Criteria for Contrast: the Phoneme System
The construction of rules stating allophonic distributions is based on the identification of phonemes. This relies on two fundamental tools: predictability of occurrence and invariance of meaning. If two sounds occur in non-overlapping, predictable sets of contexts, and if substituting one for the other does not make a semantic difference, then those two sounds must necessarily be allophones of a single phoneme. This diagnosis is confirmed by the commutation test, which involves putting different sounds in a particular context to see if minimal pairs result. An example for English consonants is the context of -at ( pat /p/ - bat /b/ - mat /m/ - fat /f/ - vat /v/....). Accidental gaps in the English vocabulary mean that no lexical item (*jat, or "*lat, or *dat) is available. However, minimal pairs can be found in slightly different contexts to establish them as as consonant phonemes of English. Considering a range of contexts provides evidence for all the consonant phonemes of English.
Minimal pairs and the commutation test can generally suffice to establish the members of a phoneme system. However, there are some circumstances where phonemes cannot be established by minimal pairs alone, and we need supplementary criteria for phonemicisation. For example, /h/ contrasts with a number of English consonant phonemes word-initially; but there is no minimal pair for /ɳ/. Conversely, in word-final position, it is straightforward to find contrasts for /ɳ/, as in rang, ran, ram , rat, rack, rag, rap, rash; but there is no equivalent minimal pair for /h/. The generalisation extractable from this is that /h/ appears only before a stressed vowel (or at the beginning of a syllable), as in hat, ahead, apprehensive, vehicular (but not vehicle , where /h/ appears in the spelling, but there is no /h/, as the stress here falls on the first vowel). On the other hand, /ɳ/ is not permissible syllable-initially: it can appear only at the end of a syllable, either alone, as in rang, hanger, or before a velar plosive as in rink, stinker, finger, stronger.
The additional criterion for allophony states that all the allophones of a phoneme must be phonetically similar. Using distinctive features, the allophones of a single phoneme must be different by no more than three features. Two sounds are highly unlikely to be allophones of the same phoneme if the number of contrasting feature values is higher than the number of shared ones. For /h/ and /ɳ/, this produces an unambiguous result: both are consonants, but there the similarity ends, [h] is a voiceless fricative, while [ɳ] is a voiced stop; /h/ is oral, while /ɳ/ is nasal; /h/ is glottal, while /ɳ/ is velar; /h/ is an obstruent, while /ɳ/ is a sonorant. Invoking phonetic similarity allows us to justify regarding /h/ and /ɳ/ as distinct phonemes, despite the lack of minimal pairs.
Phonetic similarity also helps in cases where a single allophone could theoretically be assigned to more chan one possible phoneme. This situation is commonly encountered when members of a natural class of phonemes undergo the same rule. This solution is supported by a requirement of phonetic similarity, involving the assignment of the two most similar allophones, to a single phoneme. An example is the case with the voiceless stop phonemes and their aspirated and unaspirated allophones.
In the normal course, some realisation of every phoneme in a language to appear in every possible environment: initially, medially, and finally in the word, and also before and after other consonants in clusters. However, there are two types of exception to this sweeping generalisation.
First, there are the phonotactic constraints of a language, which spell out which combinations of sounds are possible. In English, only three-consonant clusters are permissible; and the first consonant in the sequence must always be /s/. More specifically, /v/ and /m/ cannot be the first member of any initial consonant cluster. Phonotactic statements of this kind restrict the length and composition of possible clusters, on a language-specific (and period-specific) basis.
Secondly, some phonemes have defective distributions. They are not only restricted in the combinations of consonants they can form, but are simply absent from some positions in the word. English /h/ and /ɳ/ both fall into this category. The former is available only syllable-initially, and the latter only syllable-finally. Because those defective distributions are mutually exclusive, English |h] and /ɳ/ are in complementary distribution.
Phonemes with defective distributions like this are relatively rare. Sometimes, their defectiveness follows from their historical development. A chain of sound changes leading to the weakening and loss of /h/ before consonants and word-finally has left it 'stranded' only syllable-initially before a vowel. Often, defectively distributed phonemes are relatively new arrivals. For instance, the newest member of the English consonant system is /dʒ/, which developed in from sequences of [zj] in measure, treasure, and from French loans such as rouge, beige. The /zj/ sequence does not appear word-initially.