Incomplete neutralisation presents a problem for classical modular feed-forward grammars: it results in surface phonetic distinctions between phonologically neutralised segments. This paper argues for a model of incomplete neutralisation using two independently motivated theoretical devices: paradigm uniformity and weighted phonetic constraints. A case study is presented, showing that Japanese monomoraic lengthening results in incomplete neutralisation: when monomoraic nouns with short vowels are lengthened to fill a bimoraic minimality requirement, they reach a duration intermediate between that of unlengthened short vowels and underlyingly long vowels. The Japanese case has properties distinct from other classically cited examples of incomplete neutralisation such as final devoicing, which are not predicted by previous theories of neutralisation. The Weighted Paradigm Uniformity theory of incomplete neutralisation is shown to make four unique predictions, and is argued to better capture the typology of incomplete neutralisation.

While many phonological patterns target classes of sounds that can be defined phonetically, a large number of patterns in descriptive grammars involve sounds that cannot be easily characterised in phonetic terms. This finding suggests that phonological patterns themselves must be taken into account when learning phonological representations, and that phonological classes may emerge in learning from both phonetic factors (bottom-up) and phonological patterns (top-down). The current work presents a case of a phonetically unnatural class in South Bolivian Quechua that is active in the phonology of the language, and provides experimental support that this class is referred to by speakers’ grammars. While many cases of phonetically unnatural classes have been documented in descriptions of language patterns, in most cases there is little or no evidence that these patterns or classes are represented by speakers as they are described by linguists.

This paper re-examines theoretical constructs used in the analysis of Russian word stress, employing data from speakers with acquired surface dyslexia, a symptom which is characterised by impaired lexical access and preserved grapheme–phoneme correspondence rules. Russian stems have been traditionally analysed as lexically accented or unaccented, with a default rule deriving surface stress in the latter case. In the study reported here, we found no differences in the production of accented and unaccented stems. Instead, the analysis of errors revealed that the significant factors determining stress placement include stress neighbourhood and stress position. The speakers produced fewer errors in consistently spelled words, and there was a strong tendency to shift stress to the final syllable in consonant-final words, and to the penultimate syllable in vowel-final words. These results indicate that distributional properties play an important role in stress assignment in both accented and unaccented stem types.

Ternary or three-level vowel length is typologically rare, and supporting evidence is limited. This paper presents the results of an investigation into the hypothesised case of this configuration in Shilluk. We first describe the role of vowel length in Shilluk phonology and morphology, and then report on an acoustic study in which minimal sets for vowel length (short, long, overlong) are measured for vowel duration, coda duration, vowel quality and fundamental frequency. Short, long and overlong vowels differ significantly and substantially in terms of vowel duration: 96% of the items can be classified successfully for vowel length on the basis of this measurement alone. Of the other measurements, only vowel quality is significant, and this effect is considerably smaller. The mean values for vowel duration – 68, 111 and 150 ms for short, long and overlong vowels respectively – are similar to those reported for ternary vowel length in Dinka.

We present three case studies of marginal contrasts in Tongan loans from English, working with data from three speakers. Although Tongan lacks contrasts in stress or in CC vs. CVC sequences, secondary stress in loans is contrastive, and is sensitive to whether a vowel has a correspondent in the English source word; vowel deletion is also sensitive to whether a vowel is epenthetic as compared to the English source; and final vowel length is sensitive to whether the penultimate vowel is epenthetic, and if not, whether it corresponds to a stressed or unstressed vowel in the English source. We provide an analysis in the multilevel model of Boersma (1998) and Boersma & Hamann (2009), and show that the loan patterns can be captured using only constraints that plausibly are needed for native-word phonology, including constraints that reflect perceptual strategies.