@techreport{Anastcio2022,
    Author = "Anastácio, Ricardo and Ferreira, Luís and Mónica, Figueiredo and Luís, Conde Bento",
    title = "Ai4edgept Submission to {DCASE} 2022 Low Complexity Acoustic Scene Classification Task1",
    institution = "DCASE2022 Challenge",
    year = "2022",
    month = "June",
    abstract = "This report details the submission to task1 of DCASE2022 competition. The task aims to classify acoustic scenes using devices with low computational power and memory. We propose two ensemble models for scene classification. The first model clusters classes into 2 groups, each of a two-network ensemble being responsible for intra-group discrimination, i.e. discriminating between the classes that are most related in the confusion matrix. The second model implements a canonical one-versus-all ten-network ensemble architecture followed by knowledge distillation, i.e. the ensemble model is used as the teacher network. The student is an optimised version of the DCASE2022 baseline architecture. In both models we resort to three different data pre-processing techniques: audio downsample; mel-spectrogram tuning; and data augmentation. We’ve used the DCASE2022 baseline for all networks - two-network ensemble, ten-network ensemble and student network - on which we have conducted an architecture’s hyperparameter search to identify the best performing architecture, while being compliant with DCASE2022 performance metrics. Results revealed that data pre-processing and knowledge distillation techniques improve overall performance. Nevertheless, a simple two-network ensemble without knowledge distillation, maintains the MACS and parameters size low, while achieving similar results."
}

@techreport{Cai2022,
    Author = "Cai, Yiqiang and Tang, He and Zhu, Chenyang and Li, Shengchen and Shao, Xi",
    title = "Low-Complexity Model Based on Depthwise Separable {CNN} for Acoustic Scene Classification",
    institution = "DCASE2022 Challenge",
    year = "2022",
    month = "June",
    abstract = "The task1 of DCASE 2022 put forward higher requirements for system complexity and the new datasets also brought greater challenges. We tried to reproduce several models in previous years, but did not get a good performance. Therefore, we introduced the depthwise separable CNN method to the baseline architecture, which successfully reduces the complexity and improves the accuracy. We also used three methods of data augmentation, mixup, pitch shifting and stretching to further improve the results."
}

@techreport{Cao2022,
    Author = "Cao, Wenchang and Li, Yanxiong and Huang, Qisheng and Liu, Mingle",
    title = "Low-Complexity Acoustic Scene Classification Using Broadcasted {ResNet} and Data Augmentation",
    institution = "DCASE2022 Challenge",
    year = "2022",
    month = "June",
    abstract = "Acoustic scene classification (ASC) is a task to classify each input audio recording into one class of pre-given acoustic scenes. As an important task in Detection and Classification of Acoustic Scenes and Events (DCASE), ASC has attracted a lot of attention from researchers in the community of audio and acoustic signal processing in recent years [1]-[4]. In the work of this report, we focus on the task of low-complexity ASC with multiple devices, namely, Task 1 of the DCASE2022 challenge [5]. In this task, a low-complexity model is required to classify audio recordings recorded by multiple devices (real and simulated). In the proposed ASC method, the BC-ResNet-Mod [6] is used as the backbone of our model whose training strategy is the Cross-Gradient Training (CGT) [7]. In addition, some data augmentation techniques are adopted for further improving the performance of the proposed method. The size of our model is 125.33 KB after model compression, which is lower than the size limit of 128 KB. Evaluated on the development dataset, our system obtains classification accuracy of 51.1\%."
}

@techreport{Dong2022,
    Author = "Dong, Hongxia and Zhang, Lin and Cai, Xichang and Wu, Menglong and Qiao, Ziling and Gan, Yanggang and Wu, Juan",
    title = "Acoustic Scene Classification Based on Fhr\_mobilenet",
    institution = "DCASE2022 Challenge",
    year = "2022",
    month = "June",
    abstract = "This technical report describes our submission for Task1 of DCASE2022 challenge. We calculated 128 log-mel energies under the original sampling rate of 44.1KHz for each time slice by taking 2048 FFT points with 50\% overlap. Additionally, deltas and deltadeltas were calculated from the log Mel spectrogram and stacked into the channel axis. The resulting spectrograms were of size 128 frequency bins, 43 time samples and 3 channels with each representing log-mel spectrograms, its delta features and its delta-delta features respectively. Then, the three channel feature map is fed into the mobilenet-based frequency high-resolution network. Finally, after 1 × 1 convolution and global average pooling, the classification results are obtained through softmax output. The classification accuracy of our proposed model is 53.9\% with a loss value of 1.378. The number of parameters of the model is 70.608K, where each parameter is represented using int8 and the MACs are 28.461M."
}

@techreport{Hou2022,
    Author = "Hou, YuanBo",
    title = "Low-Complexity for {DCASE} 2022 Task {1A} Challenge",
    institution = "DCASE2022 Challenge",
    year = "2022",
    month = "June",
    abstract = "This technical report describes the systems for task1/subtask A of the DCASE 2022 challenge. In order to reduce the number of model parameters and improve accuracy. In this work, I use a simple neural networks with causal convolution and bottleneck structure. The log-mel spectrograms are extracted to train the acoustic scene classification model. The mix-up and Specaugmentation are used to augment the acoustic features. My system achieves higher classification accuracies and lower log loss in the development dataset than baseline system."
}

