@techreport{Du2023,
    Author = "Yan, Genwei and Wang, Ruoyu and Zou, Liang and Du, Jun and Wang, Qing and Gao, Tian and Fang, Xin",
    title = "MULTI-TASK FRAME LEVEL SYSTEM FOR FEW-SHOT BIOACOUSTIC EVENT DETECTION",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "June",
    abstract = "This technical report describes our new frame-level embedding learning system for DCASE2023 Task5: few-shot bioacoustic event detection. In the previous work, we proposed the frame-level embedding learning system and achieved the best performance of the DCASE 2022 Task5. In this work, we utilize several techniques to improve upon our previous work. Additionally, we introduce multi-task learning and Target Speaker Voice Activity Detection (TS-VAD) strategies to transform our previous system into a new multi-task frame-level embedding learning system. Compare to our previous work, our new system can achieve a better result (Fmeasure 75.74\%, No ML) on the official validation set"
}

@techreport{Gelderblom2023,
    Author = "Gelderblom, Femke and Cretois, Benjamin and Johnsen, Pal and Remonato, Filippo and Reinen, Tor Arne",
    title = "FEW-SHOT BIOACOUSTIC EVENT DETECTION USING BEATS",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "June",
    abstract = "Our method for the DCASE Challenge 2023 combines BEATs with Prototypical Networks. BEATs, standing for Bidirectional Encoder representation from Audio Transformers, is a newly-released architecture by Microsoft for audio tokenisation and classification. BEATs combines a tokenizer and a semi-supervised audio classifier which learn from each other to improve the classification of audio samples. Prototypical Networks, instead, can be briefly described as a neural network-based clustering algorithm. Somewhat resembling a K-means clustering, Prototypical Networks classify samples based on their distance from the classes’ prototypes (what would be the centroids in a K-means setting). Since the prototypes are constructed from a small set of examples from each class, called the support set, Prototypical Networks are well suited to handle fewshot learning settings like the DCASE Challenge. In our method, we combine the two by using BEATs as a feature extractor, constructing informative features which are used by the Prototypical Network to perform the prototypes’ construction and subsequent classification of test audio samples. We obtain a F1 score of 0.36 on the validation dataset."
}

@techreport{Jung2023,
    Author = "Lee, Yuna and Chung, HaeChun and Jung, JaeHoon",
    title = "FEW-SHOT BIOACOUSTIC DETECTION BOOSTING WITH FINE TUNING STRATEGY USING NEGATIVE BASED PROTOTYPICAL LEARNING",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "June",
    abstract = "Few-shot sound event detection has always faced the challenge of detecting bioacoustic sound events with only a few labelled instances of the class of interest. In this technical report, we describe our submission system for DCASE2023 Task5: few-shot bioacoustic event detection. We propose a novel framework of training audio segments via contrastive learning and prototypical learning, building the network more robust to the variety of acoustic environments, even in unseen domains. In addition, a finetuning strategy based on the novel loss functions is introduced. Our final systems achieves an f-measure of 83.08 on the DCASE task 5 validation set, outperforming the baseline performance and last year’s first place by a large margin."
}

@techreport{Moummad2023,
    Author = "Moummad, Ilyass and Serizel, Romain and Farrugia, Nicolas",
    title = "SUPERVISED CONTRASTIVE LEARNING FOR PRE-TRAINING BIOACOUSTIC FEW SHOT SYSTEMS",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "June",
    abstract = "We show in this work that learning a rich feature extractor from scratch using only official training data is feasible. We achieve this by learning representations using a supervised contrastive learning framework. We then transfer the learned feature extractor to the sets of validation and test for few-shot evaluation. For fewshot validation, we simply train a linear classifier on the negative and positive shots and obtain a F-score of 63.46\% outperforming the baseline by a large margin. We don’t use any external data or pretrained model. Our approach doesn’t require choosing a threshold for prediction or any post-processing technique"
}

@techreport{Wilkinghoff2023,
    Author = "Wilkinghoff, Kevin and Cornaggia-Urrigshardt, Alessia",
    title = "FEW-SHOT BIOACOUSTIC EVENT DETECTION",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "June",
    abstract = "This report describes the Fraunhofer FKIE submission for task 5 “Few-shot Bioacoustic Event Detection” of the DCASE challenge 2023. The submitted system is an adaptation of a few-shot keyword spotting system that uses embeddings with a temporal resolution suitable for template matching with dynamic time warping. The embedding model is trained to not only predict the sound event class but also the temporal position of a segment in a sound event using the angular margin loss TempAdaCos. At inference, embeddings are extracted and segment-wise cosine distances between the recording to be searched in and the provided templates are calculated. The resulting cost matrices are processed by applying a logistic regression model that is trained to discriminate between positive and negative frames. Lastly, dynamic time warping in combination with peak-picking and using a decision threshold is applied to detect on- and offsets of bioacoustic events. As a result, the presented system significantly outperforms both baseline systems."
}

@techreport{XuQianHu2023,
    Author = "Liu, Junyan and Zhou, Zikai and Sun, Mengkai and Xu, Kele and Qian, Kun and Hu, Bian",
    title = "SE-PROTONET: PROTOTYPICAL NETWORK WITH SQUEEZE-AND-EXCITATION BLOCKS FOR BIOACOUSTIC EVENT DETECTION",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "June",
    abstract = "In this technical reprot, we describe our submission system for DCASE2023 Task5: Few-shot Bioacoustic Event Detection. We propose a metric learning method to construct a novel prototypical network, based on adaptive segment-level learning and Squeezeand-Excitation (SE) blocks. We make better utilization of the negative data, which can be used to construct the loss function and provide much more semantic information. Most importantly, we propose to use SE blocks to adaptively recalibrate channel-wise feature response, by explicitly modeling interdependencies between channels, which improves f-measure to 63.94 \%. For the input feature, we use combination of per-channel energy normalization (PCEN) and delta mel-frequency cepstral coefficients (∆MFCC). Our system performs better than the baseline given by the officials, on the DCASE task 5 validation set. Our final score reaches an f-measure of 65.49 \%, outperforming the baseline performance by 30.18 \%."
}