Sonorant+Obstruent Stem-final Clusters

The obstruent in such clusters may combine with a nasal or a liquid [r, l]. First, nasal+obstruent stems followed by the preterite morpheme show no variation across speakers or verbs; the obstruent invariably deletes, as shown in 2.¹

The different sources adopt different transcription conventions, and often ones that do not conform to the IPA system. I use here uniform IPA transcriptions. Icelandic stops are all phonetically voiceless but show a contrast in aspiration. Authors vary in their transcription of unaspirated stops: Rögnvaldsson (1989) systematically uses [

], while Richter (1982), Helgason (1993), and Gussmann (2002) write [p, t, k]. Einarsson (1945) distinguishes the underlyingly unaspirated [

], from the deaspirated [p, t, k]. Blöndal (1920) and Gibson (1997) do not note devoicing of orthographic <b, d, g> and simply transcribe them [b, d, g]. For the purposes of this study it does not really matter which system we adopt, as long as it is used consistently. For the sake of maintaining phonetic accuracy while avoiding unnecessary diacritics, I simply use [p, t, k] (and [c], palatal stops being predictable from the context) for all unaspirated stops.

Stems ending in [l, r]+obstruent show more variation in the preterite form, both within and across speakers. Unlike with nasals, the stem obstruent is not automatically lost. In the [l]+obstruent forms in 3 the obstruent is obligatorily omitted. But in those in 4 the velar obstruent is variably retained by speaker O; Blöndal and Einarsson also note this variable retention. Speaker H, however, treats all [l]+obstruent forms alike and uniformly omits the obstruent.

When preceded by [r], stem-final obstruents only variably delete for all speakers, as shown in 5. Variable obstruent retention does not necessarily hold for every verb and every speaker. For instance, speaker H retains the whole cluster in [εrkðI] ‘tease-pret’. (5c), optionally deletes the stop in [pεr(k)ðI] ‘taste-pret’. (5a), and obligatorily omits it in [

] ‘mark-pret’. (5e). Individual verbs may behave differently, but this range of behaviors indicates that obstruent deletion is globally variable for stems ending in [r]+obstruent sequences. The same word may also be treated differently by different speakers. For instance, the form verpti ‘lay eggs-pret’. (5g) was reported with a stable clustermedial stop by speaker O [

] and with obligatory deletion by speaker H [

], while Einarsson indicates variable deletion [

]. I largely disregard here the detailed behavior of individual verbs, which certainly depends on factors such as frequency and register.

Again, speaker H distinguishes himself from speaker O, Blöndal, and Einarsson, and the two patterns appear in distinct columns in 5. Speaker O, Blöndal, and Einarsson, although they share the same global pattern, may diverge for specific forms on whether deletion is either obligatory or optional, or optional or excluded; such variations have no effect on the issues discussed here and are not noted. In the last column both forms in 5i and [

] in 5h were reported only by speaker O; this is noted in parentheses.

Rögnvaldsson (1989: 51) notes the ungrammaticality of the pronunciations [

] or [

] According to him (p.c.), [

] is a spelling pronunciation.

Two main patterns emerge for verbs with a stem-final sonorant+obstruent sequence. One pattern is exemplified by speaker H, who drops all obstruents next to [l] but variably retains them after [r], depending on the particular sequence and verb. H, then, treats [l]+obstruent and nasal+obstruent combinations alike, with categorical obstruent deletion, and differently from [r]+obstruent sequences. The other pattern is illustrated by speaker O, whose judgments correspond fairly closely to the pronunciations indicated in Blöndal and Einarsson. For them obstruent deletion is variable next to liquids, [r] or [l], and obligatory next to nasals.

The available data suggest a hierarchy among sonorant consonants with respect to the likelihood of deletion of an adjacent obstruent. An obstruent is most likely to be omitted after a nasal, which triggers obligatory deletion for all speakers. Omission is less likely next to an [l], as it is obligatory only for a subset of speakers, and least likely after [r], which does not clearly trigger obligatory deletion for any speaker. As discussed in more detail below, this hierarchy is to be related to the level of contrast in manner of articulation between the obstruent and the sonorant.

