Learning Sound Event Classifiers from Web Audio with Noisy Labels

By Eduardo Fonseca, Manoj Plakal, D. Ellis, Frederic Font, Xavier Favory, Xavier Serra. 2019

126 citations recorded by the source corpus.

As sound event classification moves towards larger datasets, issues of label noise become inevitable. Web sites can supply large volumes of user-contributed audio and metadata, but inferring labels from this metadata introduces errors due to unreliable inputs, and limitations in the mapping. There is, however, little research into the impact of these errors. To foster the investigation of label noise in sound event classification we present FSDnoisy18k, a dataset containing 42.5 hours of audio across 20 sound classes, including a small amount of manually-labeled data and a larger quantity of real-world noisy data. We characterize the label noise empirically, and provide a CNN baseline system. Experiments suggest that training with large amounts of noisy data can outperform training with smaller amounts of carefully-labeled data. We also show that noise-robust loss functions can be effective in improving performance in presence of corrupted labels.

Subfields: Computer Vision and Pattern Recognition, Signal Processing

How researchers see this paper

The researchers whose trusted networks rank it highest, strongest advocate first. Every researcher gets a different order.

In Computer Vision and Pattern Recognition

In Signal Processing

Citations in the corpus

FSD50K: An Open Dataset of Human-Labeled Sound Events (2020, methodology): “In particular, signiﬁcant progress is being made in the development of sound event recognizers with noisy supervision [32, 42] or self-supervision [49]–[51].”
Convolutional Recurrent Neural Network and Data Augmentation for Audio Tagging with Noisy Labels and Minimal Supervision (2019, background): “Until now there are only few works addressing the impact of noisy labels for sound recognition [5].”
Mt-Gcn For Multi-Label Audio Tagging With Noisy Labels (2020, methodology): “In our experiments, we used the FSDKaggle2019 Dataset introduced in [8] for the Task 2 of DCASE 2019 Challenge.”
The Impact of Missing Labels and Overlapping Sound Events on Multi-label Multi-instance Learning for Sound Event Classification (2019, background): “While research into noisy labels has recently been growing [9, 10], the impact of missing labels specifically has not yet been investigated to the best of our knowledge.”
Image Classification with Deep Learning in the Presence of Noisy Labels: A Survey (2019, unknown): “[29, 30, 31], sound classification [185] and more.”
Learning With Out-of-Distribution Data for Audio Classification (2020, methodology): “In order to utilise much larger datasets, there has been interest in learning methods that do not rely on clean data [2, 3, 4, 5].”
SeCoST: Sequential Co-Supervision for Weakly Labeled Audio Event Detection (2019, background): “Lastly, one can observe that noisy labels are also of concern in weak label learning [15], [16].”
Unsupervised Contrastive Learning of Sound Event Representations (2020, unknown): “We use the FSDnoisy18k dataset [21] 23, 24, 25, 21, 26]).”
Universal Speech Enhancement with Score-based Diffusion (2022, methodology): “To validate our model, we use 1.5 h of clean utterances sampled from VCTK (Yamagishi et al., 2019) and Harvard sentences (Henter et al., 2014), together with noises/backgrounds from DEMAND (Thiemann et al., 2013) and FSDnoisy18k (Fonseca et al., 2019).”
Self-Supervised Learning from Automatically Separated Sound Scenes (2021, background): “For example, label noise issues such as missing, incorrect or inconsistent labels are inherent to this process [1, 2, 3, 4].”
Improving Sound Event Classification by Increasing Shift Invariance in Convolutional Neural Networks (2021, methodology): “Convolutional Neural Networks (CNNs) have been one of the cornerstones of Sound Event Classiﬁcation or Tagging (SET) in recent years [1, 2, 3, 4].”
The Benefit of Temporally-Strong Labels in Audio Event Classification (2021, background): “A performance boost is reported by adding a small amount of strongly labeled data to the weak labels, possibly due to the mitigation of the inherent problems in weakly labeled data [13, 14, 15].”
Audio tagging with noisy labels and minimal supervision (2019, background): “Hence, the label noise can vary widely in amount and type depending on the class, including in- and out-of-vocabulary noises [4].”
Model-Agnostic Approaches To Handling Noisy Labels When Training Sound Event Classifiers (2019, methodology): “Nevertheless, MAE has been reported to cause other difficulties in training that lead to performance drop [5, 14].”
A Hybrid Parametric-Deep Learning Approach for Sound Event Localization and Detection (2019, methodology): “It seems appropriate for this task since the front-end outcome can present leakage due to overlapping sources, effectively mixing two sources while only one training label is available, which can be understood as a form of label noise [19].”