Binary Attribute Embeddings for Zero-Shot Sound Event Classification

In this paper, we introduce a zero-shot learning method for sound event classification. The proposed method uses a semantic embedding of each sound event class and measures the compatibility between the semantic embedding and the input audio feature embedding. For semantic embedding, we newly define attribute vector that explains several attribute information of a sound event class, such as pitch, length, material of the sound source, etc. In the experiments, the proposed method showed higher accuracy than a conventional method using word embedding as the semantic embedding.