The proceedings of the DCASE2019 Workshop have been published as an electronic publication by New York University:
Urban Noise Monitoring in the Stadtlärm Project - A Field Report
Jakob Abeßer1, Marco Götze2, Tobias Clauß1, Dominik Zapf1, Christian Kühn1, Hanna Lukashevich3, Stephanie Kühnlenz4, and Stylianos I. Mimilakis1
1Fraunhofer IDMT, 2IMMS Institut für Mikroelektronik- und Mechatronik-Systeme gemeinnützige GmbH, 3Fraunhofer Institute for Digital Media Technology, Germany, 4Software-Service John GmbH
As noise pollution in urban environments is constantly rising, novel smart city applications are required for acoustic monitoring and municipal decision making. This paper summarizes the experiences made during the field test of the Stadtlärm system for distributed noise measurement in summer/fall of 2018 in Jena, Germany.
Urban Sound Tagging using Convolutional Neural Networks
In this paper, we propose a framework for environmental sound classification in a low-data context (less than 100 labeled examples per class). We show that using pre-trained image classification models along with usage of data augmentation techniques results in higher performance over alternative approaches. We applied this system to the task of Urban Sound Tagging, part of the DCASE 2019. The objective was to label different sources of noise from raw audio data. A modified form of MobileNetV2, a convolutional neural network (CNN) model was trained to classify both coarse and fine tags jointly. The proposed model uses log-scaled Mel-spectrogram as the representation format for the audio data. Mixup, Random erasing, scaling, and shifting are used as data augmentation techniques. A second model that uses scaled labels was built to account for human errors in the annotations. The proposed model achieved the first rank on the leaderboard with Micro-AUPRC values of 0.751 and 0.860 on fine and coarse tags, respectively.
A Multi-room Reverberant Dataset for Sound Event Localization and Detection
Sharath Adavanne1, Archontis Politis1, and Tuomas Virtanen1
This paper presents the sound event localization and detection (SELD) task setup for the DCASE 2019 challenge. The goal of the SELD task is to detect the temporal activities of a known set of sound event classes, and further localize them in space when active. As part of the challenge, a synthesized dataset where each sound event associated with a spatial coordinate represented using azimuth and elevation angles is provided. These sound events are spatialized using real-life impulse responses collected at multiple spatial coordinates in five different rooms with varying dimensions and material properties. A baseline SELD method employing a convolutional recurrent neural network is used to generate benchmark scores for this reverberant dataset. The benchmark scores are obtained using the recommended cross-validation setup.
Cites: 17 ( see at Google Scholar )
Sound Event Classification and Detection with Weakly Labeled Data
Sharath Adavanne1, Haytham Fayek2, and Vladimir Tourbabin2
1Tampere University, 2Facebook
The Sound Event Classification (SEC) task involves recognizing the set of active sound events in an audio recording. The Sound Event Detection (SED) task involves, in addition to SEC, detecting the temporal onset and offset of every sound event in an audio recording. Generally, SEC and SED are treated as supervised classification tasks that require labeled datasets. SEC only requires weak labels, i.e., annotation of active sound events, without the temporal information, whereas SED requires strong labels, i.e., annotation of the onset and offset times of every sound event, which makes annotation for SED more tedious than for SEC. In this paper, we propose two methods for joint SEC and SED using weakly labeled data: a Fully Convolutional Network (FCN) and a novel method that combines a Convolutional Neural Network with an attention layer (CNNatt). Unlike most prior work, the proposed methods do not assume that the weak labels are active during the entire recording and can scale to large datasets. We report state-of-the-art SEC results obtained with the largest weakly labeled dataset - Audioset.
Localization, Detection and Tracking of Multiple Moving Sound Sources with a Convolutional Recurrent Neural Network
Sharath Adavanne1, Archontis Politis1, and Tuomas Virtanen1
This paper investigates the joint localization, detection, and tracking of sound events using a convolutional recurrent neural network (CRNN). We use a CRNN previously proposed for the localization and detection of stationary sources, and show that the recurrent layers enable the spatial tracking of moving sources when trained with dynamic scenes. The tracking performance of the CRNN is compared with a stand-alone tracking method that combines a multi-source (DOA) estimator and a particle filter. Their respective performance is evaluated in various acoustic conditions such as anechoic and reverberant scenarios, stationary and moving sources at several angular velocities, and with a varying number of overlapping sources. The results show that the CRNN manages to track multiple sources more consistently than the parametric method across acoustic scenarios, but at the cost of higher localization error.
Cites: 1 ( see at Google Scholar )
DCASE 2019 Task 2: Multitask Learning, Semi-supervised Learning and Model Ensemble with Noisy Data for Audio Tagging
Osamu Akiyama1, and Junya Sato1
This paper describes our approach to the DCASE 2019 challenge Task 2: Audio tagging with noisy labels and minimal supervision. This task is a multi-label audio classification with 80 classes. The training data is composed of a small amount of reliably labeled data (curated data) and a larger amount of data with unreliable labels (noisy data). Additionally, there is a difference in data distribution between curated data and noisy data. To tackle this difficulty, we propose three strategies. The first is multitask learning using noisy data. The second is semi-supervised learning using noisy data and labels that are relabeled using trained models’ predictions. The third is an ensemble method that averages models trained with different time length. By using these methods, our solution was ranked in 3rd place on the public leaderboard (LB) with a label-weighted label-ranking average precision (lwlrap) score of 0.750 and ranked in 4th place on the private LB with a lwlrap score of 0.75787. The code of our solution is available at https://github.com/OsciiArt/Freesound-Audio-Tagging-2019.
Polyphonic Sound Event Detection and Localization using a Two-Stage Strategy
Yin Cao1, Qiuqiang Kong1, Turab Iqbal1, Fengyan An2, Wenwu Wang1, and Mark Plumbley1
1University of Surrey, 2Qingdao University of Technology
Sound event detection (SED) and localization refer to recognizing sound events and estimating their spatial and temporal locations. Using neural networks has become the prevailing method for SED. In the area of sound localization, which is usually performed by estimating the direction of arrival (DOA), learning-based methods have recently been developed. In this paper, it is experimentally shown that the trained SED model is able to contribute to the direction of arrival estimation (DOAE). However, joint training of SED and DOAE degrades the performance of both. Based on these results, a two-stage polyphonic sound event detection and localization method is proposed. The method learns SED first, after which the learned feature layers are transferred for DOAE. It then uses the SED ground truth as a mask to train DOAE. The proposed method is evaluated on the DCASE 2019 Task 3 dataset, which contains different overlapping sound events in different environments. Experimental results show that the proposed method is able to improve the performance of both SED and DOAE, and also performs significantly better than the baseline method.
