Semi-Supervised NMF-CNN For Sound Event Detection

Chan, Teck Kai^1,2 and Chin, Cheng Siong¹ and Li, Ye^2
1Newcastle University Singapore, Faculty of Science, Agriculture, and Engineering, Singapore, ²Xylem Water Solution Singapore Pte Ltd, 3A International Business Park,Singapore

Abstract

For the DCASE 2020 Challenge Task 4, this paper proposed a combinative approach using Nonnegative Matrix Factorization (NMF) and Convolutional Neural Network (CNN). The main idea begins with utilizing NMF to approximate strong labels for the weakly labeled data. Sub-sequently, based on the approximated strongly labeled data, two different CNNs are trained using a semi-supervised framework where one CNN is used for clip-level prediction and the other for frame-level prediction. Using this idea, the best model trained can achieve an event-based F1-score of 45.7% on the validation dataset. Using an ensemble of models, the event-based F1-score can be increased to 48.6%. By comparing with the baseline model, the proposed model outperforms the baseline model by a margin of over 8%.

System characteristics

Combined Sound Event Detection And Sound Event Separation Networks For DCASE 2020 Task 4

Chen, You-Siang¹ and Lin, Zi Jie¹ and Li, Shang-En² and Koh, Chih-Yuan¹ and Bai, Mingsian R.¹ and Chien, Jen-Tzung² and liu, Yi-Wen^1
1National Tsing Hua University, Hsinchu, ²National Chiao Tung University, Hsinchu, Taiwan

Abstract

In this paper, we propose a hybrid neural network (NN) to handle the tasks of sound event separation (SES) and sound event detection (SED) in Task 4 of DCASE 2020 challenge. The convolutional time-domain audio separation network (Conv-TasNet) is employed to extract the foreground sound events defined in DCASE challenge. By comparing the baseline SED network with various training strategies, we demonstrate that the SES network is capable of enhancing the SED performance effectively in terms of several event-based performance metrics including macro F1 and poly-phonic sound detection score (PSDS).

System characteristics

Sound Event Detection And Classification Using CWT Scalograms And Deep Learning

Copiaco, Abigail¹ and Ritz, Christian² and Fasciani, Stefano³ and Abdulaziz, Nidhal^4
1Engineering and Information Sciences Dept., University of Wollongong, Northfields Wollongong, Australia, ²School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, Northfields Wollongong, Australia, ³, Musicology Dept., University of Oslo, Norway, ⁴University of Wollongong in Dubai, Dubai

Abstract

This report describes our proposed system for the DCASE 2020 Task 4 challenge. In this work, we examine the combination of signal energy and spectral centroid features with 0.05 s of time windowing for the detection of sound events. Along with this, spectro-temporal features extracted from Fast Fourier Transform (FFT) based wavelet coefficients of the audio files were used for classification. These coefficients are mapped into images called scalograms, which are fed into the layers of AlexNet, a pre-trained Deep Convolutional Neural Network (DCNN), for transfer learning. Through the validation set, this method gathered an average F1-score of 74% amongst the 10 classes of the DESED database for weak labelling. However, this technique is not deemed to be suitable for classification with strong time stamps labelling, gathering an F1-score of a mere 11.21%

System characteristics

Detecting And Classifying Separated Sound Events Using Wavelet_Based Scalograms And Deep Learning

Copiaco, Abigail¹ and Ritz, Christian² and Fasciani, Stefano³ and Abdulaziz, Nidhal^4
1Engineering and Information Sciences Dept., University of Wollongong, Northfields Wollongong, Australia, ²School of Electrical, Computer and Telecommunications Engineering, University of Wollongong, Northfields Wollongong, Australia, ³, Musicology Dept., University of Oslo, Norway, ⁴University of Wollongong in Dubai, Dubai

Abstract

This report describes our proposed system submitted to the DCASE 2020 Task 4 challenge that includes sound separation and detection. In this work, we examine the use of a Deep Neural Network (DNN) based sound separation system as a preprocessing technique to the sound event classification technique used. For the detection of sound events, a combination of signal energy and spectral centroid features with 0.05-s of time windowing was utilized. Along with this, spectro-temporal features extracted from a Fast Fourier Transform (FFT) based wavelet coefficients of the audio files were used for classification. These coefficients are mapped into images called scalograms, which are fed into the layers of AlexNet, a pre-trained Deep Convolutional Neural Network (DCNN), for transfer learning. Through the validation set, this method gathered an average F1-score of 70% amongst the 10 classes of the DESED database for weak labelling However, this technique is not deemed to be suitable for classification with strong time stamps labelling, gathering an F1-score of a mere 8.73%.

