This paper adopts Latent Semantic Analysis (LSA) for longterm analysis of voice quality, in particular creakiness. Each automatically labeled creaky instance (word) is modeled as a document and different prosodic and syntactic cues as terms. This framework attempts to automatically identify the most salient correlates, or latent factors, of creakiness, and further assign each creaky instance (word) to one of the latent factors. The algorithm implemented in this study identifies at least two correlates of creakiness in Switchboard: (1) particles, coordinating conjunctions in repair/repeat locations, and filled pauses; (2) starts of various sentence/clause structures, such as Whadverb phrases, sentences and asides with sentence restarts at repair/repeat locations. Such automatic long-term voice quality analysis could pave the way for better incorporating voice quality in speech recognition, among other speech applications.