4.1. Subjective ratings of ‘prefixedness’

Rating is used as a gradient measure of degree of prefixedness. To address the issue of gradience of these word forms, objective measures can and have been previously used (e.g., Hay, 2007; Plag & Ben Hedia, 2018; Stein, 2023; Wurm, 1997), but examining gradience in this way is often complex and open to conflicting interpretations. We wanted a simple global measure that would reflect native speakers’ intuitive feelings about the extent to which each word was experienced as being prefixed or unprefixed. Accordingly, we used an online rating task and solicited the opinions of volunteers in NZ, Great Britain, USA, Canada and Australia via social media and a few professional mailing lists.

The questionnaire was run in Qualtrics (https://www.qualtrics.com/uk/). Participants rated how much the word seems to them to contain a prefix. Full instructions are given in Appendix B. To keep the rating task to about 20 minutes, 14 words were removed. These 14 were inflectional [s] variants, chosen for exclusion provided that the base word was represented in the ranking task, and that neither the base nor the inflected variant are ambiguous as to part of speech. Pilot work indicated that the addition of suffixes does affect ratings, but the above constraints on removing words in effect made minimal changes to the word’s meaning and maintained the same number of syllables for each word type. Excluded words were: disabilities, disagreements, disappears, disappointments, discoveries, discussions, disruptions, disrupts, distinctions, distractions, distributors, misdemeanours, misfortunes, mispronunciations. In later analyses, these words were assigned the mean rating given to their noninflected counterparts. This left 214 words, which are listed in Appendix C.

Each of the remaining 214 words was presented individually on the computer screen, centred above a numerical scale ranging in 11 steps of 10 from 0 on the left to 100 on the right. There was a horizontal slider below these numbers. Respondents used their mouse to slide a short vertical bar on the slider to the desired numerical value. Spatial configuration was not counterbalanced: 0 was always designated ‘Completely unprefixed’, and 100 always ‘Completely prefixed’. At the start of each trial, the slider was centered at 50, which was labelled ‘Neutral’ on the numerical scale. A response was required for each word. A horizontal bar near the bottom of the screen showed how much of the task had been completed. The task was self-paced: the participant clicked on a ‘Next’ button when ready to proceed to the next item. After rating all 214 words, respondents completed a short demographic questionnaire. This did not allow respondents to be identified, but did allow their responses to be sorted by regional accent of English and age.

After completing the questionnaire, the participant was informed on the screen that pressing the final ‘Submit’ button constituted consent for their data to be used (anonymously) for research purposes. Ethical approval was granted by the University of Canterbury, Christchurch, and by the Faculty of Music, University of Cambridge, UK.

Ratings were analysed from the 78 respondents deemed suitable: native speakers of English, defined as having English as their main language since infancy (with regional dialect immaterial). Mean degree of ‘felt prefixedness’ was calculated for each word within each geographical subgroup. Figure 1 shows density plots of the mean responses per word, with one panel and colour for each region. Responses to words classed as prefixed in Appendix C are shown with solid lines and to those classed as unprefixed are shown with dashed lines. For all regions, the distribution is roughly bimodal. There are many judgments close to 100% for words felt to be prefixed, whereas there seems to be less agreement about unprefixed words: there are relatively fewer judgments close to 0%, and the spread of the lower (more unprefixed) judgments is wider than that of the higher (prefixed) judgments. There also seems to be some regional variation in terms of how prefixed the words seem overall. In particular, the judgments from the United States skew more strongly toward prefixedness than those of the other regions, with a wide spread of ratings amongst unprefixed forms and a somewhat narrow spread amongst prefixed forms.

Figure 1

Density plots of mean word prefixedness Rating scores, by geographical group. Solid lines: ratings of prefixed forms. Dashed lines: unprefixed forms. Vertical lines under each curve show quartiles. Individual datapoints below x axes: circles = classed in Appendix C as prefixed; x = classed as unprefixed.

Despite the fact that the shape of the distribution of responses varies slightly across different regions, the rankings of the words were remarkably aligned. Figure 2 shows these same data as bar charts (undifferentiated by formal prefixed/unprefixed classification) separately for each geographical subgroup, together with Pearson r and Spearman r_s correlations across words between geographical subgroups. The correlation coefficients between different geographical groups are all positive and highly significant, with r_s ranging between .9 and .94, indicating significant consistency between each geographical subgroup’s mean ratings. This indicates that individuals are able to rank the words in terms of how prefixed they feel, and that the ranking is extremely consistent, even when compared across speakers of different dialects of English.

