In this paper, I investigate the notion of speaker identifiability, a term that is strongly associated with the pragmatic approach to specificity. Following Haspelmath 1997, I provide evidence from Russian for the linguistic relevance of speaker identifiability. In particular, I discuss two series of existential indefinites, koe- items and -to items, which are inherently specified as identifiable or not identifiable to the speaker. This specification is shown to be independent of such phenomena as the free-choice effect or narrow scope relative to another operator in the logical form of the sentence. I propose a formal analysis of speaker identifiability formulated within the framework of possible-world semantics. According to this account, an NP is speaker-identifiable if and only if it picks up the same individual in every possible world that is compatible with the speaker's worldview. Speaker identifiability is analyzed as a condition on the relative scope of an existential operator that ranges over individuals and a universal quantifier which quantifies over a set of possible worlds introduced by the context. I also argue that the speaker (non-)identifiability meaning component contributed by the investigated items constitutes a conventional implicature.