DP-Parse: Finding Word Boundaries from Raw Speech with an Instance Lexicon

Abstract Finding word boundaries in continuous speech is challenging as there little or no equivalent of a ‘space’ delimiter between words. Popular Bayesian non-parametric models for text segmentation (Goldwater et al., 2006, 2009) use Dirichlet process to jointly segment sentences and build lexicon types. We introduce DP-Parse, which uses similar principles but only relies on an instance tokens, avoiding the clustering errors that arise with On Zero Resource Speech Benchmark 2017, our model sets new state-of-the-art 5 languages. The algorithm monotonically improves better input representations, achieving yet higher scores when fed weakly supervised inputs. Despite lacking type lexicon, DP-Parse can be pipelined language learn semantic syntactic representations assessed by spoken embedding benchmark. 1