3.1 Previous descriptions of stress

Primary stress in Tongan is straightforward, falling on the second-to-last mora of a domain that includes stems, suffixes and some enclitics (Churchward Reference Churchward1953: 4–5). That is, there is a moraic trochee at the end of each phonological word (Anderson & Otsuka Reference Anderson and Otsuka2006). The words in (2) illustrate these generalisations. We indicate foot boundaries with parentheses.

1. (2)

   

Secondary stress has also been described (Churchward Reference Churchward1953, Feldman Reference Feldman1978, Schütz Reference Schütz2001). Feldman states that ‘secondary stress falls on every second vowel from the end of each morpheme’ (Reference Feldman1978: 134), unless the next vowel has primary stress (e.g. in [ˌfaka-aˈŋa-hi] ‘criticise’ the final vowel of the root /aŋa/ receives primary stress because it is penultimate in the word, and therefore the penultimate vowel of /aŋa/ does not receive secondary stress). Feldman's examples illustrate ways that his generalisation cashes out, as shown in (3).

1. (3)

   

Long monomorphemes are rare and often contain reduplication-like repetitions or strings that exist as affixes in the language (/taki/ and /maa/, in (3a), are affixes that form verbs or adjectives with unpredictable meaning change; Churchward Reference Churchward1953: 243–244, 259–260).

Native Tongan vocabulary is silent on secondary-foot alignment: in hypothetical five-mora /tatatatata/, should there be an ‘initial dactyl’ (Prince Reference Prince1983, Hayes Reference Hayes1995), as in [(ˌtata)ta(ˈtata)], or should all feet be aligned to the right, as in [ta(ˌtata)(ˈtata)]? We failed to find convincingly monomorphemic native words of this shape (or ones with seven moras). Because of epenthesis, monomorphemic loans from English offer many five-mora cases and longer, and as we will see, both stress patterns occur.

Our purpose in this section is to investigate what determines whether an English loan is stressed as [(ˌtata)ta(ˈtata)] or [ta(ˌtata)(ˈtata)]. We show that secondary stress avoids falling on an epenthetic vowel.

3.2 Materials and methods

Churchward's (Reference Churchward1959) dictionary marks loans, and some additional words that Churchward and Tongan-speaking collaborator Feleti Vī thought the language might need to adopt, such as /palafini/ paraffin. We entered all these loans and suggested loans into a spreadsheet.

In each session with a consultant, we went through a few dozen words to check that the word was familiar, find out the consultant's preferred spelling and check the gloss. Words deemed familiar were then recorded in a quiet room using Praat (Boersma & Weenink Reference Boersma and Weenink2006) and a Logitech USB desktop microphone. Words were recorded in citation form, with two repetitions each. If both tokens were untranscribable because of hesitation, too much creaky voicing or other problems, that word was re-recorded in a subsequent session. We recorded about 500 loans in total; after the various exclusions discussed below, 134 were used to investigate secondary stress.

To objectively transcribe secondary stress, we first considered the realisation of primary stress. Vicenik & Kuo (Reference Vicenik and Kuo2010), working with the same consultants as we did, found that the primary-stressed syllable was usually associated with a pitch rise that begins around the beginning of the syllable and peaks around the end. We see this in Fig. 1 in a token of [(ˌɁas)toloo(ˈnoma)] astronomer from Speaker 2, where pitch rises over the primary-stressed syllable [no].

Figure 1

Pitch track of [(ˌɁas)toloo(ˈnoma)] astronomer (Speaker 2). Primary stress is boxed. All pitch tracks, spectrograms and waveforms were made using Praat. The frequency range of the spectrogram superimposed for reference is 0 to 5000 Hz.

Preliminary investigation found that there was usually a similar pitch rise on syllables where we heard secondary stress, as can be seen at the beginning of astronomer. (This token is also typical in having no clear sign of a tertiary stress, even though there is room for another foot in the middle of the word. Clear cues to tertiary stresses were rare, and we did not attempt to analyse it.) We therefore took rising pitch followed by a fall to indicate secondary stress on the syllable with the rising pitch. Using citation forms, rather than frame sentences, maximised the chance that secondary stress would display this cue. Doubtless there are other cues, but they are weaker and less reliable: Garellek & White (Reference Garellek and White2015) found that fundamental frequency, RMS energy, duration, F1 and voice quality were all reliably different between primary stressed and unstressed vowels in Tongan. But looking at secondary stress vs. no stress (in four-mora words, where the position of secondary stress is uncontroversial), they found that only fundamental frequency was significantly different for all five vowels.

In Fig. 2a we see secondary stress on the first syllable: pitch rises on [mo], then falls on [no]. (The overall pitch of this token declines, so that the F0 peak of the secondary stress in [mo] is higher than that of the primary stress in [la].) In Fig. 2b, by contrast, we see secondary stress on the second syllable: pitch rises on [lo], then falls on the next syllable, [ni].

Figure 2

Pitch tracks of (a) [(ˌmono)ka(ˈlame)] monogram (Speaker 1) and (b) [ko(ˌloni)(ˈtini)] quarantine (Speaker 2). Secondary and primary stresses are boxed.

