In this work, we propose a hierarchical subspace model for acoustic unit\ndiscovery. In this approach, we frame the task as one of learning embeddings on\na low-dimensional phonetic subspace, and simultaneously specify the subspace\nitself as an embedding on a hyper-subspace. We train the hyper-subspace on a\nset of transcribed languages and transfer it to the target language. In the\ntarget language, we infer both the language and unit embeddings in an\nunsupervised manner, and in so doing, we simultaneously learn a subspace of\nunits specific to that language and the units that dwell on it. We conduct our\nexperiments on TIMIT and two low-resource languages: Mboshi and Yoruba. Results\nshow that our model outperforms major acoustic unit discovery techniques, both\nin terms of clustering quality and segmentation accuracy.\n