@techreport{Jin2023,
    Author = "Jin, Yongbin and Chen, Minjun and Shao, Jun and Liu, Yangyang and Peng, Bo and Chen, Jie",
    title = "{DCASE} 2023 Challenge Task4 Technical Report",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "May",
    abstract = "We describe our submitted systems for DCASE2023 Task4 in this technical report: Sound Event Detection with Weak Labels and Synthetic Soundscapes (Subtask A), and Sound Event Detection with Soft Labels (Subtask B). We focus on construct a CRNN model, which fuses the embedding extracted by the BEATs or AST pre-trained model，and use the frequency dynamic convolution(FDY-CRNN) and channel-wise selective kernel attention (SKA) for having adaptive receptive field. To get multiple models of different architectures for making an ensemble, we fine-tune multiple BEATs model on the SED dataset also. In order to make use of the weak labeled and unlabeled subset of DESED dataset further, we pseudo labels these subsets by a multiple iterative of self-training. We also use a small part of audio files from the Audioset dataset, and this part of data following the same self-training procedure. We train these models using two different settings, one setting for optimizing PSDS1 score, and the other for optimizing PSDS2 score. Our proposed systems achieve poly-phonic sound event detection scores (PSDS-scores) of 0.570 (PSDS-scenario1) and 0.889 (PSDS-scenario2) respectively on development dataset of subtask A, and macro-average F1 score with optimum threshold per class (F1MO) 49.70 on development dataset of subtask B."
}

@techreport{Liu2023,
    Author = "Liu, Shutao and Zhang, Peihong and Yang, Fulin and Zhu, Chenyang and Li, Shengchen and Shao, Xi",
    title = "Sound Event Detection System Using a Modifided Mvit for {DCASE} 2023 Challenge Task 4b",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "May",
    abstract = "In this report, we describe our submissions for the task 4b of Detection and Classification of Acoustic Scenes and Events (DCASE) 2023 Challenge: Sound Event Detection with Soft Labels. We use a MViT model based on frequency dynamic convolution. While preserving the advantages of multi-scale feature extraction by MViT, frequency dynamic convolution is used to overcome the translation invariance of image feature extraction caused by MViT model to improve the ability of the model in terms of extracting frequency dimension features. Without using any external datasets or pretrain model, our system trained only on the provided soft-label dataset, and the final F1-m score and F1-MO score are 80.52 and 63.43, respectively, both higher than the baseline system."
}

@techreport{Min2023,
    Author = "Min, Deokki and Nam, Hyeonuk and Yong-Hwa, Park",
    title = "Application of Spectro-Temporal Receptive Field for {DCASE} 2023 Challenge Task4 B",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "May",
    abstract = "Spectro-Temporal Receptive Field (STRF) is a linear function which describe the relationship between sound stimulus and primary auditory cortex (A1) neural response. By means of convolution with sound spectrogram and STRF, it is possible to predict the A1 cells response. A1 is known to estimates the spectro-temporal modulation information of input sound, and by considering the characteristic of A1, STRF is designed to capture this both spectral and temporal modulation information. In this work, we used STRF as a CNN kernel and construct the two-branch deep learning model. One branch is named as STRFNet whose first CNN layer kernel is STRF. This branch extracts the neuroscience-inspired spectro-temporal modulation information. The other branch is CRNN which has deeper layer than the baseline, and extracts the general time-frequency information of input spectrogram. Two-branch model is named as STRFaugNet, and its performance outperforms the baseline by 6.9\%."
}

@techreport{Nhan2023,
    Author = "Nhan, Tri-Do and Param, Biyani and Zhang, Yuxuan",
    title = "Sound Event Detection with Soft Labels Using Self-Attention Mechanisms for Global Scene Feature Extraction",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "May",
    abstract = "This paper presents our approach to Task 4b of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2023 Challenge, which focuses on Sound Event Detection with Soft Labels. Our proposed method builds upon a CRNN backbone model and leverages the benefits of data augmentation techniques to improve model robustness. Furthermore, we introduce self-attention mechanisms to capture global context information and enhance the model's ability to predict soft label segments more accurately. Our experiments demonstrate that incorporating soft labels and self-attention mechanisms result in significant performance gains compared to traditional methods on data varying across different scenarios"
}

@techreport{Xuenan2023,
    Author = "Xuenan, Xu and Ziyang, Ma and Fei, Yang and Guanrou, Yang and Mengyue, Wu and Xie, Chen",
    title = "Sound Event Detection by Aggregating Pre-Trained Embeddings From Different Layers",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "May",
    abstract = "This technical report is the system description of the X-Lance team submission to the DCASE 2023 task 4b challenge: sound event detection with soft labels. Our submissions focus on incorporating informative audio representations from self-supervised learning. The embeddings from different layers of the pre-trained models are aggregated as the input of our model. Since the occurrence of sound events in different scenes is imbalanced, for each scene we train our models using all the audio files. Finally, models of different architectures trained under different scenes are ensembled with learned weights."
}

@techreport{Yin2023,
    Author = "Yin, Han and Bai, Jisheng and Huang, Siwei and Chen, Jianfeng",
    title = "How Information on Soft Labels and Hard Labels Mutually Benefits Sound Event Detection Tasks",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "May",
    abstract = "This technical report describes our submission to DCASE 2023 Task 4 Subtrack B: Sound event detection (SED) with soft labels. We propose different architectures to explore how both soft and hard labels can jointly improve the performance of SED. And we use temporal mixup for data augmentation and k-fold cross-validationapproaches to solve the problem of too little training data. Our systems are built upon the Convolutional Recurrent Neural Network (CRNN) proposed by the baseline and the Conformer structure. We conduct extensive ablation experiments to compare the advantages and disadvantages of different information fusion strategies."
}

@techreport{Zhang2023,
    Author = "Zhang, Haiyue and Zuo, Liangxiao and Chen, Jingxuan and Cai, Xichang and Wu, Menglong",
    title = "Sound Event Detection Based on Soft Label",
    institution = "DCASE2023 Challenge",
    year = "2023",
    month = "May",
    abstract = "This report focuses on an in-depth study of DCASE 2023 Task 4b. In contrast to previous tasks, this task provides a dataset with soft labels, aiming to explore how to improve the performance of the baseline system using soft labels. The report primarily employs two effective enhancement methods. Firstly, to balance the dataset, the report expands the original dataset. Given the significant differences in sound events within the dataset, an augmentation method is employed to generate additional samples and equalize the dataset. Secondly, to further enhance the system performance, a model ensemble approach is utilized. By combining the predictions of multiple models, their individual strengths can be effectively utilized to improve overall performance."
}