The three authors independently transcribed each token. There were tokens with a flat pitch contour over the first two syllables and thus no discernable secondary stress according to our criteria. Such tokens were not used in the analysis below. There were also tokens where one of the first two vowels was deleted, such as the one shown in [ˌɁas_tolooˈnoma] in Fig. 1 above, from /Ɂasitoloonoma/; we coded the deleted vowel as unstressed, but the coding of stress on the other vowel still depended on whether it bore a pitch rise. In our original set of 306 tokens that meet the further criteria set out below, we disagreed on 19 (6%), in almost every case because one or two transcribers detected no clear secondary stress on either of the first two vowels, while the rest did detect some. There were only three cases where we transcribed opposite secondary stress; these were excluded. For the remaining 16 items, we retranscribed, and went with majority rule if there was still disagreement (two items were transcribed by only two researchers).

Once we had our speakers’ pronunciations of all the words, we excluded certain word shapes. (Because speakers often disagreed slightly with the dictionary form, we would not want to exclude items a priori on the basis of their dictionary form.) We included only words of the form #(C)V₁(C)V₂(C)V₃…V́(C)V#, such as /Ɂolokaholo/ alcohol. Requiring at least one vowel between V₂ and the main stress means that there is enough room before the primary stress for either V₁ or V₂ to be stressed: actual [(ˌɁolo)ka(ˈholo)] and hypothetical [Ɂo(ˌloka)(ˈholo)] are both well footed. In shorter words like /Ɂamelali/ admiral, there is only one place for a non-final foot, [(ˌɁame)(ˈlali)]; such words tells us nothing about subtler footing preferences, and are excluded.

We placed additional restrictions on vowel sequences, because of additional stress generalisations we noticed. We often observed ‘broken’ vowels (Poser Reference Poser1985, Mester Reference Mester1992) like [eˈe], with stress on the second half, in a word's last foot ([meˈesi] mercy), or the last foot of an obvious compound member ([haˌafe-kalaˈuni] half-crown). But we did not observe broken vowels earlier in a word: we transcribed 43 tokens like [(ˌɁaa)ke(ˈtika)] architect, and none like *[Ɂa(ˌake)(ˈtika)]. Even when neither half of the sequence was stressed, it was rare for a foot boundary to split it: we observed 48 tokens like [pa(ˌloo)(ˈmesi)] promise and just two like [(ˌseva)a(ˈniti)] servant, 53 tokens like [(ˌɁele)(faa)(ˈnite)] (with no clear tertiary stress, but at least the possibility that the two /a/’s are footed together) and none like *[Ɂe(ˌlefa)a(ˈnite)]. There must be a penalty against […a(ˈa…] and […a)a…].Footnote ⁴ Therefore, if any two of the first four vowels of a word were identical and immediately adjacent, we excluded the word (e.g. /Ɂaaketika, paloomesi, elefaanite/).

For similar reasons, we did not allow two adjacent vowels (in the word's first four vowels) to be of rising height; i.e. /ai ae au ao ei eu oi ou/ (e.g. /tainamiki/ dynamics, /kalaimeti/ climate, /Ɂeponaite/ ebonite) were excluded. This is because sequences of rising height may have special stress behaviour too. Churchward (Reference Churchward1953: 4) claims that in a word like /tauhi/ ‘to keep’, ‘stress may fall either on the u or on au as a whole’, and similarly for words with ao, ai or ei; Schütz (Reference Schütz2001: 319–321) suggests that the variation for these ‘potential diphthongs’ is conditioned by phrase position and speech style and rate. Our transcription of primary stress in words ending with sequences like /…auCV/ was often uncertain. Garellek & White (Reference Garellek and White2010) find acoustic evidence that VV sequences with rising height tend to belong to a single syllable, while VV sequences with falling height tend to belong to two separate syllables. We might therefore expect that *[ta(ˌina)(ˈmiki)] is ill-formed because it splits an /ai/ sequence into separate syllables. This prediction was also borne out in our data: we found 36 tokens like [(ˌtai)na(ˈmiki)] and only six like [po(ˌini)se(ˈtia)] poinsettia, seven like [ka(ˌlai)(ˈmeti)] and four like [(ˌtisa)i(ˈpele)] disciple, 24 like [(ˌɁepo)ˈnaite] or [(ˌɁepo)naˈite] (leaving open where the main-stress foot boundaries would be) and only one like [si(ˌkala)(ˈipe)] scribe.

Occasionally, a speaker's two repetitions of a word differed in meeting the criteria above, such as [(ˌsee)keli(ˈtali), (ˌseke)li(ˈtali)] secretary, where only the second pronunciation meets our criteria. We take it that there is variation or uncertainty in the underlying form, and use the repetition that meets our criteria. There were also five cases which we simply excluded, whose repetitions differed purely in the location of secondary stress, such as [(ˌkoni)sinaa(ˈnite), ko(ˌnisi)naa(ˈnite)] consonant.

We excluded four additional categories. First, we excluded probable compounds, such as /simolo-pookisi/ smallpox, because they should contain two separate stress domains. Second, we excluded a few religious words, such as /Ɂevaŋelioo/ ‘evangel, gospel’, that could not be straightforward adaptations from English (what would be the source of the final /oo/?). The source might be another language, perhaps as pronounced by English-speaking missionaries, so that we were unsure of the source stress. Third, we excluded two tokens produced with exceptional antepenultimate stress, because there were not enough vowels before the primary stress to allow both options for secondary stress: [ˌsemeneeˈlio ~ ˌsemeˈnelio] ‘seminary’ (excluded anyway because the final [o] makes the source obscure), and [ˌɁeleˈvasio] elevation.Footnote ⁵ Fourth, we excluded words that begin with /i/ or /u/ and another vowel, because they tend to become glides: /uesiliana/ [(ˌwesi)li(ˈana)] Wesleyan; we also excluded one word with a high vowel between two other vowels, /noauee/ Norway, because after gliding there were only four surface vowels: [(ˌnoa)(ˈwee)].

