularly, we evaluate diferent ways of initializing the parameters of ResNet and ConvNeXt. Results show that ResNet backbone with pretrained weights and ConvNeXt with random weights strategy outperforms others, with scores of 0.751 and 0.771, respectively (Team name: Hathaway Tan, Rank: 1455th).
The structure of this paper is as follows. Section 2 details the proposed model. Section 3 introduces the datasets, preprocessing steps and presents experimental results, followed by the conclusion section.
Considering the synthetic efect of diferent types of acoustic features in analyzing signals[
Recent studies indicate that pretrained models in the context of feature mapping have been widely used
because of its efectiveness in accelerating training[
The BirdCLEF2025 training audio dataset consists of a total of 28,564 audios, covering 206 unique species across four major taxonomic classes. Each audio recording has a duration ranging from 0.54s to 1774s. Table 1 presents the summary of dataset and Figure 2 displays the top 20 species by training audios.
To increase the number of samples for training, we use the sliding window without overlapping to segment the raw audio data into 10 seconds slices, where zero padding is applied to extend the data length if the original audio data is shorter than 10s. Finally, we get 78,579 segments in total. Figure 3 shows an example of data segment from an audio file in training dataset.
We then extract the MFCC and Mel features from each of the segments. Specifically, the MFCC features consist of the 13-dimensional MFCCs, first-order and second-order derivatives. For Melspectrogram features, we extract 128 Mel frequency bands per frame, resulting in a 128-dimensional feature vector. We set the FFT window size to 1024 and the hop length to 512. These parameters define the time-frequency resolution when computing the MFCC and Mel. We normalize the MFCCs and Mels sample by sample using Z-Score approach. The MFCCs and Mels are finally reshaped to 224 x 224 pixels.
For the training procedure, we utilize ResNet50.a1_in1k pretrained on ImageNet-1k dataset[10, 11], and use ConvNeXtv2_pico.fcmae[12]. The loss function is BCEWithLogitsLoss. The model totally runs 30 epochs with an early-stop strategy avoiding overfitting. We fine-tune the end-to-end model with the AdamW optimizer. An initial learning rate 0.001 is used and a cosine annealing learning rate scheduler is utilized, which adjusts the learning rate following a cosine curve from the initial value down to 1e-6.
As for the performance metric, a version of macro-averaged ROC-AUC that skips classes which have no true positive labels is used. We employ 3-fold stratified cross-validation for training. During training, the model achieving the highest average AUC on the validation set in each fold is saved.
During the test stage, prediction is conducted on the hidden test set, which transforms the model’s outputs to multi-class probabilities and calculates AUC scores using predictions and real multi-class labels. The submission format requires that the length of each test audio segment is 5 second, so we concatenate the 5 second segment with itself to create a 10 second slice. We opt not to train directly on 5-second segments due to the limited acoustic context they ofer, which can negatively impact classification performance.
Tables 2 and 3 present the results of diferent weight-using strategies on 34% and 66% test data respectively. In the table, “fine-tuned” means the feature mapping networks are equipped with pretrained weights; “from scratch” indicates that the parameters of feature mapping networks are randomly initialized.
The experimental results presented in Tables 2 and 3 show the AUC scores of various model combinations on two diferent test datasets (34% and 66%). Across both datasets, the model ResNet_ft(finetuned)+ConvNeXt_fs(from scratch) consistently achieves the highest AUC scores — 0.751 on the 34% test set and 0.771 on the 66% test set — indicating superior performance. In contrast, the model ResNet_fs(from scratch)+ConvNeXt_ft(fine-tuned) performs the worst. Interestingly, the combination ResNet_fs(from scratch)+ConvNeXt_ft(fine-tuned) performs worse than using ConvNeXt_fs(from scratch), which may imply that the feature in ConvNeXt has a less significant or even slightly detrimental efect when ResNet is not enhanced.
Our study reveals the following performance ordering: ResNet pretrained only > both pretrained > no pretraining > ConvNeXt pretrained only. These findings shows that visual pretrained models transfer well to spectrogram’s low-level texture and edge features but require close alignment between input representation and pretrained domain. Based on our findings, we recommend using pretrained ImageNet weights only when the input representation retains visual-like structures, such as Mel spectrograms, which benefit from learned low-level convolutional filters. On the contrary, for more abstract representations like MFCC, we advise against using pretrained visual weights.
In this work, we proposed a dual-branch architecture that leverages both Mel-spectrogram and MFCC features, processed through ResNet and ConvNeXt backbones, for passive acoustic monitoring in the BirdCLEF2025 task. Comparative experiments on the competition’s test datasets demonstrate the efectiveness of our design, with the model achieving macro-mean ROC-AUC scores of 0.751 on the 34% test set and 0.771 on the 66% test set. These results confirm that the proposed end-to-end framework can efectively capture complementary acoustic information and deliver robust multi-label bird species classification performance.
During the preparation of this work, the authors used OpenAI-GPT-4o in order to: Grammar and spelling check. After using this tool, the authors reviewed and edited the content as needed and take full responsibility for the publication’s content. [9] H. Liu, M. Long, J. Wang, M. I. Jordan, Towards understanding the transferability of deep representations, arXiv preprint arXiv:1909.12031 (2019). [10] R. Wightman, H. Touvron, H. Jégou, Resnet strikes back: An improved training procedure in timm, arXiv preprint arXiv:2110.00476 (2021). [11] K. He, X. Zhang, S. Ren, J. Sun, Deep residual learning for image recognition, in: Proceedings of the IEEE conference on computer vision and pattern recognition, 2016, pp. 770–778. [12] S. Woo, S. Debnath, R. Hu, X. Chen, Z. Liu, I. S. Kweon, S. Xie, Convnext v2: Co-designing and scaling convnets with masked autoencoders, in: Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 16133–16142.