Information Retrieval for ZeroSpeech 2021: The Submission by University of Wroclaw

We build on the In this paper we present our submission which tries to address unsupervised representations of speech proposed by the organizers all four tasks. We extend the baseline solution in several as a baseline, derived from CPC and clustered with the k-directions: we refine the intermediate representations, extracted means algorithm. We demonstrate that simple methods of refining with CPC, to directly improve the ABX scores. We show that those representations can narrow the gap, or even improve such representations can be used to perform simple fuzzy lookups upon the solutions which use a high computational budget. The in a large dataset, and even extract some common patterns results lead to the conclusion that the CPC-derived representations that serve as pseudo-words. Our approach to the semantic word are still too noisy for training language models, but stable similarity task is also based on pseudo-words.