Two additional aspects of these patterns need to be mentioned. First, speakers differ on the realization of stops in preterites with stems ending in liquid+obstruent. Stops alternate with fricatives in various contexts in Icelandic; compare sleppa ‘let go-inf.’ [stlεhpa] with sleppti ‘let go-pret.’ [stlεftI] or sagna ‘story-gen. pl.’ [sakna] with saga ‘story-nom. sg.’ [

] and sagt ‘say-past. part.’ [saxt]. In the preterite forms under consideration, fricativization applies in the speech of speaker O, as well as for Einarsson and Blöndal. The process appears to be systematic with aspirated [k] in the infinitive (aspiration resulting in the devoicing of the preceding sonorant), which alternates with [x] in the preterite, as in infinitive [

] versus preterite [

] (4b), or infinitive [

] versus preterite [

] (5e). Fricativization also variably applies to aspirated [p] for consultant O, but I found no mention of this in Blöndal or Einarsson. O accepts the alternate pronunciations [

] and [

] for skerpti (5h), [

] and [

] for skyrpti (5i), but fricativization with labials seems to be blocked in verpti (5g). Speaker H, however, appears to lack the fricativization process altogether and keeps stem-final stops as such in the preterite, for example, ergði [εrkðI] (5c).

The second point concerns the absence of voiced fricatives in the preterite forms. Underlying [v] in the infinitive is systematically omitted in the preterite, as in erfa ‘inherit-inf.’ [εrva] versus erfði ‘inherit-pret.’ [εrðI] (5j). As for [

], it could be expected to arise in the preterite from fricativization of the underlying unaspirated [k] in the infinitive (for those speakers who apply this process). If aspirated [k] in [

] yields [

] in the preterite (4a), why does [fIlca] correspond to the preterite [fIltI] (3a) and not [

]? To anticipate the analysis to come, the absence of [v] and [

] will be related to a general constraint against voiced fricatives.

A note concerning the transcriptions indicated in Blöndal and Einarsson is in order. For [l]+obstruent and [r]+obstruent stems involving voiced [l] and [r], these sources provide forms like fylgdi ‘follow-pret.’ [

], skelfdi ‘frightenpret.’ [scεl(v)tI], ergði ‘tease-pret.’ [

], and erfði ‘inherit-pret.’ [

]. Such transcriptions suggest that the obstruent is variably retained, accompanied by a process of stop fricativization, contrary to what is indicated in 3 and 5. In fact, the pronunciations with a full voiced fricative sound unnatural and hypercorrect (Gussmann 2002). According to an anonymous reviewer and E. Rögnvaldsson (p.c.), the fricatives in [

], [scεlvtI], [

], and [εrvðI] rather represent an attempt at indicating velarization or labialization on the liquid, not a distinct velar or labial segment. Consequently, I have omitted these voiced fricatives from the transcriptions. However, if they were fully pronounced in 3, 5, and 7, the generalizations and analysis to follow would not be affected in a significant way. The only modifications would concern the status of voiced fricatives, which is completely orthogonal to the analysis, and the ranking of the Sonority Sequencing Principle (section 3.4). The perceptual aspects, in particular, would remain unchanged.

Velarization and labialization of the liquid is not noted in the forms in 3–7 and will be disregarded in the discussion below. According to E. Rögnvaldsson (p.c.), velarization and labialization are variable and velarization is more prominent than labialization. He also notes that velarization and labialization are much weaker with [r] than with [l], and could be related to the spelling. The status of these secondary articulations remains unclear, in particular as it relates to fricativization and voicing (secondary articulations appear on voiced but not voiceless liquids; E. Rögnvaldsson, p.c.). Their integration into the analysis would not be problematic, however, if they were to be analyzed as a secondary retention of the point of articulation of the deleted obstruent, or as [

] and [v] becoming secondary articulations on the adjacent liquid (which could be related to the fact that these sounds are usually realized as approximants [

] and [ν]; Helgason 1993).

Obstruent+Sonorant Stem-final Clusters

All obstruent+sonorant stems obligatorily lose their obstruent in the preterite. This applies to obstruent+nasal (6) and obstruent+[l] (7) stems.⁵

With obstruent+[l] stems, Blöndal and Einarsson provide three possible pronunciations for the preterite form, for example, sigldi ‘sail-pret.’: [

], [sIltI] and, more surprisingly, [

], with apparent metathesis of the stem-final consonants. Kress (1963:41–42) notes the same alternation between these three forms. As in liquid+obstruent stems (see note 4), the voiced fricative in these transcriptions appears to indicate possible velarization/labialization. This accords with Helgason (1993), for instance, who considers [sIltI] with deletion of the obstruent to be the only natural pronunciation of sigldi. Interestingly, Einarsson (p. 14) indicates that the [

] is shortened before [lt], as in [

], which may point to its nonsegmental status. As for the reason why two distinct forms, [

] and [sIl

tI], are indicated, E. Rögnvaldsson (p.c.) suggests that they do not imply a phonetic or phonological distinction and that [sIl

tI] may be related to the fact that speakers tend to misspell this (and other similar) forms and write silgdi instead of sigldi.