Figure 2

Barplots (diagonal): distributions of words’ mean ratings of ‘felt prefixedness’, grouped by respondents’ country of origin: 14 NZ respondents, 15 UK, 34 US, and 11 Other (comprising Canada (8), Australia (3), India (1)). Correlations between words’ mean ratings according to country of respondent are shown by scatterplots (right of diagonal) with r and r_s coefficients and associated p values left of diagonal.

For the purposes of predicting acoustic realization, it is clear that there is no need to use regionally specific ratings, as the ratings are very highly correlated across regions. The prefixedness ratings were therefore pooled across geographical groups to yield a single rated value for each word. The mean rating for each word is shown in Appendix C, together with its initial classification as prefixed or unprefixed. The prefixed/unprefixed classification assigned by our formal criteria corresponded almost perfectly with the mean ratings of felt prefixedness. Only one word that we formally classified as unprefixed, misnomer, gained a higher mean rating than words we classified as prefixed, and the overlap was small: the unprefixed mistaken (53%) and misnomer (65%) were separated by just five words we formally classed as prefixed, all of which are, like misnomer, semantically opaque to many native speakers: dismantled, disconcerting, discharging, misdemeanour, discharge. All subsequent analyses of gradience use this single rating scale.

We used simple tests to assess the roles of five variables: four of the predictor variables defined above: Frequency, word Stress (whether the syllable immediately following dis-mis bears primary lexical stress), BoundaryType (CodaPossible/CodaImpossible, further subdivided due to large differences in cell numbers (see Table S3.4) into whether phoneme 4 was a Consonant or Vowel), and whether the word begins with mis or dis (MisDis). The fifth variable, word length, contrasted two versus three or more syllables. Length could be included because our analyses of influences on ratings do not require that our ratings are independent of one another, making its aforementioned interdependency with word Stress less important.

We first examined the distribution of these five properties between the two morphological categories. All differ significantly between prefixed and unprefixed word types. Prefixed forms are significantly less frequent than unprefixed forms (Unpaired Wilcoxon test, W = 2801.5, p < .0001). Prefixed forms are also significantly more likely to have the primary stress on the third syllable or later (51% vs. 11%, chi-square = 35.99, p < .0001), to be three syllables or longer (49% of prefixed forms vs. 23% of unprefixed forms, chi-square = 8.323, p < .01), and to have a vowel following the [s] (47% vs. 6%, chi-square = 38.619, p < .0001). Frequency, early/late stress and number of syllables are presumably correlated: the addition of a prefix often (though not inevitably) results in a less common word, and inevitably increases the syllable count and shifts the main lexical stress one syllable later. However, there seems to be no a priori reason for stems to begin with vowels more often than the second syllable of unprefixed forms.

We then investigated whether these five properties also influenced ratings within rather than between each morphological category. Word frequency approaches significance within prefixed forms (Spearman’s r_s = –.16, p = .09), with lower frequency words likely to be rated as more prefixed. It is not a significant predictor of variation within unprefixed forms. Wilcoxon tests show that words with later lexical primary stress are more likely to be given higher ratings, within both prefixed forms (p < .001) and unprefixed forms (p < .01). Longer words are more likely to be given higher ratings within prefixed forms (p < .001), but not within unprefixed forms. And Consonant versus Vowel as the fourth phoneme is a significant predictor of ratings for prefixed forms (p < .02), with forms with a vowel following the [s] attracting higher ratings. Very few unprefixed forms have a vowel following the [s], so this factor cannot account for variation in the ratings among unprefixed forms.

Overall, a number of linguistic properties seem to be more characteristic of prefixed forms than of unprefixed forms. These properties also drive variation within these categories in terms of native speakers’ judgments of how prefixed each word form feels. Within each category, words appear to feel maximally prefixed when they most resemble other prefixed forms phonetically and phonologically, and minimally prefixed when they most resemble unprefixed forms. That is, native speakers’ judgments are at least somewhat gradient, in a way that reflects the phonological distributions in the language.

