Unsupervised methods for evaluating speech representations

Abstract

Unsupervised representation learning using deep generative models has produced remarkable results across many domains in recent years. These methods have been applied to speech processing to learn representations useful for downstream supervised tasks like speaker, dialect, or phoneme identification. One research path has been to develop general purpose priors that select effective representations. However, many priors on good representations are difficult to incorporate into unsupervised methods because they are difficult to evaluate without supervision. This thesis proposes to use low-level acoustic features to address this problem for speech. By using techniques in acoustic processing, we develop methods for structured evaluation for speech representations. The evaluation aims both to assess the efficacy of representations for downstream tasks and to validate claims about the priors used to construct them. An evaluation suite for benchmarking and analyzing research in speech representation learning is produced and open-sourced as a result of this thesis.