Learning sound event classifiers from web audio with noisy labels

Fonseca, Eduardo; Plakal, Manoj; Ellis, Daniel P. W.; Font Corbera, Frederic; Favory, Xavier; Serra, Xavier

doi:http://dx.doi.org/10.1109/ICASSP.2019.8683158

Learning sound event classifiers from web audio with noisy labels

Citation

Fonseca E, Plakal M, Ellis DPW, Font F, Favory X, Serra X. Learning sound event classifiers from web audio with noisy labels. In: 2019 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP); 2019 May 12-17; Brighton, United Kingdom. New Jersey: Institute of Electrical and Electronics Engineers; 2019. p. 21-5. DOI: 10.1109/ICASSP.2019.8683158

Abstract

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.

Description

Comunicació presentada a: ICASSP 2019 - 2019 IEEE International Conference on Acoustics, Speech and Signal Processing celebrat de 12 al 17 de maig de 2019 a Brighton, Regne Unit.