System characteristics

The UNIVPM-INRIA Systems For The DCASE 2020 Task 4

Cornell, Samuele Cornell¹ and Pepe, Giovanni¹ and Principi, Emanuele¹ and Pariente, Manuel² and Olvera, Michel² and Gabrielli, Leonardo¹ and Squartini, Stefano^1
1Dept. Information Engineering, Università Politecnica delle Marche, Ancona, Italy, ²Dept. Information and Communication Sciences and Technologies, INRIA Nancy Grand-Est, France

Abstract

In this technical report, we propose different Sound Event Detection (SED) systems for the 2020 DCASE Task 4 challenge. Given the mismatch between synthetic labelled data and target domain data, we exploit a domain adversarial training to improve the network invariance to different types of background noise. Furthermore, we use dynamic mixing and augmentation of synthetic examples at training time as well as prediction smoothing by using Hidden Markov Models. In one system, we also show that using a learnable dynamic compression, Per-Channel Energy Normalization (PCEN) front-end improves robustness to background noise by making it Gaussian. Finally, an ensemble of models proves beneficial to improve the prediction score. Concerning joint separation and sound event detection we propose a permutation-invariant training scheme to optimize directly the Sound-Event-Detection objective.

System characteristics

Multi-Resolution Mean Teacher For DCASE 2020 Task 4

de Benito-Gorron, Diego and Segovia, Sergio and Ramos, Daniel and Toledano, Doroteo T.
Universidad Autónoma de Madrid, Madrid, Spain

Abstract

In this technical report, we describe our participation in DCASE 2020 Task 4: Sound event detection and separation in domestic environments. A multi-resolution feature extraction approach is proposed, aiming to take advantage of the different lengths and spectral characteristics of each target category. The combination of up to five different time-frequency resolutions via model fusion is able to outperform the baseline results. In addition, we propose class-specific thresholds for the F 1 -score metric, further improving the results over the Validation set.

System characteristics

Convolutional Recurrent Neural Networks For Weakly Labeled Semi-Supervised Sound Event Detection In Domestic Environments

Ebbers, Janek and Haeb-Umbach, Reinhold
Paderborn University, Paderborn, Germany

Abstract

In this report we present our system for the Detection and Classification of acoustic scenes and events (DCASE) 2020 Challenge Task 4: Sound event detection in domestic environments. We present a convolutional recurrent neural network (CRNN) with two recurrent neural network (RNN) classifiers sharing the same preprocessing convolutional neural network (CNN). Both recurrent networks perform audio tagging. One is processing the input audio signal in forward direction and the other in backward direction. The networks are trained jointly using weakly labeled data, such that at each time step an active event is tagged by at least one of the networks given that the event is either in the past captured by the forward RNN or in the future captured by the backward RNN. This way the models are encouraged to output tags as soon as possible. After training, the networks can be used for sound event detection by applying them to smaller audio segments, e.g. 200 ms. Further we propose a tag conditioned CNN as a second stage which is supposed to refine sound event detection. Given its receptive field and file tags as input it performs strong label prediction trained using pseudo strong labels provided by the CRNN system. By ensembling four CRNNs and four CNNs we obtain an event-based F-score of 48.3% on the validation set, which is 13.5% higher than the baseline. We are going to release the source code at https://github.com/fgnt/pb_sed.

System characteristics

Cross-domain sound event detection: from synthesized audio to real audio

Hao, Junyong and Hou, Zhenwei and Peng, Wang
Chongqing University, Chongqing, China

Abstract

This technical report describes some of the system information submitted to DCASE2020 task4 - Sound Event Detection in Domestic Environments. We use the dataset of DCASE2019 task4 to train our model, contains strongly labeled synthetic data, large unlabeled data, and weakly labeled data. There is a very large domain gaps in the statistical distribution between the synthesized audio and the real audio, and the performance of the SED model obtained on the synthesized audio applied to the real audio is greatly reduced. To perform this task, we propose a DA-CRNN network for joint learning of sound event detection (SED) and domain adaptation (DA).We consider the impact of the distribution within a single sound on the generalization perfor- mance of the model by mitigating the impact of complex background noise on event detection and the self-correlation consistency regularization of clip-level sound event classification, these make the intra-domain of a single sound smoother ; for cross-domain adaptation, adversarial learning through feature extraction network with frame-level domain discriminator and clip-level domain discriminator, forcing the feature extraction network to learn the invariant features of the domain, and further improve the generalization performance of the model. We did not use sound source separation, achieved an F1 score of 48.25% on the validation dataset and an F1 score of 49.43% on the public evaluation dataset.

