Sound event detection from weak annotations: weighted-GRU versus multi-instance-learning
Résumé
In this paper, we address the detection of audio events in domestic environments in the case where a weakly annotated dataset is available for training. The weak annotations provide tags from audio events but do not provide temporal boundaries. We report experiments in the framework of task four of the DCASE 2018 challenge. The objective is twofold: detect audio events (multi-categorical classification at recording level), localize the events precisely within the recordings. We explored two approaches: 1) a ”weighted-GRU” (WGRU), in which we train a Convolutional Recurrent Neural Network (CRNN) for classification and then exploit its frame-based predictions at the output of the time-distributed dense layer to perform localization. We propose to lower the influence of the hidden states to avoid predicting the same score throughout a recording. 2) An approach inspired by Multi-Instance Learning (MIL), in which we train a CRNN to give predictions at frame-level, using a custom loss function based on the weak label and statistics of the frame-based predictions. Both approaches outperform the baseline of 14.06% in F-measure by a large margin, with values of respectively 16.77% and 24.58% for combined WGRUs and MIL, on a test set comprised of 288 recordings.
Origine | Fichiers produits par l'(les) auteur(s) |
---|