@techreport{Kim2022,
    Author = "Kim, TaeSoo and Lee, GaHui and Park, JaeHan",
    title = "Kt Submission for the {DCASE} 2022 Challenge: Modernized Convolutional Neural Networks for Acoustic Scene Classification",
    institution = "DCASE2022 Challenge",
    year = "2022",
    month = "June",
    abstract = "In this technical reports, we present our team’s submission for DCASE 2022 TASK1 which is the low complexity Acoustic Scene Classification (ASC). We gradually modernized a neural network architecture design starting from the baseline model and discover several key components that contribute to the performance. To meet constraints of the model complexity, the number of parameters and the number of MACs are considered while applying each designs. As a result, our model achieves 1.2593 log-loss and 54.03\% accuracy on the development set, while having less than 114k of total parameters (including the zero-valued) and 30 million MACs."
}

@techreport{Lee2022,
    Author = "Lee, Joo-Hyun and Choi, Jeong-Hwan and Byun, Pil Moo and Chang, Joon-Hyuk",
    title = "Hyu Submission for the {DCASE} 2022: Efficient Fine-Tuning Method Using Device-Aware Data-Random-Drop for Device-Imbalanced Acoustic Scene Classification",
    institution = "DCASE2022 Challenge",
    year = "2022",
    month = "June",
    abstract = "This paper address the Hanyang University team submission for the DCASE 2022 Challenge Low-Complexity Acoustic Scene Classification task. The task aims to design a generalized audio scene classification system for various devices under low complexity and short input time conditions. We followed two strategies to achieve our goal: improving the model structure for short segmented audio and adopting transfer learning methods that are generalizable to unknown devices. Based on the BC-ResNet, which showed the best performance in DCASE 2021 challenge, we incorporated the method proposed in the field of short-duration speaker verification to secure high accuracy. In addition, we proposed a novel finetuning method using device-aware data-random-drop to get a generalized model across multiple devices. Most of the training dataset is data recorded with a specific device. We devised a fine tuning method that gradually excludes data recorded with a specific device from mini-batch during training, and this method improves generalization performance. Following the official protocol of cross validation setup from the TAU Urban Acoustic Scenes 2022 Mobile development dataset, we achieve 70.1\% accuracy and 0.835 multi class cross-entropy loss, respectively."
}

@techreport{Liang2022,
    Author = "Liang, Jiangnan and Zeng, Cheng and Shi, Chuang and Zhang, Le and Zhou, Yisen and Li, Yuehong and Zhou, Yanyu and Tan, Tianqi",
    title = "Low-Complexity Acoustic Scene Classification Based on Residual Net",
    institution = "DCASE2022 Challenge",
    year = "2022",
    month = "June",
    abstract = "This technical report describes the submitted systems for task 1 of the DCASE 2022 challenge. The log-mel energies, delta features and delta-delta features were extracted to train the model. We adopted a total of eight data augmentation methods. BC-ResNet and MobileNetV2 were used as training model. We used knowledge distillation and quantization to compress the model. Our systems achieved lower log loss and higher accuracy in the development dataset than the baseline system."
}

@techreport{Morocutti2022,
    Author = "Morocutti, Tobias and Shalaby, Diaaeldin",
    title = "Receptive Field Regularized {CNNs} with Traditional Audio Augmentations",
    institution = "DCASE2022 Challenge",
    year = "2022",
    month = "June",
    abstract = "This technical report describes our system for Task 1 (Low-Complexity Acoustic Scene Classification) of the DCASE2022 Challenge. Due to the limited allowed complexity of the model to submit, we use a teacher-student approach. The teacher is a Receptive Field (RF) regularized CNN model and the student is a simpler 5-layer CNN with batch normalization, dropout and maxpool layers. In addition, some data augmentation techniques, such as adding gaussian noise, shifting, pitch shifting and time stretching are adopted for expanding the diversity of the dataset. Our system achieves an accuracy of 53.4\% and a multiclass cross-entropy (log loss) of 1.279 on the development dataset. The student model has 21,930 parameters and a Multiply accumulate count of 9.775 million."
}

@techreport{Olisaemeka2022,
    Author = "Olisaemeka, Chukwuebuka and Babu Saheer, Lakshmi",
    title = "Submission to {DCASE} 2022 Task 1: Depthwise Separable Convolutions for Low-Complexity Acoustic Scene Classification",
    institution = "DCASE2022 Challenge",
    year = "2022",
    month = "June",
    abstract = "This technical report describes the details of the TASK1 submission to the DCASE2022 challenge. The aim of this task is to design an acoustic scene classification system that targets devices with low memory and computational allowance. The task also aims to build systems that can generalize across multiple devices. To achieve this objective, a model using Depthwise Separable Convolutional layers is proposed, which reduces the number of parameters and computations required compared to the normal convolutional layers. This work further proposes the use of dilated kernels, which increase the receptive field of the convolutional layer without increasing the number of parameters to be learned. Finally, quantization is applied to reduce the model complexity. The proposed system achieves an average test accuracy of 39\% and log loss of 1.878 on TAU Urban Acoustic Scenes 2022 Mobile, development dataset with a parameter count of 96.473k and 3.284 MMACs."
}