Of course, the factors analysed here are unlikely to be the sole ones driving this gradience. Semantic factors surely play an important role in affecting gradience, despite the fact that they are not formally analysed here. Semantic criteria were of course central to the formal binary classification, but there were nonetheless cases that were difficult to decide. The anomalous misnomer and dismantled discussed above are good examples of these, and their semantic indeterminacy is reflected in the subjective ratings. The mean word ratings derived from this task were a predictor variable—Ratings—in subsequent regression analyses.

5.1. Durations

Durations were measured only for the first three phonemes, /mɪs/ or /dɪs/. Smith et al. (2012) also measured the closure duration and VOT of the voiceless stop that was the fourth phoneme in all their words. But since in our data the fourth phoneme encompassed a wide range of vowels and consonants, and the vast majority of the tokens do not contain a following stop, closure duration and VOT are less useful in this study. We refer to the joint duration of the first three phonemes as triphone duration.

Duration measurements extracted were guided by Smith et al. (2012)’s findings: the absolute duration of the [m], [d], [ɪ], [s] and of the three segments together (i.e., the whole triphone). From these, relative durations of [s]:[ɪ] were calculated.

Some segmentation from HTK’s forced alignment is not optimal. To test whether this was likely to affect results, a small sample of the corpus was hand-corrected. The sample was 196 words in which the critical dis- or mis- is followed by a voiceless stop, for example discarded. Of these, 157 came from the NZE corpus, and 39 from the LIV corpora. (This sample was chosen because the original intention was to include measures of VOT in words where the fourth phoneme was a voiceless stop, and HTK segmentation does not distinguish stop closure from VOT. However, this plan was abandoned because there were too few such words in the corpus.) Hand correction reduced the variance in the data and brought mean durations between the first syllables of prefixed and unprefixed words slightly closer. However, in our initial investigations, hand correction made no difference to statistical results. Therefore, uncorrected HTK segmentation was used for all corpora except Buckeye.

5.2. Vowel formant frequencies

Smith et al. (2012) hand-corrected frequencies of the first three formants. For within-category comparisons of (unrounded) high front vowels, F3 frequency contributes little that cannot be inferred from the difference between F1 and F2 (F2 – F1), and automatic measures of F3 frequency tend to be unreliable. Consequently, it was decided not to try to measure F3.

Frequency measurements for F1 and F2 were made at the vowel midpoint, using Praat’s formant tracker (version 6.0.28, ‘To Formant (burg)…’) with 5 formants estimated below maxima of 5 kHz for men and 5.5 kHz for women, a 50 ms Gaussian window, pre-emphasis of +6 dB/octave, and a time step of 2.5 ms. From these midpoint frequencies, the frequency separation, F2 – F1, was calculated, to give a single measure of vowel quality for /ɪ/.

5.3. Spectral properties of [s]

Standard spectral measures of [s] (e.g., the centre of gravity (first spectral moment), and/or frequency of a spectral peak), calculated across the whole spectral range, can characterize differences between places of articulation. But we sought spectral measures that would be able to capture nuances within the spectral shape of [s] that might stem from stress- and/or duration-related production differences relevant to the prefix/nonprefix distinction. These could affect relative amplitudes and bandwidths in mid versus high frequencies, which could in turn affect perceived [s] loudness or quality.

Figure 3 illustrates Koenig’s (2013) measures, using the coding terms adopted in this paper (see also Table 3). These terms closely mirror those of Koenig but are typically shorter and without sub- or superscripts, to simplify graphs. The three frequency bands distinguished are shown with different shadings of grey under the spectral envelope. There were two measures of amplitude differences. One is the maximum amplitude difference between the two lowest frequency bands, s-ampDiff, applied to unique frequency-amplitude points shown by the vertical red line between the points within the red circles. The other is the relative amplitude over broad frequency ranges in the mid- and high-frequency bands, s-levelDiff, calculated as areaM – areaH in Figure 3, each normalized by being divided by the frequency range over which it was calculated. That is, all areaM and areaH means were divided by 4 kHz (3–7 kHz, and 7–11 kHz respectively) except for Buckeye, whose Nyquist frequency, at only 8 kHz, meant the mean of areaH was (effectively) divided by 1 kHz. This ‘normalization’ gave results in dB close to those reported by Koenig et al. (2013), and allowed us to explore very high-frequency spectral shapes when initial sampling rates allowed. Peak frequencies within the mid- and high-frequency ranges were also compared, s-freqM and s-freqH. The higher-amplitude of these two was designated s-freqMH.

