Large-scale weakly supervised
sound event detection for smart cars

@techreport{Adavanne2017,
    Author = "Adavanne, Sharath and Virtanen, Tuomas",
    title = "Sound Event Detection Using Weakly Labeled Dataset with Stacked Convolutional and Recurrent Neural Network",
    institution = "DCASE2017 Challenge",
    year = "2017",
    month = "September",
    abstract = "This paper proposes a neural network architecture and training scheme to learn the start and end time of sound events (strong labels) in an audio recording given just the list of sound events existing in the audio without time information (weak labels). We achieve this by using a stacked convolutional and recurrent neural network with two prediction layers in sequence one for the strong followed by the weak label. The network is trained using frame-wise log mel-band energy as the input audio feature, and weak labels provided in the dataset as labels for the weak label prediction layer. Strong labels are generated by replicating the weak labels as many number of times as the frames in the input audio feature, and used for strong label layer during training. We propose to control what the network learns from the weak and strong labels by different weighting for the loss computed in the two prediction layers. The proposed method is evaluated on a publicly available dataset of 155 hours with 17 sound event classes. The method achieves the best error rate of 0.84 for strong labels and F-score of 43.3\% for weak labels on the unseen test split."
}

@techreport{Badlani2017,
    Author = "Badlani, Rohan and Shah, Ankit and Elizalde, Benjamin",
    title = "{DCASE} 2017 Challenge Setup: Tasks, Datasets and Baseline System",
    institution = "DCASE2017 Challenge",
    year = "2017",
    month = "September",
    abstract = "DCASE 2017 Challenge consists of four tasks: acoustic scene classification, detection of rare sound events, sound event detection in real-life audio, and large-scale weakly supervised sound event detection for smart cars. This paper presents the setup of these tasks: task definition, dataset, experimental setup, and baseline system results on the development dataset. The baseline systems for all tasks rely on the same implementation using multilayer perceptron and log mel-energies, but differ in the structure of the output layer and the decision making process, as well as the evaluation of system output using task specific metrics."
}

@techreport{Chou2017,
    Author = "Chou, Szu-Yu and Jang, Jyh-Shing and Yang, Yi-Hsuan",
    title = "{FrameCNN}: A Weakly-Supervised Learning Framework for Frame-Wise Acoustic Event Detection and Classification",
    institution = "DCASE2017 Challenge",
    year = "2017",
    month = "September",
    abstract = "In this paper, we describe our contribution to the challenge of detection and classification of acoustic scenes and events (DCASE2017).We propose framCNN, a novel weakly supervised learning frame-work that improves the performance of convolutional neural net-work (CNN) for acoustic event detection by attending to details of each sound at various temporal levels. Most existing weakly-supervised frameworks replace fully-connected network with global average pooling after the final convolution layer. Such a method tends to identify only a few discriminative parts, leading to sub-optimal localization and classification accuracy. The key idea of our approach is to consciously classify the sound of each frame given by the corresponding label. The idea is general and can be applied to any network for achieving sound event detection and improving the performance of sound event classification. In acoustic scene classification (Task1), our approach obtained an average accuracy of 99.2\% on the four-fold cross-validation for acoustic scene recognition, comparing to the provided baseline of 74.8\%. In the large-scale weakly supervised sound event detection for smart cars(Task4), we obtained a F-score 53.8\% for sound event audio tagging (subtask A), compared to the baseline of 19.8\%, and a F-score32.8\% for sound event detection (subtask B), compared to the base-line of 11.4\%"
}

@techreport{Kukanov2017,
    Author = "Kukanov, Ivan and Hautamäki, Ville and Lee, Kong Aik",
    title = "Recurrent Neural Network and Maximal Figure of Merit for Acoustic Event Detection",
    institution = "DCASE2017 Challenge",
    year = "2017",
    month = "September",
    abstract = "In this report, we describe the systems submitted to the DCASE 2017 challenge. In particular, we explored convolutional recurrent neural network (CRNN) for acoustic scene classification (Task 1). For the weakly supervised sound event detection (Task 4), we utilized CRNN by embedding maximal figure-of-merit (CRNN-MFoM) into the binary cross-entropy objective function. On the development data set, the CRNN model achieves an average 14.7\% relative accuracy improvement on the classification Task 1, the CRNN-MFoM improves F1-score from 10.9\% to 33.5\% on the detection Task 4 compared to the baseline system."
}

@techreport{Lee2017,
    Author = "Lee, Jongpil and Park, Jiyoung and Nam, Juhan",
    title = "Combining Multi-Scale Features Using Sample-Level Deep Convolutional Neural Networks for Weakly Supervised Sound Event Detection",
    institution = "DCASE2017 Challenge",
    year = "2017",
    month = "September",
    abstract = "This paper describes our method submitted to large-scale weakly supervised sound event detection for smart cars in DCASE Challenge 2017. It it based on two techniques that have been previously suggested for music auto-tagging. One is training sample-level deep convolutional neural networks using raw waveforms as feature extractors. The other is aggregating features on multi-scaled models of the CNNs and making final predictions. With this approach, we achieved the best results, 44.3\% in F-score on subtask A (audio tagging) and 0.84 in error rate on subtask B (sound event detection). Finally, we visualize hierarchically learned filters from the challenge dataset in each layer of the raw waveform based model to explain how they discriminate the events."
}