Obstruent+Obstruent Stem-final Clusters

The only final cluster with two obstruents is [sk], which appears in two stems. The velar stop has a strong tendency to be lost in the preterite, as shown in 8. For speaker H, retention of the [k] is acceptable, though somewhat marginally, with one of the two verbs (8a). Einarsson also marks the stop as optional in this form. Speaker O (in agreement with Blöndal) omits the stop in both forms.

Sonorant+Sonorant Stem-final Clusters

These clusters include [lm], [rm] and [rn]. In the preterite form the cluster-medial nasal never deletes in any of my sources. Initial [l] also appears stable, so [lm] clusters before the preterite morpheme surface intact for my two consultants, and neither Einarsson nor Rögnvaldsson, who otherwise give a long list of cases of consonant deletion, note the dropping of a consonant in such forms. This is shown in 9.

Postvocalic [r], however, may be dropped before a nasal. For my two consultants, as well as Einarsson (see also Gussmann 2002:206), [rm] stems behave like [lm] ones above and tolerate no simplification (10). Only Rögnvaldsson indicates the deletion of the initial [r] in similar forms.⁶

With [rn] stems (11), [r]-deletion is more frequent and occurs not only in Rögnvaldsson (R), who cites the form [stIntI] (11a), but also in consultant H's speech. H considers that deletion is optional but preferred in this case. The possibility of [r]-dropping is also noted in Blöndal (B) and Einarsson (E) (p. 82). Speaker O, unlike all the others, rejects the [r]- less outputs.

What is of particular interest to us here is the variation observed in the domain of application of [r]-deletion. For Rögnvaldsson, it applies before [n] and [m] alike, whereas for speaker H and Einarsson it is restricted to [n]. Again, I suggest that the distinction between [n] and [m] with respect to the likelihood of omitting the preceding [r] relates to the role of contrast between adjacent segments in consonant deletion, here contrast in place of articulation.

General Considerations on Consonant Deletion and Perceptibility

I view segmental processes and consonant phonotactics as constrained or motivated by requirements on the perceptibility of segments and features. Consonants in particular should be perceptually salient, and deletion applies when a consonant lacks perceptual salience and becomes more easily confusable with nothing, that is when the cues that permit a listener to detect its presence are diminished.⁹

The perceptual salience of a segment—or its degree of confusability with zero—is a function of the quantity and quality of the auditory cues that signal its presence in the speech stream. For consonants, cues can be grouped into two categories: internal cues produced during the closure part of the consonant, and contextual cues that originate from neighboring segments. In addition, an important cue to stops is their release burst. (See Wright 1996 for a summary of available cues to consonants' place and manner of articulation). I hypothesize that there is a correlation between the perceptibility scale of consonants and the likelihood that they delete. In other words, the likelihood that a certain consonant deletes depends on the quality and quantity of the auditory cues associated with it in a given context. In my view, however, the effect of perception on consonant deletion is not direct but mediated by phonological features, or other phonological categories, which have perceptual correlates (see Côté in press b for discussion).

Important cues to consonants are found in neighboring vowels. Formant transitions from and to adjacent vowels provide optimal contextual cues to consonants because of their high amplitude and dynamic pattern, which gives information about the changing configuration of the vocal tract. They provide cues to all aspects of the articulation of consonants: place, manner, and laryngeal settings. The major role played by vocalic transitions entails the desirability for consonants to be adjacent to at least one vowel (see, for example, Wright 2001). In Icelandic, word-internal cluster simplification applies in sequences of three (or more) consonants (generalization 12a), that is, when there is at least one consonant that is not adjacent to a vowel, and therefore lacks the associated perceptual cues. Consonant deletion does not take place in word-internal two-consonant sequences, in which both consonants are flanked by a vowel. This is a common situation crosslinguistically, which is expressed here in terms of adjacency to vowels (rather than avoidance of complex syllabic constituents, for example).

The requirement that consonants be adjacent to a vowel is expressed in the markedness constraint in 13a, also used independently by Steriade (1999c) and Fleischhacker (2001). Consonant deletion in triconsonantal clusters is obtained by ranking C[harr ]V above MAXC, the faithfulness constraint against consonant deletion (13b). This is shown in the tableau in 14.