To obtain these [s] spectral measures, sound files were first downsampled as necessary to 22 kHz (e.g., Quakebox, MAONZE, and Glasgow were all sampled at > 44 kHz). Buckeye, whose sampling rate was 16 kHz, was unchanged. Then, for each [s], multitaper spectra were calculated over a 25 ms Hamming window centred in the middle of the [s] segment, with a time bandwidth of 4.

Using an R implementation kindly adapted by Patrick Reidy from Reidy (2015), methods followed Koenig’s (2013) exactly except in three respects. First, as noted above, the mid- and high-frequency areas were normalized by dividing by the frequency range contributing to them. Second, following C. Shadle’s recommendation (pers. comm. 2018), we introduced two more parameters, s-freqH and s-freqMH, to compare peak frequencies above 7 kHz and in the mid-frequency range, 3–7 kHz. Lastly, while Koenig et al. (2013) measured spectra at the beginning, middle and end of their fricatives, we took spectra only from the middle of the [s]. This was for two reasons. The first reason was phonetic: Koenig et al. (2013) observed that the greatest differences due to articulatory differences tended to be in the middle of the fricative, which they interpreted as due to maximum constrictions being achieved at this point. Narrow constrictions for an articulation like [s] increase amplitudes of frication at high frequencies by increasing air particle velocity, which affects the shape of the source spectrum, and also by increasingly decoupling the front and back cavities, which reduces low-frequency spectral energy contributed by back-cavity resonances. The second reason was pragmatic: forced alignment of segments in our database tends to produce segmentation errors at the edges, which would be undesirable in these nuanced measures. A dynamic approach to measuring [s] might differentiate prefix status in dis-mis-, but might be better addressed in a more controlled study.

5.4. Overall Predictions

Our overall prediction is that all dialects will make a phonetic distinction between prefixes and phonemically identical nonprefixes, and be subject to the same nonacoustic influences, including any effects of frequency. The dialects differ in how they realize the phonemes involved, and need not necessarily use the same pattern of acoustic parameter values to mark prefixedness. Broadly, however, we expect the internal acoustic structure of nonprefixes to reflect the pattern typical of more reduced syllables for each dialect, compared with that of prefixes. But since prefixes are also relatively weak syllables, how exactly this will manifest is uncertain because not enough is known about this type of variation amongst weak dis and mis syllables in each dialect. For two reasons, we did not make specific predictions about the effect of prefixedness on the acoustic dependent variables, either separately or once the PCA had identified dimensions of acoustic covariation in the dataset. First, there are no relevant data for all the new dimensions for [s]. More importantly, because we are dealing with multiple interacting parameters and a wide range of influences on articulation, there are usually arguments that support at least two conflicting patterns. Some of these influences are discussed when we explain observed patterns.

8.1. PC1: “[s] duration and peakiness”

8.1.1. Acoustic interpretation of PC1

PC1 captures 20% of the variance. Our shorthand for the direction of PC1 is indicated by “longer peakier [s]” and “shorter less peaky [s]”. Figure 5A shows that the acoustic variables driving PC1 are related both to duration of [s] and to its spectral shape. Stronger PC1 loadings are associated with positive values for s-dur and s:i ratio, meaning that tokens with high positive loadings on PC1 have an [s] that is long both in absolute duration, and relative to the duration of the [ɪ], compared with other triphones in the dataset.

S-ampDiff also has a positive loading on PC1, and s-var a negative loading, albeit with a large confidence interval, which very slightly overlaps the null distribution. This indicates that, a very small proportion of the time, a subsample of the full dataset does not result in a significant contribution of s-var to this PC. S-ampDiff captures the peak amplitude in the mid-frequency band (3–7 kHz) minus the minimum amplitude in the low-frequency band. The observed inverse relationship between s-ampDiff and s-var suggests that, when s-ampDiff is large, most of the spectral energy is typically concentrated in a comparatively narrow bandwidth (low variance), representing a relatively prominent spectral peak in the frequency range 3–7 kHz.

We refer to low (negative) PC1 [s], then, as being less peaky. High (positive) PC1 [s] is more peaky. Example spectra from our data are shown in Figure 6, with the same three frequency bands as those in Figure 3. The ‘peaky’ high-PC1 example in the upper panel has a notable mid-frequency peak that the low-PC1 example in the bottom panel lacks.

