Data Cleaning of Sound Data with Label Noise Using Self Organizing Map

The noise label of data is a problem that can cause low performance of deep learning. It is difficult to manually relabel due to huge amounts of data. In addition, there are much more problems due to the similarity of sounds that are difficult to manually distinguish and label sound data. We proposed a data cleaning method using SOM (Self-Organizing Map), one of the unsupervised learning methods. In order to extract compact features from audio, densely connected layer with log scaled Mel-spectrogram is used. Data selection is performed based on the Euclidean distance of each Best matching unit (BMU) derived through the SOM. We also experiment with various grid sizes for SOM to find an efficient grid size. In addition, an appropriate distance finding experiment is conducted. This method is evaluated in sound classification using a pre-trained DenseNet model.