Audio tagging with noisy labels and minimal supervision

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

103 citations recorded by the source corpus.

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 in 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.

Subfields: Artificial Intelligence, 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 Artificial Intelligence

In Signal Processing

Citations in the corpus

FSD50K: An Open Dataset of Human-Labeled Sound Events (2020, methodology): “As we have seen above, several datasets containing Freesound audio have been widely used by the research community [22, 32, 33, 39, 54, 55], showing its usefulness for dataset creation.”
The Benefit of Temporally-Strong Labels in Audio Event Classification (2021, background): “lωlrap [22] indicates the precision of lists of ranked labels obtained for each test clip, and thus, unlike d ′, reflects the relative calibration of score values from different class outputs.”
Model-Agnostic Approaches To Handling Noisy Labels When Training Sound Event Classifiers (2019, background): “Recently, the topic of label noise was included for the first time as one of the research problems in the DCASE2019 Challenge [17].”
Convolutional Recurrent Neural Network and Data Augmentation for Audio Tagging with Noisy Labels and Minimal Supervision (2019, methodology): “In this contribution we tackle the DCASE 2019 Challenge Task 2 [10], where the main research objective is the following: How can we train a high performance Audio Tagging system given relatively little data with reliable labels but a larger amount of mis-matched data with noisy labels?”
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.”
Unsupervised Contrastive Learning of Sound Event Representations (2020, background): “…websites is typically accompanied by metadata, it is usually too sparse for a meaningful mapping to a sound event label set (in the case of Flickr [8]), or sometimes the user-provided meta-data can be an underrepresentation of the actual acoustic content (in the case of a portion of Freesound).”
Audio Retrieval with Natural Language Queries (2021, unknown): “These include TUT Acoustic scenes [14], CHIME-Home [15], ESC-50 [16], FSDKaggle [17], and AudioSet [10].”
SONYC-UST-V2: An Urban Sound Tagging Dataset with Spatiotemporal Context (2020, methodology): “For each of the two levels, we compute three metrics: macroaveraged AUPRC, micro-averaged AUPRC, and label-weighted label-ranking average precision (LWLRAP) [15].”