This study investigates the task-dependence of standard cepstral peak prominence (CPP) computation methods, and the advantages conferred by an open-source method of excluding unvoiced regions in CPP computation. We use Praat and a public dataset (Perceptual Voice Qualities Database, consisting of 295 speakers) to assess how well a voice-only CPP algorithm identifies voice disorders, identifies perceived dysphonia, and correlates with dysphonia severity. Results indicate that, compared to standard CPP computation, voice-only CPP is (1) less affected by unvoiced regions in the speech signal and (2) better reflects clinical outcomes (i.e., voice disorder diagnosis and dysphonia severity) for data sets that contain varying speech tasks. We expect voice-only CPP to be particularly useful for assessing speech that contains unknown or heterogeneous utterance types, as well as for speakers whose voice signal is affected by involvement of other speech subsystems (e.g., articulatory impairment).