@techreport{bhattacharjee2025Task1,
    Author = "Alongi, Gabriela and Ferguson, Liz and Sugarman, Peter",
    title = "BioDCASE Baleen Whale Deep Learning Detection and Classification Network",
    institution = "Ocean Science Analytics, Acoustic Interactions",
    year = "2025",
    month = "May",
    abstract = "DeepAcoustics is a deep learning-based tool designed for the detection and classification of underwater acoustic signals, with a focus on marine mammal vocalizations. Developed with a modular architecture, DeepAcoustics allows users to configure, train, and apply custom neural network models for sound event detection. DeepAcoustics supports a range of deep learning architectures for acoustic detection and classification, including Tiny YOLO, ResNet, and Darknet-based networks, allowing users to select the model best suited for their specific data and computational needs."
}

@techreport{biodcase2025Task1,
    Author = "Lastname, Firstname",
    title = "BioDCASE 2025 Task 1",
    institution = "BioDCASE 2025 Challenge",
    year = "2025",
    month = "May",
    abstract = "Coming soon!"
}

@techreport{harjuLandmark2025,
    Author = "Harju, Manu and Mesaros, Annamaria",
    title = "Landmark-based synchronization for Drifting Audio Recordings",
    institution = "BioDCASE 2025 Challenge",
    year = "2025",
    month = "May",
    abstract = "This technical report describes our submissions to the multichannel alignment task in the BioDCASE 2025 Challenge. Our system is based on matching and aligning audio landmarks, which are simple structures extracted from the spectrogram representations. Our code and the configuration used is available on GitHub."
}

@techreport{nihalBeats2025,
    Author = "Nihal, Ragib Amin and Yen, Benjamin and Ashizawa, Takeshi and Nakadai, Kazuhiro",
    title = "BEATs with Cross-Attention for Multi-Channel Audio Alignment",
    institution = "BioDCASE 2025 Challenge",
    year = "2025",
    month = "May",
    abstract = "We modified the provided BEATs baseline with three main changes. First, we added cross-attention layers that allow audio embeddings from different channels to interact before making alignment predictions. Second, improved the training process with better data sampling, conservative augmentation (amplitude scaling and noise addition), and AdamW optimization with learning rate scheduling. Third, replaced the baseline's binary counting similarity metric with confidence-weighted scoring that uses the full range of model outputs. The system uses the same candidate generation approach as the baseline but processes alignment decisions differently. On validation data, the method achieved MSE of 0.099 for ARU and 0.521 for zebra finch."
}