3.3 Results

All this culling left us with 294 tokens from the three speakers, covering 134 types. Of these, 189 (64%) have stress on V₁ and 105 (36%) have stress on V₂. The correspondent of V₁ or V₂ in the English source can be a stressed vowel, an unstressed vowel or no vowel at all. We want to know if this affects where secondary stress falls.

We give the raw results in Table I, in which the percentage of words in each category with secondary stress on V₁ is shown; remaining words have stress on V₂. An example word is given for each stress pattern. Unfortunately, some of the cells are empty – for example, there were no qualifying items where both V₁ and V₂ are epenthetic – but we can still make some direct comparisons.

Table I

Percentage of items stressed on V₁. Columns give the status of V₁; rows the status of V₂.

Comparing the cells in the bottom row, where V₂ corresponds to a stressed English vowel, we see that it matters whether V₁ is epenthetic (words like /falakiseni/ fraction) or corresponds to an unstressed English vowel (/Ɂapenitiki/ appendix). Epenthetic V₁'s are stressed only 26% of the time, as compared to 66% for non-epenthetic V₁'s.

In the last two columns we can also make direct comparisons along the V₂ dimension. V₁ is stressed more often when V₂ is epenthetic (words like /Ɂasitoloonoma/ astronomer (87%) and /kaŋikaluu/ kangaroo (83%)) than when it is not (words like /Ɂapenitiki/ appendix (66%) and /monokalame/ monogram (69%)).

Are these differences significant? The cells of the table may be unbalanced with respect to which speaker the data come from, and the qualities of V₁ and V₂ (/a e i o u/). Therefore, we used regression to control for these factors.

The logistic regression model shown in Table II below was fitted using the bayesGLM() function of the arm package (Gelman et al. Reference Gelman, Su, Yajima and Hill2010) in R (R Core Team 2014). This function has the advantage of handling the ‘separation’ found in the data (cases where some combination of independent variables completely predicts the outcome, such as V₁ status = epenthetic and V₂ status = unstressed).

Table II

Logistic regression model for secondary stress. Dependent variable is whether stress falls on V₁.

The dependent variable is whether stress falls on V₁ (as opposed to V₂); variables that promote stress on V₁ have positive coefficients, and those that promote stress on V₂ have negative coefficients. The independent variable of interest encodes the status of the first two vowels: rather than treat the rows and columns of Table I as two separate variables, with a lack of data for certain combinations of values, we use a single variable ‘V status’, with values like ‘V₁ is epenthetic, V₂ corresponds to stressed’. We excluded blancmange, to avoid having only one token whose value of this variable is ‘V₁ is epenthetic, V₂ corresponds to unstressed’.

Four independent variables were included to control for possible skews in the data. First is Speaker, in case speakers have different baseline rates of stressing V₁.Footnote ⁶ Second is a binary variable encoding whether the word begins with a possible English cluster. The idea is that in a word like /kolonitini/ quarantine, even though the first vowel is not epenthetic, speakers might have formed a generalisation, based on other loans, that a vowel in the environment #k_l tends to be epenthetic, and thus might treat it differently. If the first two consonants in the word were /p-l, f-l, t-l,Footnote ⁷ k-l, s-l, s-m, s-n, s-p, s-t, s-k/, this variable was coded as ‘yes’. Third and fourth are vowel quality for V₁ and V₂, each with five levels: /a e o i u/, where /a/ is the reference level.

We used R's step() function for model selection, which tries adding and omitting various factors, seeking improvement in the Akaike Information Criterion. The function started from a model with all the variables listed above, plus an interaction between speaker and every other variable. The best model found is shown in Table II.

The intercept represents the predictions for the default (reference) level of all independent variables, and the coefficients for the other factors are added to the intercept when relevant. For an example like /kasitomaa/ customer uttered by Speaker 1, Speaker and V₁ have their default values, and V status has the value ‘V₁ corresponds to stress, V₂ is epenthetic’, so we add the intercept, ―1.110, and the coefficient, 2.397. The model's predicted rate of stressing V₁ is 1/(1 + e^{−(−1.110+2.397)}) = 78% (the actual rate for the 13 such words is 69%).

Stepping through the coefficients, we see that Speaker 3 stresses V₁ significantly more often than Speaker 1 (the reference level). Quality of V₁ has some effect: /i/ is stressed significantly less often than /a/, and /e/ significantly more often than /a/.Footnote ⁸ Turning to the factor of interest, V status, we see that every other value has a higher rate of stressing V₁ than does the reference level, where V₁ is epenthetic and V₂ corresponds to a stressed English vowel. Using the function glht() from the multcomp package (Hothorn et al. Reference Hothorn, Bretz and Westfall2008), we find that none of the other pairwise differences between levels of this variable are significant. In terms of minimal comparisons within a row or column of the table in Table I, where only one vowel's status is changed at a time, this means there is a significant difference between words like ˈp_roblem and poˈtassium – that is, between an epenthetic and non-epenthetic V₁.

