FixMatch on Audio data
A re-implementation and extension of the FixMatch approach for semi-supervised learning
We study the semi supervised approach FixMatch applied on audio data. Our experiments consist of reproducing the results of the original paper on three data sets using different amounts of labeled data points, and audio classification tasks on two audio data sets, UrbanSound8K (UBS8K) and Google Speech Commands data set (GSC). The audio data is converted to Log Mel Spectrograms in our experiments and are treated as images in FixMatch. The results show that FixMatch is an approach that results in good classification performance despite only using few labeled examples. Our experiments on audio data shows that a neural network benefits from training using the FixMatch approach on the GSC data set, but they do not show the same impact in performance when training on the UBS8K data set.