If languages are fundamentally symbolic systems, as it is assumed in cognitive linguistics, then it follows that what speakers learn in the course of language acquisition are symbolic systems. It is not unequivocally established that linguistically untrained speakers even have spontaneously emerging awareness of meaningless entities like the phonemes (Liberman et al. 1980, Lotto and Holt 2000, Port and Leary 2005, Read et al. 1986, Välimaa-Blum in press). In this paper, I will argue (i) that knowledge of individual phonemes and their prototypes is metalinguistic, (ii) that the hyperspeech variants of words constitute their prototypes, and (iii) that these should be explicitly represented in the grammar. Lakoff (1993) reintroduces three levels into cognitive phonology - morpheme, word and utterance levels, which I interpret in an exemplar-theoretical framework as representing three kinds of knowledge that speakers must have of the phonology and morpholexicon of their language. The morpheme level contains an exemplar-based lexicon with all the non-automatic allomorphy and word formation principles, the word level articulates the hyperspeech forms of isolated words, and the utterance level spells out the stochastically varying hypospeech shapes of the same in continuous speech. Phonologies only having an abstract 'underlying' level and a phonetic surface have no place for the hyperspeech forms, which, however, are cognitively real to speakers. In cognitive views, the prototypes of phonemes, and hence of words as well, tend to be schematic (Langacker 1987, Mompeán-González 2004, Nathan 1996, 2006, 2007, Taylor 2003), and consequently they are never instantiated as such. I consider it unlikely that a purely abstract, non-instantiated sound shape be the prototype of a category, just as unlikely as it would be for frequency to establish them. If we accept the distinct word and utterance levels, we introduce a specific point in the grammar - the word level - that spells out the hyper-articulated, best exemplars of words.