Cites: 11 ( see at Google Scholar )
SONYC Urban Sound Tagging (SONYC-UST): A Multilabel Dataset from an Urban Acoustic Sensor Network
Mark Cartwright1, Ana Elisa Mendez Mendez1, Jason Cramer1, Vincent Lostanlen2, Graham Dove1, Ho-Hsiang Wu1, Justin Salamon3, Oded Nov1, and Juan Bello1
1New York University, 2Cornell Lab of Ornithology, 3Adobe Research
SONYC Urban Sound Tagging (SONYC-UST) is a dataset for the development and evaluation of machine listening systems for real-world urban noise monitoring. It consists of 3068 audio recordings from the "Sounds of New York City" (SONYC) acoustic sensor network. Via the Zooniverse citizen science platform, volunteers tagged the presence of 23 fine-grained classes that were chosen in consultation with the New York City Department of Environmental Protection. These 23 fine-grained classes can be grouped into eight coarse-grained classes. In this work, we describe the collection of this dataset, metrics used to evaluate tagging systems, and the results of a simple baseline model.
Cites: 3 ( see at Google Scholar )
Non-Negative Matrix Factorization-Convolutional Neural Network (NMF-CNN) for Sound Event Detection
Teck Kai Chan1, Cheng Siong Chin1, and Ye Li2
1Newcastle University, 2Visenti Pte Ltd
The main scientific question of this year DCASE chal-lenge, Task 4 - Sound Event Detection in Domestic Environments, is to investigate the types of data (strongly labeled synthetic data, weakly labeled data, unlabeled in domain data) required to achieve the best performing system. In this paper, we proposed a deep learning model that integrates Non-Negative Matrix Factorization (NMF) with Convolutional Neural Net-work (CNN). The key idea of such integration is to use NMF to provide an approximate strong label to the weakly labeled data. Such integration was able to achieve a higher event based F1-score as compared to the baseline system (Evaluation Dataset: 30.39% vs. 23.7%, Validation Dataset: 31% vs. 25.8%). By com-paring the validation results with other participants, the proposed system was ranked 8th among 19 teams (inclusive of the baseline system) in this year Task 4 challenge.
Acoustic Scene Classification Based on a Large-margin Factorized CNN
Janghoon Cho1, Sungrack Yun2, Hyoungwoo Park3, Jungyun Eum1, and Kyuwoong Hwang1
1Qualcomm AI Research, 2Qualcomm korea, 3Qualcomm AI research
In this paper, we present an acoustic scene classification framework based on a large-margin factorized convolutional neural network (CNN). We adopt the factorized CNN to learn the patterns in the time-frequency domain by factorizing the 2D kernel into two separate 1D kernels. The factorized kernel leads to learn the main component of two patterns: the long-term ambient and short-term event sounds which are the key patterns of the audio scene classification. In training our model, we consider the loss function based on the triplet sampling such that the same audio scene samples from different environments are minimized, and simultaneously the different audio scene samples are maximized. With this loss function, the samples from the same audio scene are clustered independently of the environment, and thus we can get the classifier with better generalization ability in an unseen environment. We evaluated our audio scene classification framework using the dataset of the DCASE challenge 2019 task1A. Experimental results show that the proposed algorithm improves the performance of the baseline network and reduces the number of parameters to one third. Furthermore, the performance gain is higher on unseen data, and it shows that the proposed algorithm has better generalization ability.
Hierarchical Detection of Sound Events and their Localization Using Convolutional Neural Networks with Adaptive Thresholds
Sotirios Panagiotis Chytas1, and Gerasimos Potamianos1
1ECE, University of Thessaly
This paper details our approach to Task 3 of the DCASE’19 Challenge, namely sound event localization and detection (SELD). Our system is based on multi-channel convolutional neural networks (CNNs), combined with data augmentation and ensembling. Specifically, it follows a hierarchical approach that first determines adaptive thresholds for the multi-label sound event detection (SED) problem, based on a CNN operating on spectrograms over long duration windows. It then exploits the derived thresholds in an ensemble of CNNs operating on raw waveforms over shorter-duration sliding windows to provide event segmentation and labeling. Finally, it employs event localization CNNs to yield direction-of-arrival (DOA) source estimates of the detected sound events. The system is developed and evaluated on the microphone-array set of Task 3. Compared to the baseline of the Challenge organizers, on the development set it achieves relative improvements of 12% in SED error, 2% in F-score, 36% in DOA error, and 3% in the combined SELD metric, but trails significantly in frame-recall, whereas on the evaluation set it achieves relative improvements of 3% in SED, 51% in DOA, and 4% in SELD errors. Overall though, the system lags significantly behind the best Task 3 submission, achieving a combined SELD error of 0.2033 against 0.044 of the latter
Cites: 1 ( see at Google Scholar )
GCC-PHAT Cross-Correlation Audio Features for Simultaneous Sound Event Localization and Detection (SELD) on Multiple Rooms
Hector Cordourier1, Paulo Lopez Meyer1, Jonathan Huang1, Juan Del Hoyo Ontiveros2, and Hong Lu2
1Intel, 2Intel Labs
In this work, we show a simultaneous sound event localization and detection (SELD) system, with enhanced acoustic features, in which we propose using the well-known Generalized Cross Correlation (GCC) PATH algorithm, to augment the magnitude and phase regular Fourier spectra features at each frame. GCC-PHAT has already been used for some time to calculate the Time Difference of Arrival (TDOA) in simultaneous audio signals, in moderately reverberant environments, using classic signal processing techniques, and can assist audio source localization in current deep learning machines. The neural net architecture we used is a Convolutional Recurrent Neural Network (CRNN), and is tested using the sound database prepared for the Task 3 of the 2019 DCASE Challenge. In the challenge results, our proposed system was able to achieve 20.8° of direction of arrival error, 85.6\% frame recall, 86.5% F-score and 0.22 error rate detection in evaluation samples.
Language Modelling for Sound Event Detection with Teacher Forcing and Scheduled Sampling
Konstantinos Drossos1, Shayan Gharib1, Paul Magron1, and Tuomas Virtanen1
A sound event detection (SED) method typically takes as an input a sequence of audio frames and predicts the activities of sound events in each frame. In real-life recordings, the sound events exhibit some temporal structure: for instance, a "car horn" will likely be followed by a "car passing by". While this temporal structure is widely exploited in sequence prediction tasks (e.g., in machine translation), where language models (LM) are exploited, it is not satisfactorily modeled in SED. In this work we propose a method which allows a recurrent neural network (RNN) to learn an LM for the SED task. The method conditions the input of the RNN with the activities of classes at the previous time step. We evaluate our method using F1 score and error rate (ER) over three different and publicly available datasets; the TUT-SED Synthetic 2016 and the TUT Sound Events 2016 and 2017 datasets. The obtained results show an increase of 6% and 3% at the F1 (higher is better) and a decrease of 3% and 2% at ER (lower is better) for the TUT Sound Events 2016 and 2017 datasets, respectively, when using our method. On the contrary, with our method there is a decrease of 10% at F1 score and an increase of 11% at ER for the TUT-SED Synthetic 2016 dataset.