Figure 6

Example LTAS for high (positive) PC1 (top) and low (negative) PC1 (bottom). Neither example has extreme loadings for other PCs. The high-PC1 example (top) shows a prominent narrow-band peak in the 3–7 kHz range. The lower panel shows high-amplitude energy distributed more evenly over a wide frequency range. Top: “discharging”, Liverpool male. Bottom: “disadvantage”, New Zealand male.

The raw data confirm the tight relationship between [s] duration and s-ampDiff: longer [s] typically has a mid-frequency spectral peak. Koenig et al. (2013) noted that this spectral peak tends to be more prominent when [s] is longer. In describing the acoustic conditions that produce it, they relate it to a smaller cross-sectional constriction area, especially at [s] midpoint. They noted it creates louder and more sibilant percepts, where sibilance meant “goodness” of [s]-like quality.

Finally, we note a significant, but less stable, contribution of F2-F1. In the full dataset, when the [s] is longer, there is greater separation between F2 and F1. In the bootstrap analysis, however, this is not reliably present across multiple subsamples of the data.

In sum, high PC1 is associated with a longer peakier [s]. Low PC1 is associated with a shorter [s] that is less peaky. While triphone duration and the s:i ratio have high positive loadings on PC1, this is due to [s] duration. Neither vowel duration (i-dur) nor onset duration (o-dur) contribute significantly to PC1, but there is some association with vowel formants.

8.1.2. Relationship between PC1 and prefixedness

We fit a series of models to predict tokens’ PC1 values as a function of prefixedness, as outlined in Section 6. The selected regression model for the analysis of binary prefix status on PC1 is shown in Table S10.1.1. The coefficients are plotted in Figure 7. The phonological control variables of MisDis and Stress are highly predictive. The lowest PC1 (shortest, least peaky [s]), is associated with mis forms and forms where the base word contains a late stress. With these factors controlled, there is a significant PrefixStatus × Corpus interaction (p < .04), shown in Figure 8. Post hoc tests using estimated marginal means reveal that the difference between prefixes and nonprefixes only reaches significance within the NZ Corpus (p < .05). Post hoc significance levels were not corrected for multiple comparisons for two reasons: we expected differences between dialects; and the increased risk of Type II errors with small sample sizes seems unlikely to materially improve the reliability of our conclusions (Barnett et al., 2022).

Figure 7

Coefficients and 95% confidence intervals (CIs) from binary model of PC1. Formula: PC1 ~ Corpus * PrefixStatus + LogWordFreq + Stress + BoundaryType + MisDis +(1|Speaker). Significant predictors are Stress (p < .0001), MisDis (p < .001) and PrefixStatus * Corpus (p < .04). Numbers centered above the circles show each coefficient’s value: positive in blue, negative in red.

Figure 8

PC1: Model estimates and 95% CIs for binary PrefixStatus × Corpus. Higher values reflect longer, more peaky [s].

When prefixedness Rating was used as a predictor rather than the binary classification of PrefixStatus, it did not reach significance, either overall or when tested individually within prefixes and nonprefixes. Stress and MisDis remained strongly significant (t = –6.488 and –3.804 respectively), but no other main effects or interactions were significant (Tables S10.2.1–S10.2.3).

In sum, the overall pattern suggests that prefix status behaves in a weakly binary fashion for the parameters accounted for by PC1. Although Glasgow’s [s] tends to be shorter in prefixes than in nonprefixes, and the other corpora trend the other way, the only statistically significant difference between prefixes and nonprefixes is for longer [s] in prefixes within the New Zealand corpus.

8.2. PC2: “[s] spectral balance”

8.2.1. Acoustic interpretation of PC2

Figure 5B shows the loadings for PC2, which accounts for 17% of the variance. This is straightforwardly related to variation in the spectral balance of the [s]: whether there is relatively more energy in the high or mid frequencies. High PC2 relates to robustly negative values of s-levelDiff, and positive values of s-freqM and s-freqMH. There is a weaker but still significant positive effect of s-var in the full dataset, with confidence intervals that only slightly overlap the null distribution. This pattern reflects an [s] spectrum whose balance is tilted towards higher amplitudes in the high frequencies, with relatively broad spectral prominences. Conversely, low values on PC2 reflect an [s] with high-amplitude energy in the mid-frequency range, and relatively less above 7 kHz.