3.4 Discussion

The secondary stress data we have just seen indicate that Tongan speakers are sensitive to the English contrast between #CC and #CVC, such as predicate [paˌletiˈkasi] vs. paragraph [ˌpalakaˈlafi], even though there is no such contrast among native words: at least before contact with English, speakers did not have experience perceiving near-minimal pairs analogous to English please and police.

Tongan has much in common with Japanese, where perception of the CC/CVC difference has been studied in detail. Like Japanese, Tongan has extensive vowel devoicing and even some deletion (see §4), so that an English CC sequence might actually be close to a Tongan /…CVC…/ sequence that has undergone vowel deletion. In other words, the Tongan ‘V’ category in the context C_C includes a range from Ø to V, and crucially does not contrast with Ø. This should make perceiving the difference between Ø and V even more challenging in this environment: if Ø could be perceived as an extremely poor exemplar of V, the difference might be easier to perceive, but if Ø is actually a reasonably common exemplar of V, then we expect the perceptual space to be warped, such that Ø and V are perceived as more similar (e.g. Iverson & Kuhl Reference Iverson and Kuhl1996, where English speakers perceived the distance between two good tokens of /l/ as smaller than the distance between a good and a bad token; and similar results for English /i/ in Iverson & Kuhl Reference Iverson and Kuhl1995).

As mentioned in the introduction, the Japanese speakers studied by Dupoux, Kakehi et al. (Reference Dupoux, Fushimi, Kakehi and Mehler1999), even more advanced learners of English or French included in one experiment, had difficulty distinguishing CC and CVC, though they performed above chance in an ABX task. Peperkamp & Dupoux (Reference Peperkamp, Dupoux, Solé, Recasens and Romero2003) suggest that loan adaptation is difficult enough that perceptually driven nativisations like vowel insertion will happen consistently. By contrast, Kubozono (Reference Kubozono2006) argues that English CC and CVC are treated in subtly different ways in Japanese loan adaptation, and this is what we seem to see in Tongan.

4.1 Methods

The items for this analysis were taken from the same recordings as the items in the secondary stress analysis, but form a different (overlapping) subset. Nearly all qualifying deletions (using the criteria immediately below) were of /i/, so we limited our analysis to that vowel (there were five qualifying deletions of /e/, and none of other vowels). We chose items with at least one unstressed /i/, according to the consultant's spelling. To make sure that the /i/ was unstressed, the preceding or following vowel, or both, had to have been transcribed as stressed. Because final /i/’s are almost all epenthetic, we excluded them. We also excluded /i/’s in the first syllable, because whether or not they are epenthetic is strongly confounded with the surrounding consonants; for example, if the surrounding consonants are /s-k/, /s-t/ or /s-p/, the /i/ is almost always epenthetic. Therefore, we used only words with the patterns in Table III.

Table III

Tokens selected for vowel-deletion analysis.

There were 22 items with two qualifying /i/’s, such as [saˌkilifiˈsio] sacrifice.Footnote ⁹

Looking at the waveforms and spectrograms of these items, there was a clear difference in the recordings between deleted and non-deleted /i/’s, so only one author did the additional codings needed for this section. An example of the difference between deleted and non-deleted /i/ is shown in Fig. 3, for two tokens of /amipasitoa/, produced in immediate succession.

Figure 3

Pitch tracks of two tokens of /amipasitoa/ ambassador (Speaker 2).

Although word-final devoiced /i/ was common, devoiced /i/ was rare in our target contexts,Footnote ¹⁰ and we grouped it with deletion, in line with the analysis presented below in §6.1. If deletion occurred in only one of the repetitions, we coded the item as displaying deletion, with the idea that what is of interest is whether a vowel can delete (there were nine such items; the supplementary materials include versions of the regression model below in which these items are instead excluded or treated as non-deleted).

4.2 Results and discussion

From the raw data in Table IV, it is clear that epenthetic /i/ is more likely to delete, although the three speakers have different global rates of deletion. Unfortunately, there are not enough items in which unstressed word-medial /i/ corresponds to a stressed vowel in English to allow us to establish whether stress in English matters.

Table IV

Percentages of /i/'s that delete, for each speaker and each type of /i/. The bottom two rows give examples of each /i/ type with and without deletion.

The plots in Fig. 4 show the same data.

To test significance, we fitted a mixed-effects regression model, using the glmer() function in R's lme4 package (Bates et al. Reference Bates, Maechler, Bolker and Walker2014), shown in Table V. The dependent variable is whether /i/ is deleted. The independent variable of interest is whether Tongan /i/ is epenthetic or corresponds to an unstressed English vowel – we excluded the small number of cases where /i/ corresponds to a stressed English vowel. Speaker is treated as a fixed-effect independent variable with three levels. We included an independent variable coding whether the target vowel would be footed (as in [(ˌɁosi)ta(ˈlesi)]), or not (as in [(ˌlave)ni(ˈtaa)]). Finally, consonantal environment is treated as a random effect (not shown in Table V); in our data it has 36 levels, such as ‘s-t’ and ‘n-t’.

Figure 4

Deletion (light grey) and non-deletion (dark grey) for each speaker. The width of each column for each speaker (stressed (left), unstressed (middle), epenthetic (right)) reflects the proportion of cases for that speaker that come from the relevant category. The circles at the bottoms of the ‘stressed’ columns for all three speakers indicate zero deleted /i/ in this category.

Table V

Logistic regression regression model for vowel deletion. Consonantal environment (not shown) is random effect. The independent variable of interest is V status.