Convolutional Recurrent Neural Network and Data Augmentation for Audio Tagging with Noisy Labels and Minimal Supervision
Janek Ebbers1, and Reinhold Häb-Umbach2
1Paderborn University, 2University of Paderborn
In this paper we present our audio tagging system for the DCASE 2019 Challenge Task 2. We propose a model consisting of a convolutional front end using log-mel-energies as input features, a recurrent neural network sequence encoder and a fully connected classifier network outputting an activity probability for each of the 80 considered event classes. Due to the recurrent neural network, which encodes a whole sequence into a single vector, our model is able to process sequences of varying lengths. The model is trained with only little manually labeled training data and a larger amount of automatically labeled web data, which hence suffers from label noise. To efficiently train the model with the provided data we use various data augmentation to prevent overfitting and improve generalization. Our best submitted system achieves a label-weighted label-ranking average precision (lwlrap) of 75.5% on the private test set which is an absolute improvement of 21.7% over the baseline. This system scored the second place in the teams ranking of the DCASE 2019 Challenge Task 2 and the fifth place in the Kaggle competition ''Freesound Audio Tagging 2019'' with more than 400 participants. After the challenge ended we further improved performance to 76.5% lwlrap setting a new state-of-the-art on this dataset.
Cites: 1 ( see at Google Scholar )
Audio Tagging with Noisy Labels and Minimal Supervision
Eduardo Fonseca1, Manoj Plakal2, Frederic Font3, Daniel P.W. Ellis4, and Xavier Serra5
1Music Technology Group, Universitat Pompeu Fabra, 2Google, Inc., 3Music Technology Group - Universitat Pompeu Fabra, 4Google, Inc., 5Universitat Pompeu Fabra
This paper introduces Task 2 of the DCASE2019 Challenge, titled "Audio tagging with noisy labels and minimal supervision". This task was hosted on the Kaggle platform as "Freesound Audio Tagging 2019". The task evaluates systems for multi-label audio tagging using a large set of noisy-labeled data, and a much smaller set of manually-labeled data, under a large vocabulary setting of 80 everyday sound classes. In addition, the proposed dataset poses an acoustic mismatch problem between the noisy train set and the test set due to the fact that they come from different web audio sources. This can correspond to a realistic scenario given by the difficulty of gathering large amounts of manually labeled data. We present the task setup, the FSDKaggle2019 dataset prepared for this scientific evaluation, and a baseline system consisting of a convolutional neural network. All these resources are freely available.
Cites: 17 ( see at Google Scholar )
Robust Non-negative Block Sparse Coding for Acoustic Novelty Detection
Ritwik Giri1, Arvindh Krishnaswamy1, and Karim Helwani1
In this paper we address the problem of detecting previously unseen novel audio events in the presence of real-life acoustic backgrounds. Specifically, during training, we learn subspaces corresponding to each acoustic background, and during testing the audio frame in question is decomposed into a component that lies on the mixture of subspaces and a supergaussian outlier component. Based on the energy in the estimated outlier component a decision is made, whether or not the current frame is an acoustic novelty. We compare our proposed method with state of the art auto-encoder based approaches and also with a traditional supervised Nonnegative Matrix Factorization (NMF) based method using a publicly available dataset - A3Novelty. We also present results using our own dataset created by mixing novel/rare sounds such as gunshots, glass-breaking and sirens, with normal background sounds for various event to background ratios (in dB).
Sound Source Localisation in Ambisonic Audio Using Peak Clustering
Marc Green1, and Damian Murphy1
1University of York
Accurate sound source direction-of-arrival and trajectory estimation in 3D is a key component of acoustic scene analysis for many applications, including as part of polyphonic sound event detection systems. Recently, a number of systems have been proposed which perform this function with first-order Ambisonic audio and can work well, though typically performance drops when the polyphony is increased. This paper introduces a novel system for source localisation using spherical harmonic beamforming and unsupervised peak clustering. The performance of the system is investigated using synthetic scenes in first to fourth order Ambisonics and featuring up to three overlapping sounds. It is shown that use of second-order Ambisonics results in significantly increased performance relative to first-order. Using third and fourth-order Ambisonics also results in improvements, though these are not so pronounced.
Sound Event Localization and Detection Using CRNN on Pairs of Microphones
Francois Grondin1, Iwona Sobieraj2, Mark Plumbley2, and James Glass3
1Massachusetts Institute of Technology, 2University of Surrey, 3MIT
This paper proposes sound event localization and detection methods from multichannel recording. The proposed system is based on two Convolutional Recurrent Neural Networks (CRNNs) to perform sound event detection (SED) and time difference of arrival (TDOA) estimation on each pair of microphones in a microphone array. In this paper, the system is evaluated with a four-microphone array, and thus combines the results from six pairs of microphones to provide a final classification and a 3-D direction of arrival (DOA) estimate. Results demonstrate that the proposed approach outperforms the DCASE 2019 baseline system.
Multiple Neural Networks with Ensemble Method for Audio Tagging with Noisy Labels and Minimal Supervision
Kexin He1, Yuhan Shen1, and Wei-Qiang Zhang1
In this paper, we describe our system for the Task 2 of Detection and Classification of Acoustic Scenes and Events (DCASE) 2019 Challenge: Audio tagging with noisy labels and minimal supervision. This task provides a small amount of verified data (curated data) and a larger quantity of unverified data (noisy data) as training data. Each audio clip contains one or more sound events, so it can be considered as a multi-label audio classification task. To tackle this problem, we mainly use four strategies. The first is a sigmoid-softmax activation to deal with so-called sparse multi-label classification. The second is a staged training strategy to learn from noisy data. The third is a post-processing method that normalizes output scores for each sound class. The last is an ensemble method that averages models learned with multiple neural networks and various acoustic features. All of the above strategies contribute to our system significantly. Our final system achieved labelweighted label-ranking average precision (lwlrap) scores of 0.758 on the private test dataset and 0.742 on the public test dataset, winning the 2nd place in DCASE 2019 Challenge Task 2.
Acoustic Scene Classification Using Deep Learning-based Ensemble Averaging
Jonathan Huang1, Hong Lu2, Paulo Lopez Meyer1, Hector Cordourier1, and Juan Del Hoyo Ontiveros2
1Intel, 2Intel Labs
In our submission to the DCASE 2019 Task 1a, we have explored the use of four different deep learning based neural networks architectures: Vgg12, ResNet50, AclNet, and AclSincNet. In order to improve performance, these four network architectures were pre-trained with Audioset data, and then fine-tuned over the development set for the task. The outputs produced by these networks, due to the diversity of feature front-end and of architecture differences, proved to be complementary when fused together. The ensemble of these models' outputs improved from best single model accuracy of 77.9% to 83.0% on the validation set, trained with the challenge default's development split. For the challenge's evaluation set, our best ensemble resulted in 81.3% of classification accuracy.
Cites: 1 ( see at Google Scholar )
Neural Audio Captioning Based on Conditional Sequence-to-Sequence Model
Shota Ikawa1, and Kunio Kashino2
1the University of Tokyo, 2NTT Communication Science Laboratories
We propose an audio captioning system that describes non-speech audio signals in the form of natural language. Unlike existing systems, this system can generate a sentence describing sounds, rather than an object label or onomatopoeia. This allows the description to include more information, such as how the sound is heard and how the tone or volume changes over time, and can accommodate unknown sounds. A major problem in realizing this capability is that the validity of the description depends not only on the sound itself but also on the situation or context. To address this problem, a conditional sequence-to-sequence model is proposed. In this model, a parameter called ''specificity'' is introduced as a condition to control the amount of information contained in the output text and generate an appropriate description. Experiments show that the proposed model works effectively.