Figure 9 gives example spectra from a high PC2 (top) and a low PC2 (bottom) [s]. The high-PC2 example has more energy in the high frequencies, whereas the low-PC2 example has more energy in the mid frequencies. The relatively broad spectral prominence, especially in the high-PC2 case, may simply reflect the fact that a formant’s bandwidth typically widens significantly as its centre frequency increases. It may also reflect that s-var varies with PC2 and is negatively correlated with the strongly significant s-levelDiff. A high value of s-var and low s-levelDiff implies a relatively broad spectral prominence in the high frequencies.

Figure 9

Example LTAS [s] spectra for high PC2 (top) and low PC2 (bottom). Neither example has extreme values for other PCs. The high-PC2 spectrum shows more energy in the high frequencies. The low-PC2 spectrum shows more mid-frequency energy. Top: disappointing, New Zealand female. Bottom: disappear, vernacular Glaswegian male.

8.2.2. Relationship between PC2 and prefixedness

The selected regression model of binary prefixedness is shown in Table S11.1.1 and the coefficients are plotted in Figure 10. BoundaryType affects [s] spectral balance, with the spectrum skewed towards low frequencies (low PC2) for CodaPossible forms (p < .01). Figure S11.1.1 shows that PrefixStatus × Stress (p < .05) is due solely to unprefixed forms with late stress having lower PC2 than the other three conditions. That is, the nonprefix [s] spectrum skews away from the high frequencies, but only when the second syllable in the word is weak—when a word has more than two syllables and its main lexical stress falls on syllable 3 or later (unprefixed disappointing, distillation). When the second syllable carries the main lexical stress, then prefixed and unprefixed forms have essentially identical, slightly positive, loadings on PC2 (prefixed discomfort, unprefixed distinguished). This is confirmed by post hoc tests with estimated marginal means: with late Stress, prefixes have higher values of PC2 than nonprefixes (more spectral energy in the high frequencies [p < .026]) whereas prefixes and nonprefixes with early Stress have similarly high values of PC2 (p = .83); within nonprefixes, late Stress lowers PC2 values (p < .007) whereas prefixes are not affected by Stress (p > .76).

Figure 10

Coefficients and 95% CIs from binary model of PC2. Formula: PC2 ~ Corpus + LogWordFreq + PrefixStatus * Stress + BoundaryType + MisDis +(1|Speaker). Significant predictors are BoundaryType (p < .01) and PrefixStatus × Stress (p < .05). Numbers centered above the circles show each coefficient’s value: positive in blue, negative in red.

When prefixedness is viewed as a continuum, Rating does not reach significance either as a main effect or in interaction with Stress. This is true when tested over the whole dataset, and also when tested separately within prefixes and nonprefixes (Tables S11.2.1–S11.2.3).

8.3. PC3: “vowel duration”

8.3.2. Relationship between PC3 and prefixedness

The binary regression model of PC3 is shown in Table S12.1.1. The coefficients are plotted in Figure 11. Consistent with a wide range of previous research, PC3 is significantly predicted by word Frequency, with higher frequency words having shorter vowels. There is a more strongly significant effect of PrefixStatus × Corpus, shown in Figure 12. Post hoc testing with estimated marginal means reveals that prefixed forms have higher PC3 (longer vowels) than unprefixed forms for US (p < .0005), NZ (p < .05) and Liverpool (p < .01), but not Glasgow (p = .65). It also reveals that the difference between the corpora is carried by differences within the prefixed set. While there are no differences in PC3 across corpora for nonprefixes, within prefixes the US has a significantly higher PC3 than both NZ (p < .005) and Glasgow (p < .01). Figure 13 shows the PrefixStatus × BoundaryType interaction: prefixed forms have higher PC3 (longer vowels) for both BoundaryTypes. The difference is more robust for CodaPossible cases, when the fourth phoneme is /p t k/. (Posthoc estimated marginal means tests: CodaPossible p < .001, CodaImpossible p < .01).

Figure 11

Coefficients and 95% CIs from binary model of PC3. Formula: PC3 ~ Corpus * PrefixStatus + LogWordFreq + Stress + PrefixStatus * BoundaryType + MisDis +(1|Speaker). Significant predictors are LogWordFreq (p < .01), PrefixStatus * BoundaryType (p < .05) and PrefixStatus * Corpus (p < .001). Numbers centered above the circles show each coefficient’s value: positive in blue, negative in red.