System characteristics

Fine-Tuning Using Grid Search & Gradient Visualization

Hou, Bowei and Radzikwoski, Kacper and Farid, Ahmed
Waseda University, Fukuoka, Japan

Abstract

In this technical report, we briefly describe the models used in the task 4 challenge of DCASE2020. We utilized previously available models and fine-tuned them using the grid search algorithm and gradient visualization. This is the first attempt by our team to enter a competition on sound source manipulation.

System characteristics

AUTHOR GUIDELINES FOR DCASE 2020 CHALLENGE TECHNICAL REPORT

Hou, Zhenwei and Hao, Junyong and Peng, Wang
University of Chongqing, Chongqing, China

Abstract

In this technical report, we present the sound event detection system of task 4 (Sound event detection and separation in do- mestic environments) of the DCASE2020 challenge. We propose an improved CRNN that Context Gating and channel attention mechanism are co-embedded into backbone network. It aims to construct a general and efficient attention structure for extracting features of sound events, and give full play to the advantages of attention mechanism in event feature extraction. In the case of replacing the CRNN in the baseline model with the structure we designed and keeping the other parts unchanged, the macro F-score of our model on the validation set is 4 percentage points higher than the baseline.

System characteristics

Guided Multi-Branch Learning Systems For DCASE 2020 Task 4

Huang, Yuxin^1,2 and Lin, Liwei^1,2 and Ma, Shuo¹ and Wang, Xiangdong¹ and Liu, Hong¹ and Qian, Yueliang¹ and Liu, Min³ and Ouch, Kazushige^3
1Bejing Key Laboratory of Mobile Computing and Pervasive Device, Institute of Computing Technology, Chinese Academy of Sciences, Beijing, China, ²University of Chinese Academy of Sciences, Beijing, China, ³Toshiba China R&D Center, Beijing, China

Abstract

In this paper, we describe in detail our systems for DCASE2020 task 4. The systems are based on the first-place system of DCASE 2019 task 4, which adopts the multiple instance learning (MIL) framework with embedding-level attention pooling and a semi-supervised learning approach called guided learning. The multi-branch learning method is then incorporated into the system to further improve the performance. Multiple branches with different pooling strategies (embedding-level or instance-level) and different pooling modules (attention pooling, global max pooling or global average pooling) are used and shares the same feature encoder. To better exploit the synthetic data with strong labels, inspired by multi-task learning, a sound event detection (SED) branch is also added. Therefore, multiple branches pursuing different purposes and focusing on different characteristics of the data can help the feature encoder model the feature space better and avoid over-fitting. To combine sound separation with sound event detection, we train models using the output of the baseline system of sound separation and fuse the event detection results of models with of without sound separation.

System characteristics

Multi-Scale Residual CRNN With Data Augmentation For DCASE 2020 Task 4

JiaKai, Lu
PFU Shanghai Co., LTD, Shanghai, China

Abstract

In this paper, we present our neural network for the DCASE 2020 challenge’s Task 4 (Sound event detection and separation in domestic environments). This task evaluates systems for the largescale detection of sound events using weakly labeled data, and explore the possibility to exploit a large amount of unbalanced and unlabeled training data together with a small weakly annotated training set to improve system performance to doing audio tagging and sound event detection. We propose a mean-teacher model with convolutional neural network (CNN) and recurrent neural network (RNN) to maximize the use of unlabeled in-domain dataset. The architecture is based on our 2018 competition model.