RU Multichannel Domestic Acoustic Scenes 2019: A Multichannel Dataset Recorded by Distributed Microphones with Various Properties
Keisuke Imoto1, and Nobutaka Ono2
1Ritsumeikan University, 2Tokyo Metropolitan University
Acoustic scene analysis has seen extensive development recently because it is used in applications such as monitoring, surveillance, life-logging, and advanced multimedia retrieval systems. Acoustic sensors, such as those used in smartphones, wearable devices, and surveillance cameras, have recently rapidly increased in number. The simultaneous use of these acoustic sensors will enable a more reliable analysis of acoustic scenes because they can be utilized for the extraction of spatial information or application of ensemble techniques. However, there are only a few datasets for acoustic scene analysis that make use of multichannel acoustic sensors, and to the best of our knowledge, no large-scale open datasets recorded with multichannel acoustic sensors composed of different devices. In this paper, we thus introduce a new publicly available dataset for acoustic scene analysis, which was recorded by distributed microphones with various characteristics. The dataset is freely available from http://www.ksuke.net/dataset.
Shuffling and Mixing Data Augmentation for Environmental Sound Classification
Tadanobu Inoue1, Phongtharin Vinayavekhin1, Shiqiang Wang1, David Wood1, Asim Munawar1, Bong Jun Ko1, Nancy Greco1, and Ryuki Tachibana1
Smart speakers have been recently adopted and widely used in consumer homes, largely as a communication interface between human and machines. In addition, these speakers can be used to monitor sounds other than human voice, for example, to watch over elderly people living alone, and to notify if there are changes in their usual activities that may affect their health. In this paper, we focus on the sound classification using machine learning, which usually requires a lot of training data to achieve good accuracy. Our main contribution is a data augmentation technique that generates new sound by shuffling and mixing two existing sounds of the same class in the dataset. This technique creates new variations on both the temporal sequence and the density of the sound events. We show in DCASE 2018 Task 5 that the proposed data augmentation method with our proposed convolutional neural network (CNN) achieves an average of macro-averaged F1 score of 89.95% over 4 folds of the development dataset. This is a significant improvement from the baseline result of 84.50%. In addition, we also verify that our proposed data augmentation technique can improve the classification performance on the Urban Sound 8K dataset.
Distilling the Knowledge of Specialist Deep Neural Networks in Acoustic Scene Classification
Jee-weon Jung1, HeeSoo Heo1, Hye-jin Shim1, and Ha-Jin Yu1
1University of Seoul
Different acoustic scenes that share common properties are one of the main obstacles that hinder successful acoustic scene classification. Top two most confusing pairs of acoustic scenes, ‘airport- shopping mall’ and ‘metro-tram’ have occupied more than half of the total misclassified audio segments, demonstrating the need for consideration of these pairs. In this study, we exploited two specialist models in addition to a baseline model and applied the knowledge distillation framework from those three models into a single deep neural network. A specialist model refers to a model that concentrates on discriminating a pair of two similar scenes. We hypothesized that knowledge distillation from multiple specialist models and a pre-trained baseline model into a single model could gather the superiority of each specialist model and achieve similar effect to an ensemble of these models. In the results of the Detection and Classification of Acoustic Scenes and Events 2019 challenge, the distilled single model showed a classification accuracy of 81.2 %, equivalent to the performance of an ensemble of the baseline and two specialist models.
Sound Source Detection, Localization and Classification using Consecutive Ensemble of CRNN Models
Sławomir Kapka1, and Mateusz Lewandowski1
1Samsung R&D Institute Poland
In this paper, we describe our method for DCASE2019 task 3: Sound Event Localization and Detection (SELD). We use four CRNN SELDnet-like single output models which run in a consecutive manner to recover all possible information of occurring events. We decompose the SELD task into estimating number of active sources, estimating direction of arrival of a single source, estimating direction of arrival of the second source where the direction of the first one is known and a multi-label classification task. We use custom consecutive ensemble to predict events' onset, offset, direction of arrival and class. The proposed approach is evaluated on the TAU Spatial Sound Events 2019 - Ambisonic and it is compared with other participants' submissions.
Cites: 3 ( see at Google Scholar )
Receptive-Field-Regularized CNN Variants for Acoustic Scene Classification
Khaled Koutini1, Hamid Eghbal-zadeh2, and Gerhard Widmer1
1Johannes Kepler University, 2LIT AI Lab & Johannes Kepler University, Institute of Computational Perception
Acoustic scene classification and related tasks have been dominated by Convolutional Neural Networks (CNNs). Topperforming CNNs use mainly audio spectograms as input and borrow their architectural design primarily from computer vision. A recent study has shown that restricting the receptive field (RF) of CNNs in appropriate ways is crucial for their performance, robustness and generalization in audio tasks. One side effect of restricting the RF of CNNs is that more frequency information is lost. In this paper, we perform a systematic investigation of different RF configuration for various CNN architectures on the DCASE 2019 Task 1.A dataset. Second, we introduce Frequency Aware CNNs to compensate for the lack of frequency information caused by the restricted RF, and experimentally determine if and in what RF ranges they yield additional improvement. The result of these investigations are several well-performing submissions to different tasks in the DCASE 2019 Challenge.
Cites: 1 ( see at Google Scholar )
SpecAugment for Sound Event Detection in Domestic Environments using Ensemble of Convolutional Recurrent Neural Networks
In this paper, we present a method to detect sound events in domestic environments using small weakly labeled data, large unlabeled data, and strongly labeled synthetic data as proposed in the Detection and Classification of Acoustic Scenes and Events 2019 Challenge task 4. To solve the problem, we use a convolutional recurrent neural network composed of stacks of convolutional neural networks and bi-directional gated recurrent units. Moreover, we propose various methods such as SpecAugment, event activity detection, multi-median filtering, mean-teacher model, and an ensemble of neural networks to improve performance. By combining the proposed methods, sound event detection performance can be enhanced, compared with the baseline algorithm. Consequently, performance evaluation shows that the proposed method provides detection results of 40.89% for event-based metrics and 66.17% for segment-based metrics. For the evaluation dataset, the performance was 34.4% for event-based metrics and 66.4% for segment-based metrics.
Guided Learning Convolution System for DCASE 2019 Task 4
Liwei Lin1, Xiangdong Wang1, Hong Liu1, and Yueliang Qian1
1Institute of Computing Technology, Chinese Academy of Sciences
In this paper, we describe in detail the system we submitted to DCASE2019 task 4: sound event detection (SED) in domestic environments. We approach SED as a multiple instance learning (MIL) problem and employ a convolutional neural network (CNN) with class-wise attention pooling (cATP) module to solve it. By considering the interference caused by the co-occurrence of multiple events in the unbalanced dataset, we combine the cATP-MIL framework with the Disentangled Feature. To take advantage of the unlabeled data, we adopt Guided Learning for semi-supervised learning. A group of median filters with adaptive window sizes is utilized in post-processing. We also analyze the effect of the synthetic data on the performance of the model and finally achieve an event-based F-measure of 45.43% on the validation set and an event-based F-measure of 42.7% on the test set. The system we submitted to the challenge achieves the best performance compared to those of other participants.