The intercept represents the case where the speaker is Speaker 1, the target vowel is not footed and the target vowel is not epenthetic (in an average consonantal context). To calculate the model's predicted rate of /i/-deletion in that case, take 1/(1 + e^{−(−2.927)}) = 5%. As we saw in Table IV, the speakers differ in their baseline rates of deletion, and this is significant in the model: Speaker 2 has a significantly lower rate of deletion than Speaker 1, and Speaker 3 a significantly higher rate of deletion than Speaker 1.

As for our variable of interest, if the target vowel is epenthetic, it has a significantly higher rate of deleting. The supplementary materials include comparison with an inferior model with a non-significant interaction between Speaker and V status. The lack of significant interaction means there is no evidence that the three speakers had different degrees of sensitivity to the English pronunciation.

As in the case of secondary stress, Tongan speakers are showing sensitivity to whether a vowel in a loan is epenthetic. In the secondary stress case, we saw that Tongan speakers were distinguishing CC from CVC near the beginning of a word, though native Tongan words have no such contrast. Here, we see that they are also distinguishing CC from CVC further into the word.

5.1 Background and methods

Tongan can be analysed as having either a contrast between short and long vowels (Churchward Reference Churchward1953, Poser Reference Poser1985) or sequences of vowels, in separate syllables, which sometimes happen to be identical (Feldman Reference Feldman1978, Schütz Reference Schütz2001, Taumoefolau Reference Taumoefolau2002, Anderson & Otsuka Reference Anderson and Otsuka2006). The examples in (5) show transcriptions under each view. Transcriptions differ when stress falls on the first half of the vowel (sequence); [ˈaː] vs. [ˈa.a]; when stress falls on the second half, both views treat the sequence as disyllabic: [a.ˈa]. In this paper we mostly use the ‘VV’ transcription style.

1. (5)

   

English vowels can be adapted as short or long, probably partly determined by the vowel's duration in English. English stressed and/or tense vowels, being longer than unstressed or lax vowels,Footnote ¹¹ tend to be perceived by Tongan speakers as long. For example, the long vowel in [(ˌkaŋi)ka-(ˈluu)] kangaroo is both stressed and tense in English ([ˌkæŋgəˈɹuː]).

But unstressed English schwa can be adapted as either long or short, as in [(ˌheli)kope(ˈtaa)] and [ta(ˌlani)si(ˈmita)], from English [ˈhεlɪˌkɒptə] and [ˌtrɑːnsˈmɪtə]. It seems unlikely that there is much length difference in the final vowels of the two English words. The important difference is, as we demonstrate in this section, that in transmitter, the final short vowel causes primary stress to fall on the preceding syllable's /i/, which corresponds to English [ˈɪ]; in helicopter, by contrast, a final short vowel, *[(ˌheli)ko(ˈpeta)], would cause stress to fall on /e/, which has no English correspondent. Schütz (Reference Schütz, Wurm and Laycock1970) calls this tendency:

Schütz notes that this phenomenon is not regular enough to write a rule for it.

Kenstowicz (Reference Kenstowicz2007) analyses vowel length in Fijian loans from English (using Schütz's Reference Schütz and Schütz1978 corpus), and finds something very similar. Fijian has a similar prosodic system to Tongan, with contrastive length and a moraic trochee at the right edge of the word. Kenstowicz shows that length is deployed to achieve a better match between the prosody of the Fijian and the English forms, as listed in (6) (forms from Kenstowicz Reference Kenstowicz2007: 319, 324).

1. (6)

   

This section shows that both Schütz's and Kenstowicz's observations on Tongan and Fijian epenthetic penults also hold in our corpus of Tongan loans. We further show that final vowel length is sensitive to the English stressed/unstressed distinction: even when the penult is not epenthetic, final vowel length is more likely if the penult was unstressed in English than if it was stressed. Thus, the vowel-length data provide (further) evidence that Tongan speakers perceive both the difference between English CVC and CC and the difference between English stressed and unstressed vowels.

5.2 Results

We took from our recorded corpus all those words where the final Tongan vowel corresponds to an unstressed vowel in English, and analysed the length of that final vowel. There were 114 relevant tokens. Looking at the raw numbers, we see a clear three-way distinction, as plotted in Fig. 5. If the preceding vowel has no correspondent in English (i.e. is epenthetic), the final vowel was adapted as long in 75% of the cases. If the preceding vowel corresponds to an unstressed vowel it was adapted as long in 35% of the cases, and if it corresponds to a stressed vowel it was never adapted as long.

Figure 5

Rate of adaptation with final long vowel for English words that end in an unstressed vowel (the circle at bottom of the third column represents token count of zero).

Results are similar for all three speakers, as shown in Table VI, where we also see that the results are similar whether the final unstressed English vowel is [ə] (72 tokens) or a tense vowel ([oʊ, i]; 42 tokens).

Table VI

Breakdown of Fig. 5, showing how often final V is long, for each speaker and for English schwa vs. tense vowels.

To test the significance of the status of the penult, we used a logistic regression model, shown in Table VII. Because there are no random effects in this case, we were once again able to use the function bayesGLM(), which can handle the zero rates of lengthening for stressed English penults. The binary dependent variable is whether the word is adapted with a long final vowel (rather than a short one). The independent variable of interest is the status of the preceding vowel.Footnote ¹²

Table VII

Logistic regression model for final V length. Dependent variable is whether final V is adapted as long (vs. short).

