Transfer learning of weakly labelled audio
Aleksandr Diment, Tuomas I. Virtanen · 2017
Many machine learning tasks have been shown solvable with impressive levels of success given large amounts of training data and computational power. For the problems which lack data sufficient to achieve high performance, methods for transfer learning can be applied. These refer to performing the new task while having prior knowledge of the nature of the data, gained by first performing a different task, for which training data is abundant. Shown successful for other machine learning tasks, transfer learning is now investigated in audio analysis. We propose to solve the weakly labelled problem of sound event tagging with small amounts of training data by transferring the abstract knowledge about the nature of audio data from another tagging task. The proposed methods constitute pre-training of a recurrent neural network or its parts to perform one tagging task given abundant and diverse training data, and then using it or its parts for a new task of tagging sound events of different nature, for which the data is limited. Several architectures for such transfer are proposed and evaluated, showing impressive classification accuracy of 83.4% with gains of up to 20 percentage points over the baseline given as little as 36 training samples for the target task.