Cites: 3 ( see at Google Scholar )
Crowdsourcing a Dataset of Audio Captions
Samuel Lipping1, Konstantinos Drossos1, and Tuomas Virtanen1
Audio captioning is a novel field of multi-modal translation and it is the task of creating a textual description of the content of an audio signal (e.g. "people talking in a big room"'). The creation of a dataset for this task requires a considerable amount of work, rendering the crowdsourcing a very attractive option. In this paper we present a three steps based framework for crowdsourcing an audio captioning dataset, based on concepts and practises followed for the creation of widely used image captioning and machine translations datasets. During the first step initial captions are gathered. A grammatically corrected and/or rephrased version of each initial caption is obtained in second step. Finally, the initial and edited captions are rated, keeping the top ones for the produced dataset. We objectively evaluate the impact of our framework during the process of creating an audio captioning dataset, in terms of diversity and amount of typographical errors in the obtained captions. The obtained results show that the resulting dataset has less typographical errors than the initial captions, and on average each sound in the produced dataset has captions with a Jaccard similarity of 0.24, roughly equivalent to two ten-word captions having in common four words with the same root, indicating that the captions are dissimilar while they still contain some of the same information.
Cites: 1 ( see at Google Scholar )
Long-distance Detection of Bioacoustic Events with Per-channel Energy Normalization
Vincent Lostanlen1, Kaitlin Palmer2, Elly Knight3, Christopher Clark1, Holger Klinck1, Andrew Farnsworth1, Tina Wong4, Jason Cramer4, and Juan Bello4
1Cornell Lab of Ornithology, 2San Diego State University, 3University of Alberta, 4New York University
This paper proposes to perform unsupervised detection of bioacoustic events by pooling the magnitudes of spectrogram frames after per-channel energy normalization (PCEN). Although PCEN was originally developed for speech recognition, it also has beneficial effects in enhancing animal vocalizations, despite the presence of atmospheric absorption and intermittent noise. We prove that PCEN generalizes logarithm-based spectral flux, yet with a tunable time scale for background noise estimation. In comparison with pointwise logarithm, PCEN reduces false alarm rate by 50x in the near field and 5x in the far field, both on avian and marine bioacoustic datasets. Such improvements come at moderate computational cost and require no human intervention, thus heralding a promising future for PCEN in bioacoustics.
Acoustic Scene Classification from Binaural Signals using Convolutional Neural Networks
Rohith Mars1, Pranay Pratik1, Srikanth Nagisetty2, and Chongsoon Lim3
1Panasonic R&D Center Singapore, 2Panasonic R&D Center SIngapore, 3Panasonic Singapore Laboratory
In this paper, we present the details of our proposed framework and solution for the DCASE 2019 Task 1A - Acoustic Scene Classification challenge. We describe the audio pre-processing, feature extraction steps and the time-frequency (TF) representations employed for acoustic scene classification using binaural recordings. We propose two distinct and light-weight architectures of convolutional neural networks (CNNs) for processing the extracted audio features and classification. The performance of both these architectures are compared in terms of classification accuracy as well as model complexity. Using an ensemble of the predictions from the subset of models based on the above CNNs, we achieved an average classification accuracy of 79.35% on the test split of the development dataset for this task. In the Kaggle’s private leaderboard, oursolution was ranked 4th with a system score of 83.16% — an improvement of ≈ 20% over the baseline system.
First Order Ambisonics Domain Spatial Augmentation for DNN-based Direction of Arrival Estimation
Luca Mazzon1, Yuma Koizumi2, Masahiro Yasuda2, and Noboru Harada2
1University of Padova, 2NTT
In this paper, we propose a novel data augmentation method for training neural networks for Direction of Arrival (DOA) estimation. This method focuses on expanding the representation of the DOA subspace of a dataset. Given some input data, it applies a transformation to it in order to change its DOA information and simulate new potentially unseen one. Such transformation, in general, is a combination of a rotation and a reflection. It is possible to apply such transformation due to a well-known property of FirstOrder Ambisonics (FOA). The same transformation is applied also to the labels, in order to maintain consistency between input data and target labels. Three methods with different level of generality are proposed for applying this augmentation principle. Experiments are conducted on two different DOA networks. Results of both experiments demonstrate the effectiveness of the novel augmentation strategy by improving the DOA error by around 40%.
The Impact of Missing Labels and Overlapping Sound Events on Multi-label Multi-instance Learning for Sound Event Classification
Maarten Meire1, Peter Karsmakers2, and Lode Vuegen2
1KULeuven, 2KU Leuven
Automated analysis of complex scenes of everyday sounds might help us navigate within the enormous amount of data and help us make better decisions based on the sounds around us. For this purpose classification models are required that translate raw audio to meaningful event labels. The specific task that this paper targets is that of learning sound event classifier models by a set of example sound segments that contain multiple potentially overlapping sound events and that are labeled with multiple weak sound event class names. This involves a combination of both multi-label and multi-instance learning. This paper investigates two state-of-theart methodologies that allow this type of learning, LRM-NMD and CNN. Besides comparing the accuracy in terms of correct sound event classifications, also the robustness to missing labels and to overlap of the sound events in the sound segments is evaluated. For small training set sizes LRM-NMD clearly outperforms CNN with an accuracy that is 40 to 50% higher. LRM-NMD does only minorly suffer from overlapping sound events during training while CNN suffers a substantial drop in classification accuracy, in the order of 10 to 20%, when sound events have a 100% overlap. Both methods show good robustness to missing labels. No matter how many labels are missing in a single segment (that contains multiple sound events) CNN converges to 97% accuracy when enough training data is available. LRM-NMD on the other hand shows a slight performance drop when the amount of missing labels increases.
Acoustic Scene Classification in DCASE 2019 Challenge: Closed and Open Set Classification and Data Mismatch Setups
Annamaria Mesaros1, Toni Heittola1, and Tuomas Virtanen1
Acoustic Scene Classification is a regular task in the DCASE Challenge, with each edition having it as a task. Throughout the years, modifications to the task have included mostly changing the dataset and increasing its size, but recently also more realistic setups have been introduced. In DCASE 2019 Challenge, the Acoustic Scene Classification task includes three subtasks: Subtask A, a closed-set typical supervised classification where all data is recorded with the same device; Subtask B, a closed-set classification setup with mismatched recording devices between training and evaluation data, and Subtask C, an open-set classification setup in which evaluation data could contain acoustic scenes not encountered in the training. In all subtasks, the provided baseline system was significantly outperformed, with top performance being 85.2% for Subtask A, 75.5% for Subtask B, and 67.4% for Subtask C. This paper presents the outcome of DCASE 2019 Challenge Task 1 in terms of submitted systems performance and analysis.
