Alternative analyses
Scobbie (1992), Donegan (1993)
Scobbie (1992) claims that forms with intrusive [r] have been assimilated over time from an /r/-less to an /r/-ful class. That is, the underlying representations for idea, saw, baa and Canada are now /aɪdiər/, /sɔ:r/, /bɑ:r/ and /kænədər/ for speakers with both linking and intrusive [r], while etymological /r/ remains in clear, soar, spar, letter. Speakers then select a `weak (less occlusive, strident, consonantal or long) phonetic interpretation of coda /r/' (1992: 9).

Similarly, Donegan assumes that `r in a syllable-fall (before consonant or pause) loses its r-coloring, becoming ə̭' (1993: 117). Because this lenition is exceptionless in non-rhotic varieties, speakers hearing final schwa (or /ɑ: ɔ:/) will `undo' the weakening to arrive at underlying final /r/. However, they may also, inappropriately, hypothesize lenition in comma, spa, saw words, and therefore assume underlying /r/ here, too. In Donegan's opinion, `The ``intrusive r'' does not, then, intrude because the speaker makes up an r-insertion rule. Instead, the r appears by analysis, when speakers assume that, because some final schwas represent /r/'s, other final schwas do so as well' (1993: 119). That is, the change is purely perceptual: `speakers with intrusive r's perceive final [ə]'s as /r/'s . . . Speakers without intrusive r's can perceive final schwas as /ə/'s (or they can ignore them)' (ibid.).

Both Scobbie and Donegan assume that intrusive [r]s arise from underlying /r/s innovated either by analogical extension, or by perceptual recategorization. In Present-Day English, /r/ is then weakened in codas. Invoking weakening rather than deletion is in itself problematic, since /r/ cannot be assumed to become schwa in every instance: many speakers lack a schwa offglide after /ɑ: ɔ:/, while both schwa and [r] surface in linking and intrusive contexts in the idea is, or farther away: [r] and [ə] are clearly not in complementary distribution. Donegan's perceptual interpretation raises further questions. Since Donegan assumes that linking [r] preceded intrusion historically, it seems reasonable to claim that speakers might learn underlying /r/ in alternating letter, where [r] would surface only prevocalically. However, we must then accept that subsequent generations could acquire underlying /r/ in comma, by perceptual reanalysis of the final [ə], despite its lack of alternation and hence of surface [r]. That is, the deletion accounts assume acquisition of underlying /r/ in non-alternating forms, which thereby become alternating. In such a phonology, it seems hard to see how we are to rule out underlying /r/, which would then delete categorically, in final clusters like harp, beard. Donegan's account of the diachronic development of the non-rhotic system also seems incomplete: she argues that some speakers perceive [ə] as /r/, while others perceive it as /ə/ (or, indeed, ignore it). However, there is no insight into why this discrepancy should arise; and surely, if we are to ascribe present-day features to historical developments, we should make some attempt to understand the history.

However, there are also more general difficulties, which are shared by Harris (1994) and McCarthy (1991). All these accounts assume a piecemeal, analogical extension of underlying /r/ to words with final /ɑ: ɔ: ə/. Of course, this is highly likely to represent the starting point of the generalization of [r], and is consistent with my hypothesis of subsequent rule inversion; but that rule inversion, or some parallel regularizing force, is necessary to account for the great regularity and productivity of linking and intrusion for many non-rhotic speakers now. The underlying /r/ analyses put intrusive [r] for non-rhotic speakers on a par with the fact that some rhotic speakers, Scots for instance, happen to have categorical [aɪdiər] for idea, or with West Country hyperrhoticity (Wells 1982), where rhotic speakers again may pronounce [r] medially in kha[r]ki, or finally in comma[r], Anna[r], regardless of the following context. But this is truly a sporadic, speaker-specific phenomenon, affecting individual lexical items or lexical sets; in all probability, it simply reflects the exposure of British rhotic speakers to quantities of non-rhotic speech, for instance via the spoken media. When rhotic speakers hear [r], they are likely to assume underlying and hence categorically pronounced /r/, and some confusion is inevitable. Intrusive and linking [r] for non-rhotic speakers do not share these characteristics: even when intrusion is suspended, the deciding factors seem to be socio-linguistic rather than lexical.

It is true that a change which proceeded analogically might conceivably result in a regular system; but analogy usually leaves tell-tale gaps. Even more damaging, however, is the question of where intrusive [r] does appear, rather than where it does not. Underliers like /rɔ:r/ `raw', /kæfkər/ `Kafka', / ʃɑ:r/ `Shah' may simply be unfamiliar, but others are downright improbable. We would have to assume, for instance, that a speaker hearing Stella Artois, or BUPA, or dona (in Latin dona eis requiem), or even seeing them on the printed page, would immediately set up underlying forms with final /r/. Any phonologist wishing to derive past tense forms of strong verbs from present tense bases would also have to posit /r/ in the underlying representation of see, because of intrusive [r] in saw[r]it (Johansson 1973). In this case, we would have to assume that there was no underlying /r/ historically, because of the absence of a centring diphthong in see (as opposed to seer), but that /r/ was innovated here, in an entirely inappropriate phonological context, in order to allow [r] to surface prevocalically in the past tense form. Forms with optional reduction, like tomato, potato, raise parallel problems: presumably, these would require alternative lexical entries with either a final rounded vowel, or final schwa plus /r/, rather than two productive and interacting processes of vowel reduction and [r]-insertion. Children's errors, like [ə'ræpl] `an apple' (Johansson 1973: 61), the[r] animals, a[r]aeroplane (Foulkes 1997: 76) can also be explained most easily given an insertion rule: the child may not yet have learned the prevocalic allomorph of the indefinite article, and the schwa-final article creates a hiatus which provides an appropriate context for [r]. In a deletion account, we must assume that the child has set up an allomorph of the article with final /r/: since such forms are reasonably frequent for children, but exceptionally rare for adults, why does this allomorph not persist? Finally, slower speech seems to produce fewer [r]s; insertion rules are typically constrained by pauses, presumably in this case because the hiatus is less likely to be perceived under these conditions. However, a deletion or weakening rule would have to operate more frequently in slower speech, contradicting the normal association of deletion with fast and casual registers. These shared problems should be borne in mind as we turn to two further analyses assuming underlying /r/ in unetymological contexts.