Figure 12

PC3: Model estimates and 95% CIs for binary PrefixStatus × Corpus interaction. Color: prefix status. Higher values reflect a longer vowel.

Figure 13

PC3: Model estimates and 95% CIs for binary PrefixStatus × BoundaryType. Left: CodaImpossible. Right: CodaPossible. Color: prefix status. Higher values reflect a longer vowel.

The ratings model also contains a significant Rating × Corpus interaction (Table S12.2.1, Figure S12.2.1). The pattern is much the same as in the binary model. However when tested separately within the prefixes and nonprefixes, Rating does not reach significance within either subset (Tables S12.2.2–S12.2.3). There is thus no evidence that there is a separate gradient effect that explains variation within either the prefixed or the unprefixed words.

8.4. PC4: “F2-F1 spacing in the vowel”

8.4.2. Observed relationship between PC4 and prefixedness

The binary regression model of PC4 is shown in Table S13.1.1, and the coefficients are plotted in Figure 14.

Figure 14

Coefficients and 95% CIs from binary model of PC4. Formula: PC4 ~ Corpus + PrefixStatus + LogWordFreq + Stress + PrefixStatus * BoundaryType + MisDis +(1|Speaker). Significant predictors are BoundaryType (p < .01), PrefixStatus * BoundaryType (p < .01) and MisDis (p < .001). Numbers centered above the circles show each coefficient’s value: positive in blue, negative in red.

Words beginning with mis- have significantly higher PC4 values. This is presumably due to different coarticulatory constraints imposed by the /m/ or /d/ onset, which could affect formant spacing in the vowel. However, the prefix effect is not an artefact of the strong mis-dis- effect: There are no substantive differences in parameter values between models with mis- included versus excluded. When mis- is excluded, the three estimated parameters of greatest interest are virtually identical to those in Table S13.1.1. (PrefixStatus: 0.025, SE = 0.137, t = 0.180. BoundaryType: 0.384, SE = 0.139, t = –2.765. PrefixStatus × BoundaryType: 0.774, SE = 0.265, t = 2.927.)

PrefixStatus interacts significantly with BoundaryType (Figure 15). Post hoc estimated marginal means tests confirm that PC4 differentiates PrefixStatus in CodaPossible words (p < .002), but not in CodaImpossible words (p = .89) which form a majority of tokens within the prefixed category (345 vs. 43, Table S3.3). In other words, the formant frequencies suggest that, for triphones followed by /p t k/, the vowel is more open and/or centralized in prefixes, and more close or fronted in nonprefixes.

Figure 15

PC4: Model estimates and 95% CIs for binary PrefixStatus × BoundaryType interaction. Left: CodaImpossible. Right: CodaPossible. Color: prefix status. Higher values reflect smaller F2-F1 frequency spacing.

To check whether it is specifically a following phonologically voiceless stop, rather than a following consonant of any sort, that creates the prefix effect on vowel formant spacing, we tested models that distinguished all vowels from all consonants following the triphone. These showed no difference between boundary type, confirming the current analysis of PC4: prefixes and nonprefixes were distinguished only when phoneme 4 is a stop consonant.

The Ratings model shows no overall significant gradient effect of prefixedness on PC4 (Table S13.2.1). However, when prefixes and nonprefixes are tested separately, Rating is significant within the prefixed forms, but not within unprefixed forms (Tables S13.2.2.–S13.2.3). Figure 16 shows that prefixed forms that are rated as more prefixed have higher PC4 (Rating estimate = 0.0267, SE = 0.008, t = 3.3). There is no significant interaction with BoundaryType. However, most prefixed forms are CodaImpossible (345) as opposed to CodaPossible (43).

Figure 16

PC4: Significant effect of Rating within prefixes only.

8.4.3. PC4: Discussion

PC4 captures a pattern in which prefixes have less separation between F2 and F1. How do we reconcile this with PC1, in which there is a significant loading of F2-F1, and which seems to capture a pattern in which prefixes have more separation between F2 and F1? This is the only variable which is significantly loaded onto different PCs in opposite directions, so requires a little unpacking.