Cites: 2 ( see at Google Scholar )
OtoMechanic: Auditory Automobile Diagnostics via Query-by-Example
Max Morrison1, and Bryan Pardo1
Early detection and repair of failing components in automobiles reduces the risk of vehicle failure in life-threatening situations. Many automobile components in need of repair produce characteristic sounds. For example, loose drive belts emit a high-pitched squeaking sound, and bad starter motors have a characteristic whirring or clicking noise. Often drivers can tell that the sound of their car is not normal, but may not be able to identify the cause. To mitigate this knowledge gap, we have developed OtoMechanic, a web application to detect and diagnose vehicle component issues from their corresponding sounds. It compares a user's recording of a problematic sound to a database of annotated sounds caused by failing automobile components. OtoMechanic returns the most similar sounds, and provides weblinks for more information on the diagnosis associated with each sound, along with an estimate of the similarity of each retrieved sound. In user studies, we find that OtoMechanic significantly increases diagnostic accuracy relative to a baseline accuracy of consumer performance.
Hierarchical Sound Event Classification
Eric Nichols1, Daniel Tompkins2, and Jianyu Fan3
1Microsoft Corporation, 2Micro, 3Simon Fraser University
Task 5 of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2019 challenge is "urban sound tagging''. Given a set of known sound categories and sub-categories, the goal is to build a multi-label audio classification model to predict whether each sound category is present or absent in an audio recording. We developed a model composed of a preprocessing layer that converts audio to a log-mel spectrogram, a VGG-inspired Convolutional Neural Network (CNN) that generates an embedding for the spectrogram, a pre-trained VGGish network that generates a separate audio embedding, and finally a series of fully-connected layers that converts these two embeddings (concatenated) into a multi-label classification. This model directly outputs both “fine” and “coarse” labels; it treats the task as a 37-way multi-label classification problem. One version of this network did better at the coarse labels (CNN+VGGish1); another did better with fine labels on Micro AUPRC (CNN+VGGish2). A separate family of CNN models was also trained to take into account the hierarchical nature of the labels (Hierarchical1, Hierarchical2, and Hierarchical3). The hierarchical models perform better on Micro AUPRC with fine-level classification.
Onsets, Activity, and Events: A Multi-task Approach for Polyphonic Sound Event Modelling
Arjun Pankajakshan1, Helen Bear1, and Emmanouil Benetos1
1Queen Mary University of London
State of the art polyphonic sound event detection (SED) systems function as frame-level multi-label classification models. In the context of dynamic polyphony levels at each frame, sound events interfere with each other which degrade a classifier’s ability to learn the exact frequency profile of individual sound events. Frame-level localized classifiers also fail to explicitly model the long-term temporal structure of sound events. Consequently, the event-wise detection performance is less than the segment-wise detection. We define ‘temporally precise polyphonic sound event detection’ as the subtask of detecting sound event instances with the correct onset. Here, we investigate the effectiveness of sound activity detection (SAD) and onset detection as auxiliary tasks to improve temporal precision in polyphonic SED using multi-task learning. SAD helps to differentiate event activity frames from noisy and silence frames and helps to avoid missed detections at each frame. Onset predictions ensure the start of each event which in turn are used to condition predictions of both SAD and SED. Our experiments on the URBAN-SED dataset show that by conditioning SED with onset detection and SAD, there is over a three-fold relative improvement in event-based F -score.
TrellisNet-Based Architecture for Sound Event Localization and Detection with Reassembly Learning
This paper proposes a deep learning technique and network model for DCASE 2019 task 3: Sound Event Localization and Detection. Currently, the convolutional recurrent neural network is known as the state-of-the-art technique for sound classification and detection. We focus on proposing TrellisNet-based architecture that can replace the convolutional recurrent neural network. Our TrellisNet-based architecture has better performance in the direction of arrival estimation compared to the convolutional recurrent neural network. We also propose reassembly learning to design a single network that handles dependent sub-tasks together. Reassembly learning is a method to divide multi-task into individual sub-tasks, to train each sub-task, then reassemble and fine-tune them into a single network. Experimental results show that the proposed method improves sound event localization and detection performance compared to the DCASE 2019 baseline system.
Weakly Labeled Sound Event Detection using Tri-training and Adversarial Learning
Hyoungwoo Park1, Sungrack Yun2, Jungyun Eum3, Janghoon Cho3, and Kyuwoong Hwang3
1Qualcomm AI research, 2Qualcomm korea, 3Qualcomm AI Research
This paper considers a semi-supervised learning framework for weakly labeled polyphonic sound event detection problems for the DCASE 2019 challenge's task4 by combining both the tri-training and adversarial learning. The goal of the task4 is to detect onsets and offsets of multiple sound events in a single audio clip. The entire dataset consists of the synthetic data with a strong label (sound event labels with boundaries) and real data with weakly labeled (sound event labels) and unlabeled dataset. Given this dataset, we apply the tri-training where two different classifiers are used to obtain pseudo labels on the weakly labeled and unlabeled dataset, and the final classifier is trained using the strongly labeled dataset and weakly/unlabeled dataset with pseudo labels. Also, we apply the adversarial learning to reduce the domain gap between the real and synthetic dataset. We evaluated our learning framework using the validation set of the task4 dataset, and in the experiments, our learning framework shows a considerable performance improvement over the baseline model.
A Hybrid Parametric-Deep Learning Approach for Sound Event Localization and Detection
Andres Perez-Lopez1, Eduardo Fonseca2, and Xavier Serra3
1UPF - Eurecat, 2Music Technology Group, Universitat Pompeu Fabra, 3Universitat Pompeu Fabra
This work describes and discusses an algorithm submitted to the Sound Event Localization and Detection Task of DCASE2019 Challenge. The proposed methodology relies on parametric spatial audio analysis for source localization and detection, combined with a deep learning-based monophonic event classifier. The evaluation of the proposed algorithm yields overall results comparable to the baseline system. The main highlight is a reduction of the localization error on the evaluation dataset by a factor of 2.6, compared with the baseline performance.
Classifying Non-speech Vocals: Deep vs Signal Processing Representations
Fatemeh Pishdadian1, Prem Seetharaman1, Bongjun Kim1, and Bryan Pardo1
Deep-learning-based audio processing algorithms have become very popular over the past decade. Due to promising results reported for deep-learning-based methods on many tasks, some now argue that signal processing audio representations (e.g. magnitude spectrograms) should be entirely discarded, in favor of learning representations from data using deep networks. In this paper, we compare the effectiveness of representations output by state-of-the-art deep nets trained for a task-specific problem, to off-the-shelf signal processing encoding. We address two tasks: query by vocal imitation and singing technique classification. For query by vocal imitation, experimental results showed deep representations were dominated by signal-processing representations. For singing technique classification, neither approach was clearly dominant. These results indicate it would be premature to abandon traditional signal processing in favor of exclusively using deep networks.