System characteristics

Universal sound separation

Kavalerov, Ilya^1,2 and Wisdom, Scott¹ and Erdogan, Hakan¹ and Patton, Brian¹ and Wilson, Kevin¹ and Le Roux, Jonathan³ and Hershey, John R^1
1Google Research, Cambridge MA, ²Department of Electrical and Computer Engineering, UMD, ³Mitsubishi Electric Research Laboratories (MERL), Cambridge, MA

Abstract

Recent deep learning approaches have achieved impressive performance on speech enhancement and separation tasks. However, these approaches have not been investigated for separating mixtures of arbitrary sounds of different types, a task we refer to as universal sound separation, and it is unknown how performance on speech tasks carries over to non-speech tasks. To study this question, we develop a dataset of mixtures containing arbitrary sounds, and use it to investigate the space of mask-based separation architectures, varying both the overall network architecture and the framewise analysis-synthesis basis for signal transformations. These network architectures include convolutional long short-term memory networks and time-dilated convolution stacks inspired by the recent success of time-domain enhancement networks like ConvTasNet. For the latter architecture, we also propose novel modifications that further improve separation performance. In terms of the framewise analysis-synthesis basis, we explore both a short-time Fourier transform (STFT) and a learnable basis, as used in ConvTasNet. For both of these bases, we also examine the effect of window size. In particular, for STFTs, we find that longer windows (25-50 ms) work best for speech/non-speech separation, while shorter windows (2.5 ms) work best for arbitrary sounds. For learnable bases, shorter windows (2.5 ms) work best on all tasks. Surprisingly, for universal sound separation, STFTs outperform learnable bases. Our best methods produce an improvement in scale-invariant signal-to-distortion ratio of over 13 dB for speech/non-speech separation and close to 10 dB for universal sound separation.

Polyphonic Sound Event Detection Based On Convolutional Recurrent Neural Networks With Semi-Supervised Loss Function For DCASE Challenge 2020 Task 4

Kim, Nam Kyun¹ and Kim, Hong Kook^1,2
1School of Electrical Engineering and Computer Science, Gwangju, South Korea, ²AI Graduate School (GIST), Gwangju Institute of Science and Technology, Gwangju, South Korea

Abstract

This report proposes a polyphonic sound event detection (SED) method for the DCASE 2020 Challenge Task 4. The proposed SED method is based on semi-supervised learning to deal with the different combination of training datasets such as weakly labeled dataset, unlabeled dataset, and strongly labeled synthetic dataset. Especially, the target label of each audio clip from weakly labeled or unlabeled dataset is first predicted by using the mean teacher model that is the DCASE 2020 baseline. The data with predicted labels are used for training the proposed SED model, which consists of CNNs with skip connections and self-attention mechanism, followed by RNNs. In order to compensate for the erroneous prediction of weakly labeled and unlabeled data, a semi-supervised loss function is employed for the proposed SED model. In this work, several versions of the proposed SED model are implemented and evaluated on the validation set according to the different parameter setting for the semi-supervised loss function, and then an ensemble model that combines five-fold validation models is finally selected as our final model.

System characteristics

Sound Event Detection By Consistency Training And Pseudo-Labeling With Feature-Pyramid Convolutional Recurrent Neural Networks

Koh, Chih-Yuan¹ and Chen, You-Siang¹ and Li, Shang-En² and Liu, Yi-Wen¹ and Chien, Jen-Tzung² and Bai, Mingsian R.^1
1National Tsing Hua University, Hsinchu, ²National Chiao Tung University, Hsinchu, Taiwan

Abstract

A event detection system for DCASE 2020 task 4 is presented. To efficiently utilize large amount of unlabeled in-domain data, three semi-supervised learning strategies are applied: 1) interpolation consistency training (ICT), 2) shift consistency training (SCT), 3) weakly pseudo-labeling. In addition, we propose FP-CRNN, a convolutional recurrent neural network which contains feature-pyramid components and is based on the provided baseline. In terms of event-based F-measure, these approaches outperform the baseline, at 34.8%, by a large margin, with an F-measure of 48.4% for the baseline network which is trained with the combination of all three strategies and 49.6% for FP-CRNN with the same training strategies.

System characteristics

Mean Teacher With Sound Source Separation And Data Augmentation For DCASE 2020 Task 4

Liang, Chuming and Ying, Haorong and Chen, Yueyi and Wang, Zhao
Xiaomi AI Lab, Wuhan, China

Abstract

In this paper, we present our system for the DCASE 2020 challenge Task4(Sound event detection and separation in domestic environments). The target of this task is to provide time boundaries of multiple events in an audio recording using a system trained with unlabeled, weakly-labeled and synthetic data. Also, sound source separation is encouraged to use in the system. We propose a mean-teacher model with convolutional and recurrent neural network(CRNN) structure and adopt data augmentation and sound source separation technique to improve the performance of sound event detection.

