Pronunciation variation is extremely widespread and one of the reasons for recognition errors. In this paper we explore how similar-recognized-words can be used to construct or expand more accurate grammars in a specific domain. The domain that serves as framework for this research is the assessment of depression. Assessment of depression is done via a system that verbally administers a discrete choice questionnaire over the telephone. Several experiments carried out have shown that in spite of the limited number of valid responses, responses uttered by the same speaker may be both correctly or incorrectly recognized. Analysis of the responses incorrectly recognized has provided the elements to formulate new grammar rules. To test the grammars thus expanded a new set of experiments was completed and the results obtained are presented and discussed.