Sound Event Localization and Detection using CRNN Architecture with Mixup for Model Generalization
Pranay Pratik1, Wen Jie Jee2, Srikanth Nagisetty3, Rohith Mars3, and Chongsoon Lim4
1Panasonic R&D Center Singapore, 2Nanyang Technological University, 3Panasonic R&D Center SIngapore, 4Panasonic Singapore Laboratory
In this paper, we present the details of our solution for the IEEE DCASE 2019 Task 3: Sound Event Localization and Detection (SELD) challenge. Given multi-channel audio as input, goal is to predict all instances of the sound labels and their directions-of-arrival (DOAs) in the form of azimuth and elevation angles. Our solution is based on Convolutional-Recurrent Neural Network (CRNN) architecture. In the CNN module of the proposed architecture, we introduced rectangular kernels in the pooling layers to minimize the information loss in temporal dimension within the CNN module, leading to boosting up the RNN module performance. Data augmentation mixup is applied in an attempt to train the network for greater generalization. The performance of the proposed architecture was evaluated with individual metrics, for sound event detection (SED) and localization task. Our team’s solution was ranked 5th in the DCASE-2019 Task-3 challenge with an F-score of 93.7% & Error Rate 0.12 for SED task and DOA error of 4.2° & frame recall 91.8% for localization task, both on the evaluation set. This results showed a significant performance improvement for both SED and localization estimation over the baseline system.
Exploiting Parallel Audio Recordings to Enforce Device Invariance in CNN-based Acoustic Scene Classification
Paul Primus1, Hamid Eghbal-zadeh2, David Eitelsebner1, Khaled Koutini3, Andreas Arzt3, and Gerhard Widmer3
1JKU, 2LIT AI Lab & Johannes Kepler University, Institute of Computational Perception, 3Johannes Kepler University
Distribution mismatches between the data seen at training and at application time remain a major challenge in all application areas of machine learning. We study this problem in the context of machine listening (Task 1b of the DCASE 2019 Challenge). We propose a novel approach to learn domain-invariant classifiers in an end-to-end fashion by enforcing equal hidden layer representations for domain-parallel samples, i.e. time-aligned recordings from different recording devices. No classification labels are needed for our domain adaptation (DA) method, which makes the data collection process cheaper. We show that our method improves the target domain accuracy for both a toy dataset and an urban acoustic scenes dataset. We further compare our method to Maximum Mean Discrepancy-based DA and find it more robust to the choice of DA parameters. Our submission, based on this method, to DCASE 2019 Task 1b gave us the 4th place in the team ranking.
MIMII Dataset: Sound Dataset for Malfunctioning Industrial Machine Investigation and Inspection
Harsh Purohit1, Ryo Tanabe2, Takeshi Ichige3, Takashi Endo3, Yuki Nikaido3, Kaori Suefusa3, and Yohei Kawaguchi3
1Hitachi Ltd., 2 Hitachi, Ltd., 3Hitachi, Ltd.
Factory machinery is prone to failure or breakdown, resulting in significant expenses for companies. Hence, there is a rising interest in machine monitoring using different sensors including microphones. In scientific community, the emergence of public datasets has been promoting the advancement in acoustic detection and classification of scenes and events, but there are no public datasets that focus on the sound of industrial machines under normal and anomalous operating conditions in real factory environments. In this paper, we present a new dataset of industrial machine sounds which we call a sound dataset for malfunctioning industrial machine investigation and inspection (MIMII dataset). Normal and anomalous sounds were recorded for different types of industrial machines, i.e. valves, pumps, fans and slide rails. To resemble the real-life scenario, various anomalous sounds have been recorded, for instance, contamination, leakage, rotating unbalance, rail damage, etc. The purpose of releasing the MIMII dataset is to help the machine-learning and signal-processing community to advance the development of automated facility maintenance.
Sound Event Detection and Direction of Arrival Estimation using Residual Net and Recurrent Neural Networks
Rishabh Ranjan1, Sathish Jayabalan1, Thi Ngoc Tho Nguyen2, and Woon Seng Gan3
1NTU Singapore, 2Nanyang Technological University, 3NTU
This paper presents deep learning approach for sound events detection and localization, which is also a part of detection and classification of acoustic scenes and events (DCASE) challenge 2019 Task 3. Deep residual nets originally used for image classification are adapted and combined with recurrent neural networks (RNN) to estimate the onset-offset of sound events, sound events class, and their direction in a reverberant environment. Additionally, data augmentation and post processing techniques are applied to generalize and improve the system performance on unseen data. Using our best model on validation dataset, sound events detection achieves F1-score of 0.89 and error rate of 0.18, whereas sound source localization task achieves angular error of 8° and 90% frame recall.
Open-set Evolving Acoustic Scene Classification System
Fatemeh Saki1, Yinyi Guo1, Cheng-Yu Hung1, Lae-hoon Kim1, Manyu Deshpande1, Sunkuk Moon1, Eunjeong Koh2, and Erik Visser1
Most audio recognition/classification systems assume a static and closed-set model, where training and testing data are drawn from a prior distribution. However, in real-world audio recognition/classification problems, such a distribution is unknown, and training data is limited and incomplete at training time. As it is difficult to collect exhaustive train-ing samples to train classifiers. Datasets at prediction time are evolving and the trained model must deal with an infinite number of unseen/unknown categories. Therefore, it is desired to have an open-set classifier that not only accurately classifies the known classes into their respective classes but also effectively identifies unknown samples and learns them. This paper introduces an open-set evolving audio classification technique, which can effectively recognize and learn unknown classes continuously in an unsupervised manner. The proposed method consists of several steps: a) recognizing sound signals and associating them with known classes while also being able to identify the unknown classes; b) detecting the hidden unknown classes among the rejected sound samples; c) learning those novel detected classes and updating the classifier. The experimental results illustrate the effectiveness of the developed approach in detecting unknown sound classes compared to extreme value machine (EVM) and Weibull-calibrated SVM (W-SVM).
HODGEPODGE: Sound Event Detection Based on Ensemble of Semi-Supervised Learning Methods
Ziqiang Shi1, Liu Liu1, Huibin Lin1, Rujie Liu2, and Anyan Shi3
1Fujitsu Research & Development Center, 2Fujitsu Research & Development Center Co., Ltd., 3ShuangFeng First
In this paper, we present a method called HODGEPODGE for large-scale detection of sound events using weakly labeled, synthetic, and unlabeled data proposed in the Detection and Classification of Acoustic Scenes and Events (DCASE) 2019 challenge Task 4: Sound event detection in domestic environments. To perform this task, we adopted the convolutional recurrent neural networks (CRNN) as our backbone network. In order to deal with a small amount of tagged data and a large amounts of unlabeled in-domain data, we aim to focus primarily on how to apply semi-supervise learning methods efficiently to make full use of limited data. Three semi-supervised learning principles have been used in our system, including: 1) Consistency regularization applies data augmentation; 2) MixUp regularizer requiring that the predictions for a interpolation of two inputs is close to the interpolation of the prediction for each individual input; 3) MixUp regularization applies to interpolation between data augmentations. We also tried an ensemble of various models, which are trained by using different semi-supervised learning principles. Our proposed approach significantly improved the performance of the baseline, achieving the event-based f-measure of 42.0% compared to 25.8% event-based f-measure of the baseline in the provided official evaluation dataset. Our submissions ranked third among 18 teams in the task 4.
