@techreport{bordoux2024,
    Author = "Bordoux, Valentin",
    title = "ADAPTABLE INPUT LENGTH USING MODEL TRAINED ON WAVEFORM Technical Report",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "This report presents a method for bioacoustic sound event detec- tion using few-shot learning, developed for the DCASE 2024 Task 5. Our approach experiments with pretrained models that take waveforms as input. These models serve as feature extrac- tors, and prototypical loss is used for prediction. Initially, we employed direct predictions with openly available pretrained models. Subsequently, we attempted to fine-tune the models for each file, using only the first five annotations as training set. The direct prediction system achieved 40\% F-measure score, 12 points under the baseline system proposed by the organizers. Fine-tuning did not improve the model's performance over direct prediction.While our proposed method can be applied directly without extensive parameter tuning or additional training, the results indicate that it does not achieve the generalizability required for this challenge when compared to the baseline method. This work suggests how state-of-the-art models, despite their high perfor- mance on other datasets or benchmarks, may still perform subop- timal on sound event detection using few-shot learning for cer- tain taxonomic groups present in the DCASE challenge datasets."
}

@techreport{hoffman2024,
    Author = "Hoffman, Benjamin and Robinson, David",
    title = "Toward in-context bioacoustic sound event detection",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "We introduce an in-context learning approach to bioacoustic sound event detection. Our approach consists of a large pre-trained transformer model which, when prompted with a small amount of labeled audio, directly predicts de- tection labels on unlabeled audio. To train our model, we constructed a large audio database, which we used to generate acoustic scenes with temporally fine-grained detection labels. On the validation set for the 2024 DCASE Few-shot bioacous- tic event detection challenge, our best-performing submission achieves an average F1 score of 0.584, improving on the challenge baseline by 0.063."
}

@techreport{kao2024,
    Author = "Kao, Sheng-Lun and Liu, Yi-Wen",
    title = "Cosine similarity based Few-shot Bioacoustic Event Detection with Automatic Frequency Range Identification in Mel-Spectrograms Technical Report",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "In response to the Few-shot Bioacoustic Event Detection chal- lenge, we have developed a detection system comprising three key components. First, an algorithm has been devised for auto- matically identifying frequency ranges of the positive (POS) sig- nal within the mel-spectrogram. Secondly, the cosine similarity between POS and negative (NEG) events is computed across the entire audio file. Thirdly, predictions of POS events are made based on the results of cosine similarity. Remarkably, this ap- proach does not rely on any training data from the development dataset, external data, or pretrained models. The proposed system achieved an F1-score of 44.187\% on the 2023 validation set."
}

@techreport{latifi2024,
    Author = "Bidarouni, Amir Latifi and Abeßer, Jakob",
    title = "CROSS-ADAPT: CROSS-DATASET GENERATION AND DOMAIN ADAPTATION TECHNIQUE FOR FEW-SHOT LEARNING Technical Report",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "Bioacoustic monitoring is an invaluable tool for understanding wildlife well-being. However, the scarcity of annotated data for effective model training coupled with domain shifts resulting from data recorded at various sensor locations with diverse acoustic en- vironments poses significant challenges for deep learning-based audio classification systems. In this paper, we propose a novel cross-dataset data augmentation technique designed to effectively use the limited annotated data available, exemplified by the few- shot learning task 5 of the DCASE challenge. Furthermore, we employ Instance-wise Feature Projection-based Domain Adapta- tion (IFPDA) to mitigate the domain shifts caused by variations in recording locations or devices. We use a modified ResNet model architecture for a multitask learning setting, which combines multi- class species classification on a patch level and binary classification for frame-level sound event detection."
}

@techreport{liu2024,
    Author = "Liu, Wei and Liu, Hy",
    title = "FEW-SHOT BIOACOUSTIC EVENT DETECTION AT THE DCASE 2024 CHALLENGE Technical Report",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "In this technical report, we describe the submission system for DCASE2024 Task 5: Few-shot Bioacoustic Event Detection. In previous work, we proposed a frame-level embedding learning system and achieved the best performance in DCASE2022 Task 5. In this task, we propose several methods to improve the representational capacity of embeddings under limited positive samples. Three methods are proposed based on the pre-training fine-tuning process, including the AAPM segment-level embedding learning method, the Baseline framework-level embedding learning method, and the Unet network-based framework-level embedding learning method. Compared to our previous work, our new system achieved better results on the official 2023 validation set (F-measure 76.8\%, No ML). The proposed system was evaluated on the newly released official 2024 validation set, with a best overall F-measure score of 70.56\%."
}

@techreport{lu2024,
    Author = "Zhao, Peng Yuan and Lu, Cheng Wei and Zou, Liang",
    title = "FEW-SHOT BIOACOUSTIC EVENT DETECTION WITH FRAME-LEVEL EMBEDDING LEARNING SYSTEM Technical Report",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "In this technical report, we describe our submission system for the few-shot bioacoustic event detection in the DCASE2022 task5. Participants are expected to develop a few-shot learning system for detecting mammal and birds sounds from audio recordings. In our system, Prototypical Networks are used to embed spectrograms into an embedding space and learn a non-linear mapping between data samples. We leverage various data augmentation techniques on Mel-spectrograms and introduce a ResNet variant as the classifier. Our experiments demonstrate that the system can achieve the F1-score of 47.88\\% on the vali-dation data."
}

@techreport{qianhu2024,
    Author = "Sun, Mengkai and Zhang, Haojie and Qian, Kun and Hu1, Bin",
    title = "LIF-PROTONET: PROTOTYPICAL NETWORK WITH LEAKY INTEGRATE-AND-FIRE NEURON AND SQUEEZE-AND-EXCITATION BLOCKS FOR BIOACOUSTIC EVENT DETECTION Technical report",
    institution = "DCASE2024 Challenge",
    year = "2024",
    month = "June",
    abstract = "In this technical repot, we describe our submission system for DCASE2024 Task5: Few-shot Bioacoustic Event Detection. We propose a metric learning method to construct a novel prototypical network, based on Leaky Integrate-and-Fire Neuron and Squeeze- and-Excitation (SE) blocks. We make better utilization of the nega- tive data, which can be used to construct the loss function and pro- vide much more semantic information. Most importantly, we pro- pose to use SE blocks to adaptively recalibrate channel-wise feature response, by explicitly modeling interdependencies between chan- nels, which improves f-measure to 53.72 \%. For the input feature, we use combination of per-channel energy normalization (PCEN) and delta mel-frequency cepstral coefficients (∆MFCC), then the features were first transformed through Leaky Integrate-and-Fire Neuron to mimic brain function. Our system performs better than the baseline given by the officials, on the DCASE2024 task 5 vali- dation set. Our final score reaches an f-measure of 55.49 \%, outper- forming the baseline performance."
}