The model shows that, compared to the baseline case where the preceding vowel corresponds to an unstressed English vowel, a long final vowel is significantly more likely when the preceding vowel is epenthetic (positive coefficient), and significantly less likely when the preceding vowel corresponds to a stressed English vowel (negative coefficient). That is, final vowel length is indeed favoured if it avoids stress on a preceding vowel that corresponds to an English unstressed vowel, and even more if it avoids stress on a preceding epenthetic vowel.

5.3 Discussion

In this case, Tongan speakers are making a three-way distinction among English CV́C/CV̀C, CVC and CC, even though none of those distinctions is contrastive in Tongan.

What is more surprising is that the length effects are seen not on the vowel that itself is stressed, unstressed or absent in English, but on the subsequent vowel. It's not just that stressed vowels are more likely than unstressed to be perceived as long, or that actual vowels are more likely to be perceived as long vowels than C-C transitions are. Rather, a subsequent vowel is being treated as long if doing so prevents stress on a vowel whose English correspondent is unstressed or non-existent.

6.1 Vowel deletion

We saw in §4 that unstressed /i/ is more likely to be deleted when epenthetic with respect to English. For example, we observed undeleted /i/ in [paˌlesiˈteni] president and [ˌmekesiˈkoo] Mexico, but deleted /i/ in some tokens of [kaˌlisˈtala] crys_tal and [ˌminisˈtaa] minis_ter. Overall, non-epenthetic /i/ deleted 14% of the time, and epenthetic /i/ 46% of the time.

We pursue an analysis with an underlying difference between full vowels and weak vowels, as in /palesi_Fteni_W/ vs. /kalisi_Wtala/, including analysis of how speakers end up with these underlying forms.

We adopt Boersma & Hamann's (Reference Boersma, Hamann, Calabrese and Wetzels2009) multilevel model of loan adaptation, where the grammar regulates a series of mappings, as listed in (7). An English utterance like [ˈkɹɪstəl] must first be mapped by the listener onto a perceived surface form, which is the same type of object as a produced surface form, including stress and foot structure: e.g. /ka_W(ˌli_Fsi_W)(ˈta_Fla_W)/. (In Boersma & Hamann's notation, phonological surface forms are given between / /, and underlying forms between | |.) The perception grammar governing this mapping is the same one used for native-language perception. To arrive at an underlying form – whether for a known word or, in the case of initial adaptation, a new word whose underlying form must be constructed – a recognition grammar is used, again the same one as for native words. When the listener wishes to say the word, a production grammar maps the underlying form to a produced surface form, and a phonetic implementation grammar maps this surface form into articulation, perhaps realising /i_W/ as overlapped or reduced to the point of deletion, as in the example in (7).

1. (7)

   

Nowhere in this model is there any special mechanism for loans, although some constraints may be more relevant for loans than for non-loans. We combine all of these mappings into a single grammar. The grammar includes markedness constraints that penalise structures in surface forms – applicable to both perception and production – and constraints regulating mappings between each adjacent pair of levels.

Weak vowels can occur in both underlying and surface forms. The full/weak distinction can be thought of as a feature, like voiced/voiceless. Unlike the excrescent or intrusive vowels investigated by Hall (Reference Hall2003, Reference Hall2006), which originate as gaps between consonant constrictions, Tongan weak vowels do count for footing and stress. (If they didn't, then feet like (ˈte_Fni_W) would be problematic – the /n/ would be a coda consonant. An analysis along these lines is possible, but would be different from ours, where NoCoda is undominated.) The idea, however, is similar: there is a perceptual continuum from zero to a full vowel, and the zero in an illegal consonant sequence will tend to be perceived as some kind of vowel (Hall Reference Hall2006: 9, Davidson Reference Davidson2007, Davidson & Shaw Reference Davidson and Shaw2012, Broselow Reference Broselow, Hsiao and Wee2015).

In our analysis, the difference between epenthetic and non-epenthetic vowels arises in the perceptual mapping. Tongan listeners hear instances of consonants not followed by a full vowel – CC, CV̥C, C# and CV̥# – in native words, loanwords and English speech. As shown in (8) (where full/weak status is shown only for the relevant vowel), under a simplified scenario with no variation, high-ranking /NoCoda/ and */Complex/ (Prince & Smolensky Reference Prince and Smolensky1993) force these consonant transitions, consonant releases and devoiced vowels to be perceived as vowels (similar to Boersma & Hamann's Reference Boersma, Hamann, Calabrese and Wetzels2009 analysis of epenthesis in Korean loans). Under the ranking shown here, perceiving such sounds as weak vowels is preferable to perceiving them as full vowels.Footnote ¹³

1. (8)

   

The term ‘consonant release or transition’ covers a lot of ground, from a stop's release burst, as in [pɹ], to a sudden decrease in amplitude, as in [st] or [l#]. In a more detailed model, constraints would distinguish these cases. However, we did not find systematic effects of consonant sequence, so our analysis does not break down release/transition types.Footnote ¹⁴

When there is a devoiced vowel, the weak vowel's perceived quality is straightforward; when there is no vowel, as in crystal, the weak vowel's quality will be whatever is perceived as closest to the consonant transition (see Uffmann Reference Uffmann2007 on epenthetic loan vowels).

Assuming that the mapping from perceived surface form to lexical entry is faithful, we then have underlying |palesi_Fteni| and |kalisi_Wtala|. In production, constraints against weakening underlying full vowels and vice versa maintain the difference, as in (9).

