Training General-Purpose Audio Tagging Networks with Noisy Labels and Iterative Self-Verification.

This paper describes our submission to the first Freesound generalpurpose audio tagging challenge carried out within the DCASE 2018 challenge. Our proposal is based on a fully convolutional neural network that predicts one out of 41 possible audio class labels when given an audio spectrogram excerpt as an input. What makes this classification dataset and the task in general special, is the fact that only 3,700 of the 9,500 provided training examples are delivered with manually verified ground truth labels. The remaining non-verified observations are expected to contain a substantial amount of label noise (up to 30-35% in the “worst” categories). We propose to address this issue by a simple, iterative self-verification process, which gradually shifts unverified labels into the verified, trusted training set. The decision criterion for self-verifying a training example is the prediction consensus of a previous snapshot of the network on multiple short sliding window excerpts of the training example at hand. On the unseen test data, an ensemble of three networks trained with this self-verification approach achieves a mean average precision (MAP@3) of 0.951. This is the second best out of 558 submissions to the corresponding Kaggle challenge.