@techreport{Lee2017a,
    Author = "Lee, Donmoon and Lee, Subin and Han, Yoonchang and Lee, Kyogu",
    title = "Ensemble of Convolutional Neural Networks for Weakly-Supervised Sound Event Detection Using Multiple Scale Input",
    institution = "DCASE2017 Challenge",
    year = "2017",
    month = "September",
    abstract = "In this paper, we use ensemble of convolutional neural network models that use the various analysis window to detect audio events in the automotive environment. When detecting the presence of audio events, global input based model that uses the entire audio clip works better. On the other hand, segmented input based models works better in finding the accurate position of the event. Experimental results for weakly-labeled audio data confirm the performance trade-off between the two tasks, depending on the length of input audio. By combining the predictions of various models, the proposed system achieved 0.4762 in the clip-based F1-score and 0.7167 in the segment-based error rate."
}

@techreport{Salamon2017,
    Author = "Salamon, Justin and McFee, Brian and Li, Peter",
    title = "{DCASE} 2017 Submission: Multiple Instance Learning for Sound Event Detection",
    institution = "DCASE2017 Challenge",
    year = "2017",
    month = "September",
    abstract = "This extended abstract describes the design and implementation of a multiple instance learning model for sound event detection. The submitted systems use a convolutional-recurrent neural network (CRNN) architecture to learn strong (temporally localized) predictors from weakly labeled data. Four variants of the proposed methods were submitted to DCASE 2017, Task 4."
}

@techreport{Tseng2017,
    Author = "Tseng, Shaoyen and Li, Juncheng Billy",
    title = "Large-Scale Weakly Supervised Sound Event Detection",
    institution = "DCASE2017 Challenge",
    year = "2017",
    month = "September",
    abstract = "State-of-the-art audio event detection (AED) systems fully rely on supervised-learning based on strongly labeled data.The dependence on strong labels severely limits the scalability of AED work. Large-scale manually annotated datasets are difficult and expensive to collect [1], whereas weakly labeled data could be much easier to acquire. In weakly labeled data, we only need to determine whether an event in the recording is present or absent. This not only makes manual labeling significantly easier but also makes automatically infer-ring labels from online multimedia or audio meta-information(titles, tags, etc) possible [2]. This work employs a subset of Google’s AudioSet [3], which is a large number of weakly labeled YouTube video excerpts. The subset focuses on transportation and warning sounds and consists of 17 sound events divided into two categories: Warning and Vehicle.We perform experiments on 3 sets of features, including standard Mel frequency cepstral coefficients (MFCC) and log-Mel spectrograms and pre-trained embeddings extracted from a deep convolutional network. Our system employs multiple instance learning (MIL) [4] approaches to deal with weak labels by bagging them to positive or negative bags. We apply 4 models, Deep Neural Network (DNN), Recurrent Neural Network (RNN) and Convolutional Deep NeuralNetwork. Using the late-fusion approach, we improve the performance of the baseline audio tagging (Subtask A) F1score 13.1\% by 18.1\%.The embeddings extracted by the convolutional neural networks significantly boosts the performance of all the models."
}

@techreport{Vu2017,
    Author = "Vu, Toan and Dang, An and Wang, Jia-Ching",
    title = "Deep Learning for {DCASE2017} Challenge",
    institution = "DCASE2017 Challenge",
    year = "2017",
    month = "September",
    abstract = "This paper reports our results on all tasks of DCASE challenge 2017 which are acoustic scene classification, detection of rare sound events, sound event detection in real life audio, and large-scale weakly supervised sound event detection for smart cars. Our proposed methods are developed based on two favorite neural networks which are convolutional neural networks (CNNs) and recurrent neural networks (RNNs). Experiments show that our proposed methods outperform the baseline."
}

@techreport{Xu2017,
    Author = "Xu, Yong and Kong, Qiuqiang and Wang, Wenwu and Plumbley, Mark D.",
    title = "Surrey-{CVSSP} System for {DCASE2017} Challenge Task4",
    institution = "DCASE2017 Challenge",
    year = "2017",
    month = "September",
    abstract = "In this technique report, we present a bunch of methods for the task 4 of Detection and Classification of Acoustic Scenes and Events 2017 (DCASE2017) challenge. This task evaluates systems for the large-scale detection of sound events using weakly labeled training data. The data are YouTube video excerpts focusing on transportation and warnings due to their industry applications. There are two tasks, audio tagging and sound event detection from weakly labeled data. Convolutional neural network (CNN) and gated recurrent unit (GRU) based recurrent neural network (RNN) are adopted as our basic framework. We proposed a learnable gating activation function for selecting informative local features. Attention-based scheme is used for localizing the specific events in a weakly-supervised mode. A new batch-level balancing strategy is also proposed to tackle the data unbalancing problem. Fusion of posteriors from different systems are found effective to improve the performance. In a summary, we get 61\% F-value for the audio tagging subtask and 0.72 error rate (ER) for the sound event detection subtask on the development set. While the official multilayer perceptron (MLP) based baseline just obtained 13.1\% F-value for the audio tagging and 1.02 for the sound event detection."
}