This basic ranking raises two problems for the Icelandic data. First, if nonadjacency to vowels is a necessary condition for cluster simplification to take place, it is not a sufficient one. All the preterite forms examined here contain clusters of three consonants but not all of them display consonant deletion. Second, MAXC and the markedness constraint in 13a do not determine which of the consonants will actually delete. Whether the first, second or last consonants deletes, MAXC is violated and C[harr ]V is obeyed in exactly the same way.

These problems are solved by expanding MAXC into a family of constraints built on the notion of relative perceptibility. More perceptible segments or features are less susceptible to loss; formally this is accounted by associating them with higher-ranked MAX constraints, which protect them from deletion, while less perceptible elements are subject to lower-ranked constraints. For example, MAXC1>>MAXC2 if C1 is more salient than C2. If C[harr ]V ranks between MAXC1 and MAXC2, the presence of a consonant that is not adjacent to a vowel compels the deletion of C2 but not that of C1, which is protected by its greater salience and more highly ranked faithfulness constraint.

Any factor that affects perceptibility may be integrated into MAXC constraints.¹⁰

1. The nature of the consonants: Stops and nonstrident fricatives have weaker internal cues;
2. The level of contrast between adjacent segments: The more similar a consonant is to its neighboring segments, the less perceptible and the more susceptible to deletion it is.

The Special Status of Nonstrident Obstruents

The data reveal a clear hierarchy of deletion in terms of categories of consonants. The most vulnerable consonants are the stops and nonstrident fricatives [

] (the latter two only for the speakers who fricativize velar stops, namely, O, Blöndal, and Einarsson), which generally delete if they are present in the stem. If the stem contains no obstruents, only postvocalic [r] may delete (this case will be analyzed in section 3.5.). Nonstrident obstruents contrast with nasals and laterals, which are stable in all positions (generalization 13b). For instance, the cluster [lmt] in hylmdi (9) is maintained by all speakers, whereas [lkt]/

] or [

]/[

] (depending on whether fricativization applies) invariably or optionally lose their medial obstruent (6, 7). These clusters only differ in the identity of the middle consonant: A nasal in the first case, a (nonstrident) obstruent in the second. Also, in obstruent+nasal and obstruent+[l] stems, it is the postvocalic obstruent and not the cluster-medial sonorant that deletes.

The status of [s], the only strident fricative in Icelandic, with respect to deletion, is not clearly revealed by the preterite data. The only stemfinal cluster involving [s] is [sk], which normally loses its medial stop in the preterite (8). This suggests that interconsonantal [k] is weaker than postvocalic [s], which is expected based on the relative strength of their internal cues but also on the fact that [s] benefits from postvocalic transitions. However, a look at the behavior of [s] in contexts other than weak preterites shows that it is more resistant than nonstrident fricatives (section 4); this confirms that the set of deletable obstruents includes stops and nonstrident fricatives and excludes [s] as well as sonorants.

I suggest that the contrast between nonstrident obstruents, sonorants and [s] in their propensity to delete follows from the relative perceptual salience of the consonants' internal cues. Aperiodic signals, such as those found in stop release bursts and fricative noise, are more easily masked in noise than periodic signals (Wright 2001). So it may be hypothesized that segments with aperiodic signals (obstruents) are more susceptible to loss than segments with formant structure (sonorants). Among obstruents, however, stops are weaker than fricatives. Due to the absence of oral or nasal airflow, the closure part of a stop is silent or associated only with low-amplitude vocal fold vibrations. The noncontextual perceptual cues to stops are rather concentrated in their release burst, but stops not followed by a vowel do not reliably benefit from an audible burst. Among fricatives, nonstrident ones are associated with noise of low amplitude. Miller and Nicely (1955), for example, show that the distinction between stops and the weak fricatives becomes unreliable in masking noise. Strident fricatives [

] involve much stronger frication noise.

The greater resistance of sonorants and strident fricatives is accounted for by splitting MAXC into constraints specific to different categories of consonants, sonorants, obstruents, and strident fricatives (15a–c). The constraints are ranked as in 15d, which will protect sonorant and strident consonants from deletion.

Contrast and Obstruent-final Stems

If nonstrident obstruents are more vulnerable, not all of them delete with the same likelihood. In sonorant+obstruent stems, this depends essentially on the nature of the adjacent consonants in the sequence and the level of contrast in manner of articulation between them.

