Word Segmentation on Discovered Phone Units With Dynamic Programming and Self-Supervised Scoring

Recent work on unsupervised speech segmentation has used self-supervised models with phone and word modules that are trained jointly. This paper instead revisits an older approach to segmentation: bottom-up phone-like unit discovery is performed first, symbolic then top of the discovered units (without influencing lower level). To do this, I propose a new model, chain two segment speech. Both use dynamic programming minimize costs from network additional duration penalty encourages longer units. Concretely, for acoustic discovery, duration-penalized (DPDP) contrastive predictive coding model as scoring network. For segmentation, DPDP applied autoencoding recurrent neural The chained in order gives comparable results state-of-the-art joint English benchmark. On French, Mandarin, German Wolof data, it outperforms previous systems ZeroSpeech benchmarks. Analysis shows system segments shorter filler words well, but might require some external top-down signal.