Feature Learning for Bird Call Clustering

In this paper, a supervised algorithm is proposed for the identification and segmentation of bird calls with K-means clustering using features learnt by matrix factorization. Singular value decomposition is applied on pooled time-frequency vocalization in a class-wise manner to learn a class-specific feature representation. These representations show discriminative behavior even when unseen classes are represented. By combining the proposed feature representation with K-means clustering, we are able to effectively cluster and segment bird calls from multiple species, which are present in an input recording. Experimental results are provided on a small dataset of birdsong.