1. (9)

   

Finally comes phonetic implementation, where weak vowels are pronounced with gestures that are reduced in magnitude and/or duration, leading to devoicing, overlap and deletion. Again abstracting away from variation, we assume that the /i_F/ of president is realised as a full [i], and the /i_W/ of crystal is deleted, as in (10).

1. (10)

   

Weak vowels thus contrast with full vowels: both can appear in underlying forms, while faithfulness constraints discriminate between them. But the contrast is marginal, in that the faithfulness constraints are too weak to produce stably different pronunciations, and instead only affect variation rates.

To model variation, we fitted a Maximum Entropy constraint grammar (Goldwater & Johnson Reference Goldwater, Johnson, Spenader, Eriksson and Dahl2003, Jäger Reference Jäger, Zaenen, Simpson, King, Grimshaw, Maling and Manning2007, Hayes & Wilson Reference Hayes and Wilson2008), shown in (11), using the MaxEnt Grammar Tool (Hayes et al. Reference Hayes, Wilson and George2009).Footnote ¹⁵ Each constraint has a weight, and a candidate's probability is proportional to the weighted sum of its constraint violations, exponentiated. For example, for the candidates in the final block of (11a), the weighted constraint violation sums are 4.6 and 6.4; negated and exponentiated these are e^−4.6 = 0.0101 and e^−6.4 = 0.0016; the first candidate's proportion of the total, 0.0101/0.0016 = 0.86, is its probability. The last two columns show the fitted model's predicted probability, and the rate observed in our data across all three speakers. The grammar was fitted to the full data – vowel deletion, secondary stress and vowel length – but only the relevant constraints are shown here. The full grammar is provided in the supplementary materials. (The constraints /NoCoda/ and */Complex/ are omitted, on the assumption that they have such high weights as to exclude any candidates that violate them.)

1. (11)

   

The tableaux in (11) illustrate the mappings from auditory form to surface form to underlying form, and back to produced surface form. An additional markedness constraint, */UnstressedV_F/, gives even underlying full vowels a chance of being produced as weak when unstressed. Because president and crystal are only representatives of groups of words, we ignore both the status of vowels in the word other than the target and constraint violations that those other vowels incur.

The challenge to fitting a multilevel model is that variation could occur at every level. That is, a produced surface form /ka(ˌli.si_W)(ˈta.la)/ might come from underlying |kalisi_Wtala| or |kalisi_Ftala|, and each of those might have been lexicalised from perceived /ka(ˌli.si_W)(ˈta.la)/ or /ka(ˌli.si_F)(ˈta.la)/. All we can observe is the rate of vowel deletion/weakening, not the rates of variation at the intermediate stage. This is part of what Boersma (Reference Boersma, Benz and Mattausch2011) calls the problem of ‘whole-language simulation’.Footnote ¹⁶ We simplify by assuming that absent vowels in English are always perceived and lexicalised as weak vowels, and that the variation resides entirely in production. (Thus the learner is trained on rates of 1.00 and 0.00 for the first two steps of each mapping.) It is also a simplification to assume that weak vowels are deleted only at the final stage of phonetic implementation. All this simplification prevents us from achieving a close numerical fit, but recall that exact rates varied among the three speakers too (see Fig. 4). The model succeeds in capturing the difference between president, with a lower deletion rate, and crystal, with a higher rate.

6.2 Secondary stress

In §3, we saw that if the first vowel in a loan is epenthetic, it is less likely to bear secondary stress. When neither of the first two vowels in a loan that meets the shape criteria in §3.2 is epenthetic, secondary stress falls on the first vowel 68% of the time, as in [(ˌtemo)ka(ˈlati)] deˈmocracy. We therefore take initial stress as the default, illustrated in (14). Align constraints (McCarthy & Prince Reference McCarthy and Prince1993) govern foot placement: Align-R(ω,Ft) requires every prosodic word to end with a foot, and Align-L(ω,Ft) requires every word to begin with a foot.Footnote ²⁰ Countervailing Align-R(Ft,ω) requires all feet to align towards the right, but is less powerful (variation is modelled below).

1. (14)

   

Words like [faˌlakiˈseni] fraction, whose first vowel is epenthetic, depart from this pattern, with only 26% having initial secondary stress. In the analysis developed above in §6.1, English [ˈfɹækʃən] should be perceived as /fa_W(ˌlaki)(ˈseni)/, with a weak first vowel, whereas [dᵻˈmɒkrəsi] should be perceived as /(ˌte_Fmo)ka(ˈlasi)/.

The tableau in (15) illustrates the mapping from auditory to perceptual form, simplified to categorical outcomes. We add one new markedness constraint, */V́_W/, which forbids stressed weak vowels in surface forms. This is reminiscent of Kenstowicz's (Reference Kenstowicz2007) two-level constraint *ˈv, which penalises stressing an epenthetic vowel. The advantage of a markedness constraint like */V́_W/ is that it can regulate both perception and production. In production, it prevents a stressed vowel from undergoing reduction (and is therefore needed in native Tongan words too). In perception, it ensures that a reduced vowel is not perceived as stressed, which is also a reasonable strategy for native Tongan words.Footnote ²¹

1. (15)

   

The ranking in (15) allows Tongan speakers to perceive unpredictable placement of secondary-stressed feet. This is unnecessary for native monomorphemes in isolation, but a good strategy for comprehension of running speech, where word and morpheme boundaries are not known ahead of time, and stress is a cue to boundaries.