The general idea is that consonants that are more similar to adjacent segments are more readily deleted than consonants that are more contrastive. Like the greater vulnerability of obstruents, this is related to perception. After many others (for example, Kawasaki 1982, Ohala and Kawasaki 1985, Boersma 1998), I assume the existence of a correlation between the amount of acoustic modulation in a sound sequence and its perceptual salience. Contrast enhances perceptibility, and therefore favors retention. The greater the modulation, the greater the salience, the more easily perceptible the segments involved are.¹¹

We can hypothesize that, everything else being equal, sound combinations displaying a greater modulation in a given dimension or modulation in a larger number of dimensions are perceptually better, and are predicted to be more common, than other sequences with a smaller modulation in the same dimension or modulation in fewer dimensions. This can be transposed in featural terms, assuming that features are associated with some acoustic/perceptual contrast: A segment is more perceptible if it contrasts in n features with its neighboring segments than if it contrasts in n−1 features (keeping constant the nature of the contrasting features).¹²

Contrast in both manner and place of articulation is relevant in Icelandic clusters.¹³

Contrast in manner is computed by means of the major class features proposed by Clements (1990): [sonorant], [approximant], and [vocoid]. The feature specifications for different classes of segments appear in 16. In addition, obstruents are distinguished by the feature [strident], which I assume is not specified for sonorants: Stops and fricatives other than [s] are specified [-strident], while [s] is [+strident].

Crucially, [r] is specified here as [+vocoid], unlike the lateral [l], which is then the only consonant in Icelandic left in the liquid category. This assumption accords well with the fact that [r] relates to [j] in terms of phonotactic behavior in Icelandic (Árnason 1980b:38–43) (see Côté in press a for a similar proposal for French). Categorizing [r] as a glide-like consonant is also justified phonetically. This consonant is produced in Icelandic as an apical trill or tap, which involves vocalic portions between any two apico-alveolar contacts, but also often at the beginning and end of the consonant (Pétursson 1973, Ladefoged and Maddieson 1996). These vowel-like portions are consistent with a [+vocoid] specification.

Contrast in manner of articulation between two categories of segments is determined by comparing their plus- or minus-specifications in 16. In particular, a hierarchy is established between glides (including [r]), liquids, and nasals in terms of the level of similarity they display with obstruents: Glides contrast with obstruents in [vocoid], [approximant], and [sonorant]; liquids contrast with them in [approximant] and [sonorant]; nasals only in [sonorant].

The effect of contrast in cluster simplification is accounted for by integrating contrasting features into MAXC constraints, as in 17. Contrast may refer to one (17a) or many (17b) features; it may be associated with the general MAXC as well as the more specific MAXC[+son] (or MAXC[-son]) (17c). These constraints are ranked as in 17d: Consonants that contrast in more features are more salient and more resistant to deletion, and the MAXC constraints associated with them are ranked higher.

Contrast constraints may also refer specifically to the preceding or following consonant (18), as it is clear that a given consonant X may not be equally perceptible in the context XY (if followed by a segment Y) and in the context YX (if preceded by Y), even though the level of contrast is the same in both sequences. This follows from differences in the distribution of acoustic cues as well as nonlinearities in the response of the auditory system to acoustic modulation. In particular, the onset of a stimulus signal (for example, frication noise at high frequencies or formant structure at lower frequencies) is perceptually more salient than its offset (Bladon 1986, Silverman 1995, Wright 1996, Delgutte 1997).

Obstruents in particular receive better contextual cues from a following sonorant than from a preceding sonorant, due to the enhanced auditory response at the onset of formant structure (and voicing for voiceless obstruents). MAXC[-son] constraints that refer to contrast in [sonorant], [approximant], and/or [vocoid] with the following segment are therefore ranked higher than the corresponding constraints that refer to the preceding segment. This, in conjunction with the ranking schema in 17d, generates the rankings in 20 among the MAXC[-son]/CTR:F constraints defined in 19. The two vertical hierarchies express the fact that obstruents are more vulnerable if they contrast in fewer features (because the MAXC constraint they are associated with is ranked lower); the horizontal lines establish that obstruents are less marked if followed rather than preceded by a sonorant.

Prevocalic obstruents are in a perceptually optimal position since they contrast maximally with the following segment and also benefit from an enhanced auditory response. Their stability, specifically here that of the suffix consonants [t, ð], is guaranteed by the the highestranked MAXC[-son]/→CTR:[son+app+voc]. The vertical hierarchy on the right derives the correlation between the likelihood of obstruent deletion and contrast in manner between the obstruent and the preceding consonant. Recall that speaker H systematically deletes obstruents when the preceding segment in the stem is a nasal (2) or [l] (3–4), but variably retains them next to [r] (5). Speaker O, Blöndal, and Einarsson also obligatorily delete nonstrident obstruents after a nasal, but marginally retain them after [l] and more readily next to [r]. Nasals, then, trigger deletion more than [l], and [l] more than [r]. This generalization is restated in featural terms as follows. With only a contrast in [sonorant] (after a nasal), the obstruent is obligatorily deleted by all speakers, in violation of MaxC[-son]/←tr:[son]; with a larger contrast in [sonorant]+[approximant] (after [l]), the obstruent is variably retained by a subset of speakers (O, B, E) but still systematically deleted by others (H), violating MAXC[–son]/←CTR:[son+app]; with a maximal contrast in [son+app+vocoid] (after [r]), all speakers allow the optional retention of the obstruent.¹⁴

