[…] certain segments are descriptively equivalent, […] certain segments (which have proved to be not mutually substitutable) are descriptively non-equivalent or distinct (i.e. unsubstitutable). For the further analysis of the language, the explicit record of descriptive (or, as they are called, phonemic) distinctions is as important as that of descriptive equivalences. If we have a body of text in a language, and do not know which segments in it are equivalent to each other (e.g. whether a 'ɡ' in one line of the text is substitutable for a 'k' in another), we can do little in the way of further analysis. If we do not know which segments are distinct from each other, e.g. whether a word 'ɡam', or 'kam', in another, we still can do little analysis. When these two sets of data are explicitly given, however, it is possible to carry out the rest of the analysis. The fundamental data of descriptive linguistics are therefore the distinctions and equivalences among utterances and parts of utterances.
- Harris (1951), a pag.33

In most cases, if we find that one utterance is not equivalent to another, the distinction to which this non-equivalence is due remains no matter how often we have each utterance repeated […]. Non infrequently, however, we meet an utterance which is pronounced with different, non-equivalent, segments in different repetitions. In some of these cases, two segments appear equivalent in repetitions of one utterance but not in another utterance: [e] and [iy] seem equivalent if we get [ekƏnamiks] and [iykƏnamiks] 'economics' as repetitions of each other, but not in repetitions of 'even', 'ever', 'elemental'. Because of the latter, we consider [e] and [iy] to be distinct segments […]. In other cases, while the alternation of segments occurs freely in some utterances and not at all in others, it never constitutes the sole distinction between two distinct utterances […]. Thus in repetitions of 'The seat is loose'. we may obtain both [ðƏ 'siytiz'luws.] and [ðƏ 'siyr¹iz'luws.] (where [r¹] indicates a single alveolar flap of the tongue). In repetitions of 'meter' ['miyr¹Ər] we hardly ever get [t] instead of [r¹]. Thus a variation of segments which occurs apparently freely in one utterance hardly ever occurs in another.
- Harris (1951), a pag.39