How will this translate into an underlying form in a multilevel-style comprehension system? Given what we have already, the simplest answer is that the weak vowels are lexicalised as such: |fa_Wlakiseni| vs. |te_Fmokalasi|. In production, */V́_W/ and *|V_W| → /V_F/ ensure that secondary stress continues to avoid the weak vowel in fraction.

We provide a MaxEnt model of this process. We simplify by assuming that variation happens in production rather than perception. We distinguish three cases: the first vowel is epenthetic (b_lancˈmange, ˈf_raction), the second vowel is epenthetic (aˈs_tronomer, ˌkan_gaˈroo) or neither is epenthetic (deˈmocracy, ˈmonoˌgram). For the sake of space, we show only the most important aspects of the model. (The full model of all three case studies is in the supplementary materials.) High-weighted *[release] → /V_F/ ensures that consonant releases are perceived as weak vowels, and high-weighted *[V_F] → /V_W/ ensures that English vowels are perceived as full. At the recognition stage, high-weighted */V_F/ → |V_W| and */V_W/ → |V_F| ensure that these decisions about weakness are lexicalised.

The tableaux in (16) show the variation that then occurs in production. We consider four candidates for each word: with either peninitial stress or initial stress on a full vowel, and the other, unstressed, vowel weak or full. Under our assumptions about where in the grammar vowel deletion occurs, a full underlying vowel has a 16% chance of being produced weak, and a weak underlying vowel a 46% chance. We divide up the stress rates accordingly: for example, the 84% observed rate of initial stress for kan_garoo-type words (second vowel weak: |ka_Fŋi_Wkaluu|) becomes a trained rate of 0.46 × 0.84 = 0.39 for /(ˌka_Fŋi_W)ka(ˈluu)/ and 0.54 × 0.84 = 0.45 for /(ˌka_Fŋi_F)ka(ˈluu)/. To assess the success of the model, we show the combined model probabilities and combined trained rates for the two peninitial-stress candidates and the two initial-stress candidates.Footnote ²²

1. (16)

   

The model is successful in capturing the three-way distinction: kan_garoo has the highest rate of initial secondary stress, followed by democracy, followed by f_raction, although the model's rate of initial stress in f_raction is much lower than the trained rate.

6.3 Vowel length

We saw in §5 that a final vowel was always adapted as short if this would place stress on a correspondent of a stressed English vowel ([ˌsepiˈtema] Sepˈtember), less often if it would place stress on the correspondent of an unstressed English vowel ([ˌtaiaˈmita] diˈameter, [ˌkasitoˈmaa] ˈcustomer) and even less often if it would place stress on an epenthetic vowel (*[ˌminiˈsita], rather than [minisiˈtaa] ˈminis_ter).

The analysis developed so far extends easily to these data. As shown in (18) (ignoring variation for now), the transition from [s] to [t] is perceived as a weak vowel, and the markedness constraint */V́_W/ prevents stress from falling on it. The remaining options are positing a non-binary foot (d), which is not allowed in Tongan, or perceiving the final vowel as long (f).

1. (18)

   

This is also a suitable perception strategy for native words: a devoiced or deleted penultimate vowel must be unstressed, and must therefore be followed by a long final vowel. The mapping from /minisi(ˈtaa)/ to underlying |minisitaa| is then straightforward: Tongan lacks productive vowel-length alternations, so long surface vowels should always be recognised as long underlying vowels.Footnote ²³

In order to capture the difference between Sepˈtember and ˈcustomer, we need a constraint sensitive to English stress. We propose one whose sensitivity is indirect: *[V_RelProm] → /Unstressed/, which requires a vowel in the auditory form that is more prominent than either of its within-word neighbours to be perceived as stressed. This constraint is inspired by Broselow's (Reference Broselow, Boersma and Hamann2009) FinalVLong (a final vowel more prominent than the preceding vowel should be recognised as long), but applies more broadly, to vowels anywhere in the word. Like Broselow (Reference Broselow, Boersma and Hamann2009: 217), we leave open what phonetic properties of an English word are heard as a ‘prominence peak’. English and Tongan stressed vowels have many of the same acoustic properties: higher duration, energy and pitch (depending on intonation), and lower F1 (see Garellek & White Reference Garellek and White2015 for Tongan, and references therein for English). Therefore, we can assume that, using native-language cues, Tongan speakers tend to perceive English stressed vowels as prominent.

Obeying this constraint is a reasonable perception strategy for Tongan overall: a vowel that is appreciably more prominent than either of its immediate neighbours within the same word must be stressed – either it is stressed and the neighbour is unstressed, or both are stressed but the vowel in question bears primary stress and its neighbour secondary.

As illustrated schematically (leaving customer in a tie) in (19), *[V_RelProm] → /Unstressed/ requires stressing the penult in September, because the [ˈε] that it corresponds to is more prominent than both flanking vowels in the auditory form. In customer, the constraint is irrelevant – all candidates satisfy it – so there is no need to ensure stress on the penult by having a short final vowel.

1. (19)

   

Our MaxEnt grammar's model of variation is shown in (20) (which evaluates only the last two vowels of each word). In our analyses above we simplified by confining variation to the production mapping, but vowel length in Tongan is contrastive, and not subject to either free variation or systematic alternation, so we instead assign variation in vowel length to perception.

1. (20)

   

The model captures the three-way distinction between epenthetic, unstressed and stressed.Footnote ²⁴