We now have the essential ingredients to provide a formal analysis of the deletion patterns in stems ending in an obstruent (nasal+obstruent, [l]+obstruent, [r]+obstruent, and [sk]). Let us start with nasal+obstruent stems [

], which are subject to categorical obstruent deletion (2). This is shown in the tableau in 21 with [mp] as an example (V represents any vowel). Four outputs should be considered: the faithful one and the outputs with deletion of each of the three consonants in the sequence. The faithful candidate violates C[harr ]V (21a), the constraint requiring every consonant to be adjacent to a vowel. The other candidates violate at least one of the MAXC constraints. Deletion of the nasal violates MAXC[+son] (21b), that of the suffix consonant [t] MAXC[-son]/→CTR:[son+app+voc] (21d). The winning candidate [VmtI] violates MAXC[-son]/←CTR:[son] since the deleted obstruent [p] contrasts in [son] with the preceding nasal (and contrasts in no manner features with the following [t]) (21c). This constraint must rank below C[harr ]V and MAXC[+son]; 20 has already established that it is ranked below MAXC[-son]/→CTR:[son+app+voc]. This constraint is never violated in the data and candidates with deletion of prevocalic obstruents will not be shown in the following tableaux.

Verbs with a stem ending in [r]+obstruent display for all speakers variation between deletion of the medial obstruent and retention of the full cluster (5). Following Anttila (1997) and Auger (2001), among others, I assume that variation derives from rankings between constraints that remain indeterminate, specifically here the ranking between C[harr ]V and the highest MAXC constraint violated by the output with deletion. In this situation, the permissible outputs are those generated by all the possible strict rankings compatible with the ranking indeterminacy. The possible rankings are MAXC>>C[harr ]V and C[harr ]V>>MAXC, which yield variation between full cluster retention (optimal if MaxC>>C[harr ]V) and deletion (optimal if C[harr ]V>>MAXC).

Deletion of the medial obstruent violates MaxC[-son]/←Ctr:[son+app+voc] since the obstruent contrasts in all those features with [r]. Retention of the entire cluster violates C[harr ]V and the ranking between these two constraints remains indeterminate. The omission of [r] violates MAXC[+son] and, as with nasals above, this compels a fatal violation. This is illustrated in the tableau in 22 with [

] as an example (5g-i). (In this and the following tableaux, I disregard the process of fricativization used by some speakers in the preterite with liquid+obstruent stems.)

Recall that voiced fricatives never surface in the preterite forms, either underlying as in [εrva] ‘inherit-inf.’ or as a result of the fricativization of labial or velar stops for speaker O, Einarsson, and Blöndal. In fact, voiced fricatives are generally vulnerable in Icelandic and are regularly omitted in various contexts (Helgason 1993). To account for the absence of [v] and [

] I use the constraint in 23. This constraint is consistent with the relative markedness of voiced fricatives with respect to voiceless ones crosslinguistically, which may be motivated both articulatorily and perceptually. On the one hand, the production of simultaneous voicing and noise requires greater effort (Ohala 1983). On the other hand, “the strong low-frequency energy that results from voicing tends to mask the lower-amplitude frication noise in the higher frequency range” (Ladefoged and Maddieson 1996:176).

The situation is more complicated since the voiced fricatives [v, ð,

] tend to be produced as approximants (Helgason 1993). But a complete analysis of the phonetics and phonology of voiced fricatives in Icelandic would lead me too far afield.

Consider stems ending in [rv] taking [ðI] as the preterite suffix (5j,k). In order to ensure the absence of the stem [v], but not that of the suffix [ð], *[

] must rank between MAXC[-son]/→CTR:[son+app+voc] and MAXC[-son]/←CTR:[son+app+voc]. The ranking between the latter constraint and C[harr ]V is not determined (22). The candidates with deletion of the suffix [ð] are eliminated by undominated MAXC[-son]/→CTR:[son+app+voc] (24d–e), the one with [r] deletion by MAXC[+son] (24b). The ranking *[

]>>MAXC[-son]/←CTR:[son+app+voc] then favors [v] deletion [VrðI] (24c) over retention [VrvðI] (24a).

