Analyzing the Representational Geometry of Acoustic Word Embeddings