System characteristics

Semi-Supervised Sound Event Detection Based On Mean Teacher With Power Pooling And Data Augmentation

Liu, Yuzhuo^1,2 and Chen, Chengxin^1,2 and Kuang, Jianzhong^1,2 and Zhang, Pengyuan^1,2
1Institute of Acoustics, Key Laboratory of Speech Acoustics & Content Understanding, Beijing, China, ²University of Chinese Academy of Sciences, Beijing, China

Abstract

In this technical report, we describe the details of the system submitted to DCASE2020 task4: sound event detection (SED) and separation in domestic environments. We mainly focus on the scenario that recognizes sound events without source separation. The training set includes synthetic strongly labeled data, weakly labeled data and unlabeled data. For training SED models with weak labeling, a power pooling function is introduced to generate clip-level predictions from frame-level ones. Additionally, three traditional data augmentation approaches are applied on all data. We also ensemble models with different strategies. Our best system finally achieves an F1 of 49.55% on the validation set.

System characteristics

Convolution-Augmented Transformer For Semi-Supervised Sound Event Detection

Miyazaki, Koichi^{1 and Komatsu, Tatsuya2 and Hayashi, Tomoki1,3 and Watanabe, Shinji4 and Toda, Tomoki1 and Takeda, Kazuya1
1Nagoya University, Japan, 2LINE Corporation, Japan, 3Human Dataware Lab. Co., Japan, 4 Johns Hopkins University, USA}

Abstract

In this technical report, we describe our submission system for DCASE2020 Task4: sound event detection and separation in domestic environments. Our model employs conformer blocks, which combine the self-attention and depth-wise convolution networks, to efficiently capture the global and local context information of an audio feature sequence. In addition to this novel architecture, we further improve the performance by utilizing a mean teacher semi-supervised learning technique, data augmentation, and post-processing optimized for each sound event class. We demonstrate that the proposed method achieves the event-based macro F1 score of 50.7% on the validation set, significantly outperforming that of the baseline score (34.8%).

System characteristics

Joint Acoustic And Supervised Inference For Sound Event Detection

Park, Sangwook and Bellur, Ashwin and Kothinti, Sandeep and Kapurchali, Masoumeh Heidari and Elhilali, Mounya
Johns Hopkins University, Baltimore, USA

Abstract

This is a technical report about a sound event detection system for the task 4 of DCASE2020. The purpose of a sound event detection is to find event class label as well as its time boundaries. To achieve this purpose, we considered several methods such signal enhancement and event boundary detection, and built five systems by integrating these methods with supervised system trained by using Mean Teacher model. In particular, we estimate event boundaries of weakly labeled data by performing a event boundary detection. Then, we used the estimated strong label in training the supervised system. In addition, we adopt a fusion method by calculating weighted averaging posterior over the five outputs from each individual system. In experiments with validation set, we found that a final result of our system shows an improvement about 11 % in class averaging f-score compared to a baseline performance.

System characteristics

Multi-Task Learning Paradigm For Sound Event Detection

Rykaczewski, Krzysztof
Samsung R&D Institute Poland, Warsaw, Poland

Abstract

In this technical report, we describe our system submitted to DCASE2020 task 4. This task evaluates systems for the detection of sound events in domestic environments using large-scale weakly labeled data. To perform this task, we propose resdual convolutional recurrent neural networks (CRNN) as our system and trained by datasets including strong and weak labels. We also use mean-teacher model based on confidence thresholding and smooth embedding method. In addition, we also apply specaugment for the labeled data shortage problem. Finally, we achieve better performance than DCASE2020 baseline system.

System characteristics

Multi-Scale Residual CRNN With Data Augmentation For DCASE 2020 Task 4

Tang, Maolin and Guo, Longyin and Zhang, Yanqiu and Yan, Weiran and Zhao, Qijun
Sichuan University, ChengDu, China

Abstract

In this technical report, we present our method for task 4 of DCASE 2020 challenge (Sound event detection and separation in domestic environments). The goal of the task is to evaluate systems for the detection of sound events using real data either weakly labeled or unlabeled and simulated data that is strongly labeled (with timestamps). We find that models perform well on synthetic data, but may not perform well on real data. We thus improve the baseline [1] by using a variety of data augmentation methods and synthesizing more complex synthetic data for training. Moreover, we present multi-scale residual convolutional recurrent neural network (CRNN) to solve the problem of multi-scale detection. The promising results on the validation set prove the effectiveness of our method.