With [l]+obstruent sequences, we observe interspeaker variation (4). The two stem consonants contrast in [son+app], a level intermediate between the nasals and [r]. Speaker H treats these forms like the nasal+obstruent ones and deletes the obstruent; speaker O (as well as Einarsson and Blöndal) indicates variable deletion, as with [r]. We conclude that in H's grammar C[harr ]V dominates MAXC[-son]/←CTR:[son+app] and the tableau for [l]+obstruent forms would be identical to the one with nasals (21), with contrast in [son+app] rather than [son]. In O's grammar, the ranking between MAXC[-son]/←CTR:[son+app] and C[harr ]V remains indeterminate, as in 22. These tableaux are omitted.

In stems ending in [sk] (8), the stop contrasts in the feature [strident] with the preceding consonant, and its deletion violates the constraint in 25. Contrast in [strident] is independent from contrast in the other manner features in 16, and speakers may or may not consider this contrast sufficient to retain the adjacent stop (however they behave with respect to [son], [app] and [vocoid]). So the constraint in 25 remains unranked with respect to those in 19. Speaker O and Blöndal delete the stop, so C[harr ]V dominates MAXC[-son]/←CTR:[strident], as illustrated in 26. Speaker H and Einarsson optionally keep the medial obstruent, with indeterminate ranking between MAXC[-son]/←CTR:[strident] and C[harr ]V, as in 27. For all speakers MAXC[+str] is undominated and eliminates the candidate with deletion of [s].

Sonority and Obstruent+Sonorant Stems

Obstruent+sonorant stems raise a couple of additional issues. The obstruent invariably deletes, before [n] (6) or [l] (7), a fact that is not predicted by the ranking so far. Consider the verb [sIkla] ‘sail-inf.’ and its preterite [sIltI] (7a). The deleted [k] in this form appears in the context V__[l]; it contrasts in [son+app+voc] with the preceding segment and in [son+app] with the following one. Deletion then violates MAXC[-son]/←CTR:[son+app+voc] and MAXC[-son]/→CTR:[son+app], as shown in the tableau in 28. It has already been established that MAXC[-son]/←Ctr:[son+app+voc] and C[harr ]V remain unranked with respect to each other (tableau 22). This predicts, contrary to the facts, that the [k] in [sIkla] should be optionally or obligatorily retained, depending on the ranking of MAXC[-son]/→CTR:[son+app]. If ranked above C[harr ]V (ranking A), only the incorrect output [sIkltI] survives; if ranked below C[harr ]V (ranking B), or if ranking between the two constraints remains undetermined, both [sIkltI] and [sIltI] are optimal.¹⁶

The existence of morpheme-internal sequences like [

] (Richter 1982:77) also shows that stops occur before [l] in triconsonantal clusters, so [k] deletion in [sIltI] is not motivated simply by a constraint against such clusters, like C[harr ]V.

I suggest that the relevant markedness constraint ruling out [sIkltI] is the Sonority Sequencing Principle (SSP), which, in its standard syllabic definition, requires the level of sonority to go up in syllable onsets and down in codas (for example, Hankamer and Aissen 1974, Hooper 1976, Clements 1990). The faithful candidate [sIkltI] contains the sequence [klt], with a sonorant consonant trapped between two obstruents. Such a form cannot be syllabified without violating the SSP: A syllabification [sIkl.tI] contains an ill-formed coda, and [sIk.ltI] an ill-formed onset. This situation arises when a consonant is flanked on each side by a less sonorous one, that is when a consonant corresponds to a local sonority maximum.¹⁸

If SSP is ranked sufficiently high, invariable obstruent deletion is predicted in obstruent+sonorant forms. But obstruent+nasal stems raise one more issue: The obstruent deletes but its place of articulation remains on the nasal, as in infinitive [cεkna] versus preterite [cεηtI] (6) (see generalization 12g). This also has a perceptual motivation: The retention of the place feature of the stop follows from the robustness of the place cues carried by the preceding vowel. On the other hand, the nasal lacks good place cues between two obstruents. I suggest using MAXF(eature) constraints (see Zhang 2000 and the references cited there) and posit the constraint in 30a that militates against the deletion of a place feature cued by an adjacent [+vocoid] segment. This constraint ranks above the general MAX[place] constraint.