Deep Multi-view Features from Raw Audio for Acoustic Scene Classification
Arshdeep Singh1, Padmanabhan Rajan2, and Arnav Bhavsar3
1IIT MANDI, India, 2IIT mandi, 3IIT Mandi
In this paper, we propose a feature representation framework which captures features constituting different levels of abstraction for audio scene classification. A pre-trained deep convolution neural network, SoundNet, is used to extract the features from various intermediate layers corresponding to an audio file. We consider that the features obtained from various intermediate layers provide the different types of abstraction and exhibits complementary information. Thus, combining the intermediate features of various layers can improve the classification performance to discriminate audio scenes. To obtain the representations, we ignore redundant filters in the intermediate layers using analysis of variance based redundancy removal framework. This reduces dimensionality and computational complexity. Next, shift-invariant fixed-length compressed representations across layers are obtained by aggregating the responses of the important filters only. The obtained compressed representations are stacked altogether to obtain a supervector. Finally, we employ the classification using multi-layer perceptron and support vector machine models. We comprehensively perform the validation of the above assumption on two public datasets; Making Sense of Sounds and open set acoustic scene classification DCASE 2019.
Audio Tagging using Linear Noise Modelling Layer
Shubhr Singh1, Arjun Pankajakshan2, and Emmanouil Benetos2
1queen mary university of london, 2Queen Mary University of London
Label noise refers to the presence of inaccurate target labels in a dataset. It is an impediment to the performance of a deep neural network (DNN) as the network tends to overﬁt to the label noise, hence it becomes imperative to devise a generic methodology to counter the effects of label noise. FSDnoisy18k is an audio dataset collected with the aim of encouraging research on label noise for sound event classiﬁcation. The dataset contains ∼42.5 hours of audio recordings divided across 20 classes, with a small amount of manually veriﬁed labels and a large amount of noisy data. Using this dataset, our work intends to explore the potential of modelling the label noise distribution by adding a linear layer on top of a baseline network. The accuracy of the approach is compared to an alternative approach of adopting a noise robust loss function. Results show that modelling the noise distribution improves the accuracy of the baseline network in a similar capacity to the soft bootstrapping loss.
Robustness of Adversarial Attacks in Sound Event Classification
Vinod Subramanian1, Emmanouil Benetos1, and Mark B. Sandler1
1Queen Mary University of London
An adversarial attack is a method to generate perturbations to the input of a machine learning model in order to make the output of the model incorrect. The perturbed inputs are known as adversarial examples. In this paper, we investigate the robustness of adversarial examples to simple input transformations such as mp3 compression, resampling, white noise and reverb in the task of sound event classification. By performing this analysis, we aim to provide insight on strengths and weaknesses in current adversarial attack algorithms as well as provide a baseline for defenses against adversarial attacks. Our work shows that adversarial attacks are not robust to simple input transformations. White noise is the most consistent method to defend against adversarial attacks with a success rate of 73.72% averaged across all models and attack algorithms.
Improvement of DOA Estimation by using Quaternion Output in Sound Event Localization and Detection
Yui Sudo1, Katsutoshi Itoyama1, Kenji Nishida1, and Kazuhiro Nakadai2
1Tokyo Institute of Technology, 2 Honda Research Institute Japan
This paper describes improvement of Direction of Arrival (DOA) estimation performance using quaternion output in the Detection and Classification of Acoustic Scenes and Events (DCASE) 2019 Task 3. DCASE 2019 Task3 focuses on the sound event localization and detection (SELD) which is a task that simultaneously estimates the sound source direction in addition to conventional sound event detection (SED). In the baseline method, the sound source direction angle is directly regressed. However, the angle is a periodic function and it has discontinuities which may make learning unstable. Specifically, even though -180 deg and 180 deg are in the same direc-tion, a large loss is calculated. Estimating DOA angles with a classification approach instead of regression can solve such instability of discontinuities but this causes limitation of reso-lution. In this paper, we propose to introduce the quaternion which is a continuous function into the output layer of the neural network instead of directly estimating the sound source direction angle. This method can be easily implemented only by changing the output of the existing neural network, and thus does not significantly increase the number of parameters in the middle layers. Experimental results show that proposed method improves the DOA estimation without significantly increasing the number of parameters.
Sound Event Detection in Domestic Environments with Weakly Labeled Data and Soundscape Synthesis
Nicolas Turpault1, Romain Serizel2, Justin Salamon3, and Ankit Parag Shah4
1Inria, 2Université de Lorraine, 3Adobe Research, 4Carnegie Mellon University
This paper presents Task 4 of the Detection and Classification of Acoustic Scenes and Events (DCASE) 2019 challenge and provides a first analysis of the challenge results. The task is a follow-up to Task 4 of DCASE 2018, and involves training systems for large-scale detection of sound events using a combination of weakly labeled data, i.e. training labels without time boundaries, and strongly-labeled synthesized data. The paper introduces Domestic Environment Sound Event Detection (DESED) dataset mixing a part of last year dataset and an additional synthetic, strongly labeled, dataset provided this year that we'll describe more in detail. We also report the performance of the submitted systems on the official evaluation (test) and development sets as well as several additional datasets. The best systems from this year outperform last year's winning system by about 10% points in terms of F-measure.
Open-Set Acoustic Scene Classification with Deep Convolutional Autoencoders
Kevin Wilkinghoff1, and Frank Kurth2
1Fraunhofer Institute for Communication, Information Processing and Ergonomics FKIE , 2Fraunhofer FKIE
Acoustic scene classification is the task of determining the environment in which a given audio file has been recorded. If it is a priori not known whether all possible environments that may be encountered during test time are also known when training the system, the task is referred to as open-set classification. This paper contains a description of an open-set acoustic scene classification system submitted to task 1C of the Detection and Classification of Acoustic Scenes and Events (DCASE) Challenge 2019. Our system consists of a combination of convolutional neural networks for closed-set identification and deep convolutional autoencoders for outlier detection. On the evaluation dataset of the challenge, our proposed system significantly outperforms the baseline system and improves the score from 0.476 to 0.621. Moreover, our submitted system ranked 3rd among all teams in task 1C.
MAVD: A Dataset for Sound Event Detection in Urban Environments
Pablo Zinemanas1, Pablo Cancela1, and Martín Rocamora1
1Universidad de la República
We describe the public release of a dataset for sound event detection in urban environments, namely MAVD, which is the first of a series of datasets planned within an ongoing research project for urban noise monitoring in Montevideo city, Uruguay. This release focuses on traffic noise, MAVD-traffic, as it is usually the predominant noise source in urban environments. An ontology for traffic sounds is proposed, which is the combination of a set of two taxonomies: vehicle types (e.g. car, bus) and vehicle components (e.g. engine, brakes), and a set of actions related to them (e.g. idling, accelerating). Thus, the proposed ontology allows for a flexible and detailed description of traffic sounds. We also provide a baseline of the performance of state-of-the-art sound event detection systems applied to the dataset.