System characteristics

Training Sound Event Detection On A Heterogeneous Dataset

Turpault, Nicolas and Serizel, Romain
Université de Lorraine, CNRS, Inria, Loria, France

Abstract

Training a sound event detection algorithm on a heterogeneous dataset including both recorded and synthetic soundscapes that can have various labeling granularity is a non-trivial task that can lead to systems requiring several technical choices. These technical choices are often passed from one system to another without being questioned. We propose to perform a detailed analysis of DCASE 2020 task 4 sound event detection baseline with regards to several aspects such as the type of data used for training, the parameters of the mean-teacher or the transformations applied while generating the synthetic soundscapes. Some of the parameters that are usually used as default to replicate other approaches are shown to be sub-optimal.

System characteristics

Improving Sound Event Detection In Domestic Environments Using Sound Separation

Turpault, Nicolas¹ and Wisdom, Scott² and Erdogan, Hakan² and Herhey, John R.² and Serizel, Romain¹ and Fonseca, Eduardo³ and Seetharaman, Prem⁴ and Salomon, Justin^5
1Universite de Lorraine, CNRS, Inria, Loria, Nancy, France, ²Google Research, AI Perception, Cambridge, United States, ³Music Technology Group, Universitat Pompeu Fabra, Barcelona, ⁴Interactive Audio Lab, Northwestern University, Evanston, United States⁵Adobe Research, San Francisco, United States

Abstract

Performing sound event detection on real-world recordings often implies dealing with overlapping target sound events and non-target sounds, also referred to as interference or noise. Until now these problems were mainly tackled at the classifier level. We propose to use sound separation as a pre-processing stage for sound event detection. In this paper we start from a sound separation model trained on the Free Universal Sound Separation dataset and the DCASE 2020 task 4 sound event detection baseline. We explore different methods of combining separated sound sources and the original mixture within the sound event detection. Furthermore, we investigate the impact of adapting the universal sound separation model to the sound event detection data in terms of both separation and sound event detection performance.

System characteristics

Sound Event Detection In Domestic Environments Using Dense Recurrent Neural Network

Yao, Tianchu and Shi, Chuang and Li, Huiyong
University of Electronic Science and Technology of China, Chengdu, China

Abstract

In this paper, we introduce our sound events detection system using a mean-teacher model with convolutional recurrent neural network (CRNN) for DCASE 2020 Task4, which include residual convolutional block and dense recurrent neural network (DRNN) block. To improve the performance of system, we propose to use various methods such as multi-scale input layer, data augmentation, median window filters and model fusion. By combining those method, our system achieves 15% improvement on macro-averaged F-score on the development set, as compared to the baseline.

System characteristics

The Academia Sinica System Of Sound Event Detection And Separation For DCASE 2020

Yen, Hao and Ku^1,2, Pin-Jui and Yen^1,2, Ming-Chi and Lee¹, Hung-Shin and Wang^1,2, Hsin-Min Wang^1
1Institute of Information Science, Academia Sinica, Taiwan, ²Dept. Electrical Engineering, National Taiwan University, Taiwan

Abstract

In this report, we present the system of sound event detection and separation in domestic environments for DCASE 2020. The goal of the task aims to determine which sound events appear in a clip and detailed temporal ranges they occupy. The system is trained by using real data, which are either weakly-labeled or unlabeled, and synthesized data with a strongly annotated label. Our proposed model structure starts with a feature-level front-end based on convolution neural networks (CNN) followed by both embedding-level and instance-level back-end attention modules. To take full advantage of a large amount of unlabeled data, we jointly adopt guided learning mechanism and Mean Teacher, which averages model weights instead of label predictions, to carry out weakly-supervised and semi-supervised learning. A group of adaptive median windows for each sound event is also utilized in post-processing for smoothing frame-level predictions. In the public evaluation set of DCASE 2019, our best system achieves 48.50% event-based F-score, much better than the official baseline performance (38.14%) with a relative improvement of 27.16%. Moreover, in the development set of DCASE 2020, our system is also superior to the baseline while using the student model as the back-end classifier. The F1-score is relatively improved by 32.91%.

System characteristics