In the case of obstruent+[l] stems, however, the remaining lateral retains its coronal place of articulation. I assume that MAX[place]/[voc] is dominated by a constraint requiring that lateral consonants be coronal (31).¹⁹

The tableaux in 32 and 33 illustrate how the correct outputs are obtained, for obstruent+[n] and obstruent+[l] stems, respectively. The sequences [kl] (7a,b) and [kn] (6a–c) are used as illustrations (the sequences [pl] and [pn] behave identically). In 32, candidates with stop deletion each violate two MAXC[-son] constraints (32c,d). SSP must rank above them to eliminate the faithful candidate (32a), while [n] deletion is excluded by MAXC[+son] (32b). MAX[pl]/[+voc], which prevents the deletion of the place of articulation of the postvocalic [k], determines the winner with velar [N] (32d) over coronal [n] (32c). The rankings of MAXC[-son]/→CTR:[son] and MAX[pl]/[+voc] are irrelevant as long as the former remains lower than SSP and the latter dominates MAX[pl].

The tableau in 33 is almost identical. One difference concerns the constraint LAT=COR. Ranked above MAX[pl]/[+voc], it eliminates the output [V;tI] with a velar lateral (33d). The level of contrast between [k] and the following lateral is also greater than between [k] and a nasal, and [k] deletion violates MAXC[-son]/→CTR:[son+app] rather than MAXC[-son]/→CTR:[son].

[r] Deletion and Sonorant+Sonorant Stems

With sonorant+sonorant stems different deletion patterns emerge. Speaker O retains all sonorant consonants in the preterite forms, irrespective of their nature and position. For him, then, MAXC[+son] dominates C[harr ]V and no more rankings are needed to generate the correct outputs. The retention of both stem consonants in [hIlmtI] (9) is illustrated in 34; [r]-initial sequences behave identically.

Other speakers also retain the entire sequence in [hIlmtI] but allow postvocalic [r] deletion when no weak obstruent is present in the stem. This process also appears to be driven by contrast, in both manner and place of articulation (generalizations 12e and 12f). Postvocalic [r] does not contrast in manner with the preceding vowel, also specified [+vocoid,+approximant, +sonorant]. All other sonorant consonants contrast in at least one manner feature with the preceding segment; they are protected from deletion by the constraint in (35) (which stands for all the MAXC[+son]/←CTR:F constraints in which F corresponds to a manner feature or combination of features).

Deletion is also affected by contrast in place of articulation. The data in 10 and 11 show that speaker H and Einarsson omit postvocalic [r] before [n] but not before [m], which I relate to the fact that [r] and [n] share the same place of articulation. The following constraint excludes [r] deletion in [rm].

The tableau in 37 illustrates the grammar of speaker H and Einarsson for preterite forms with sonorant+sonorant stems. Deletion of the sonorants in [VlmtI] (37a) violates MAXC[+son]/←CTR:M, since [l] contrasts with the preceding vowel in [vocoid] and [m] with the preceding lateral in [approximant]. The faithful candidate violates the lower-ranked C[harr ]V and comes out as optimal. Postvocalic [r] deletion, however, does not violate MAXC[+son]/←CTR:M but the lower-ranked MAXC[+son] (37b,c). In [rm] sequences it also violates MAXC[+son]/→CTR:Pl (37b) but not in [rn] (37c). Ranking MAXC[+son]/→CTR:Pl and MAXC[+son]/←CTR:M above C[harr ]V excludes deletion in [lm] and [rm]; unrankedness between C[harr ]V and MAXC[+son] yields variable [r] deletion in [rn].²⁰

Final Ranking

To sum up the analysis, the rankings established are put together in the figure 1 below. There are three types or rankings: fixed (universal) rankings, language-specific rankings common to all Icelandic speakers, and speaker-specific rankings. They are indicated by different lines.

This grammar contains several zones of variability, which are accounted for with crucial non-ranking between constraints:

1. Indeterminate ranking between C[harr ]V and MAXC[-son]/←CTR:[son+app+voc] yields variable obstruent deletion with [r]+obstruent stems.
2. Indeterminate ranking between C[harr ]V and MAXC[-son]/←CTR:[son+app] yields variable obstruent deletion with [l]+obstruent stems (speaker O, Blöndal, and Einarsson only).
3. Indeterminate ranking between C[harr ]V and MAXC[-son]/←CTR:[strident] yields variable obstruent deletion with [sk] stems (speaker H and Einarsson only).
4. Indeterminate ranking between C[harr ]V and MAXC[+son] yields variable [r] deletion with [rn] stems (speaker H, Blöndal, and Einarsson) (also [rm] for Rögnvaldsson).

Final Constraint Ranking.