Documenting languages helps to prevent the extinction of endangered dialects,\nmany of which are otherwise expected to disappear by the end of the century.\nWhen documenting oral languages, unsupervised word segmentation (UWS) from\nspeech is a useful, yet challenging, task. It consists in producing time-stamps\nfor slicing utterances into smaller segments corresponding to words, being\nperformed from phonetic transcriptions, or in the absence of these, from the\noutput of unsupervised speech discretization models. These discretization\nmodels are trained using raw speech only, producing discrete speech units that\ncan be applied for downstream (text-based) tasks. In this paper we compare five\nof these models: three Bayesian and two neural approaches, with regards to the\nexploitability of the produced units for UWS. For the UWS task, we experiment\nwith two models, using as our target language the Mboshi (Bantu C25), an\nunwritten language from Congo-Brazzaville. Additionally, we report results for\nFinnish, Hungarian, Romanian and Russian in equally low-resource settings,\nusing only 4 hours of speech. Our results suggest that neural models for speech\ndiscretization are difficult to exploit in our setting, and that it might be\nnecessary to adapt them to limit sequence length. We obtain our best UWS\nresults by using Bayesian models that produce high quality, yet compressed,\ndiscrete representations of the input speech signal.\n