Importantly, while PC4 captures a pattern that exists within F2-F1, this does not mean that it is identical to raw F2-F1. (Pearson’s correlation between PC4 and the z-scored F2-F1 is –.81, p < .0001.) Rather, PC4 captures that aspect of F2-F1 separation that is not captured by other components. PC1 reflects that longer /s/ (and longer segments in general) are associated with more F2-F1 separation. This is as we might expect through phonetic peripheralisation. And PC1 is associated with prefixes, at least for NZ. PC4, on the other hand, captures the fact that there is substantial extra variation in formant separation that is not correlated with or explained by differences in [s] duration. To solidify the intuition that prefixes can have both more formant separation than nonprefixes, and less formant separation than expected given the duration patterns, Figure 17 shows the z-scored formant values, for the three tertiles of [s]-duration values. (Note the x-axis shows raw F2-F1 [z-scored Hz], not PC4, which is negatively associated with F2-F1).

Figure 17

Z-scored formant separation (x-axis), for different tertiles of z-scored [s] duration: long (top), mid, and short (bottom) durations. These are z-scored raw values (Hz × ms), not principal components.

A number of things are apparent. First, distributions in the top panel sit to the right of the others. This reflects the effect of PC1: tokens with a long [s] tend to have more formant separation. Second, prefixes are much more likely to be associated with the long [s] distribution (41% of prefixes) than the short distribution (17% of prefixes). This means that, across all data, prefixes have more formant separation. Third, within long-duration tokens, prefixes have more formant separation than nonprefixes. They are also likely to be longer—though Figure 17 does not show that because duration is treated as categorical.

However, in PC4, the association between duration and formant separation is controlled (by PC1). PC4 therefore represents unexpected formant separation that is not associated with duration increases. Our regression model tells us that this is associated with nonprefixes. Indeed, Figure 17 shows that tokens with a very short [s] (bottom panel) clearly differentiate prefixes from nonprefixes: most prefixes have less formant separation than most nonprefixes, though a few prefixes have greater F2-F1 separation. For mid-duration tokens, prefixes have a broader range of formant separation than nonprefixes—some with more separated formants, others with less.

This seems to reveal a second pattern of production for prefixed words. Most often, they are associated with a stronger ‘beat’. But they also show less formant separation than we would expect given the degree of ‘beatness’, and this is reflected in the association of triphone duration with PC4 (Figure 5), longer triphones being associated with reduced formant separation.

Our binary model tells us that the pattern of reduced formant separation in prefixes is prevalent in CodaPossible prefixed words (where phoneme 4 is a stop e.g., misplacement, mischaracterization, discontented). Moreover, words rated ‘more prefixed’ are more likely to be CodaPossible. The mis-dis- in these words may be most easily recognized as a unit—semantically, and inasmuch as the vowel may differ maximally between prefix and nonprefix due to phoneme 4 being syllabified with the triphone in nonprefixed words but not in prefixed ones.

What, then, would that syllabification difference imply for realisation of the vowel? More canonical vowels are expected in prefixed words. What counts as more canonical is complex in this dataset of multiple dialects. Our prefixed CodaPossible tokens come mainly from NZ (65%) and Glasgow (14%) speakers, whose canonical KIT vowel we know (from the literature and our own impressionistic listening) approximates schwa. Wells (1982, vols 1 & 3) states that NZ [ɪ] is lowered and centralised and does not functionally contrast with schwa, while Glasgow /ɪ/ can be as open as [ʌ] or [ɛ]. Since only direction of change is important to the current argument, we term all these sounds schwa. When canonical KIT vowels are realised as close to [ə], then, effectively destressing the dis-mis- syllables may shift their phonetic quality closer to [ɪ]—i.e., less-stressed triphones will have greater F2-F1 formant spacing in NZ and Glasgow (Figure S13.3.1). The five CodaPossible prefixed forms in our US data are consistent with a trend in the opposite direction, as would be expected since Ohio KIT approximates IPA [ɪ] and its high vowels are typically centralised when unstressed.

In short, although small numbers mean that this conclusion can only be tentative, the observed patterns in PC4 are consistent with an interpretation in which, for all dialects, the change in vowel quality from prefix to nonprefix (within CodaPossible) represents a relative “relaxation”, or less extreme articulation: less centralized than expected for NZ and Glasgow, and trending towards more centralized than expected for Ohio.

What the PC4 pattern highlights is that, while prefixes are phonetically distinct from nonprefixes, it does not follow that there is only one way to be distinct. There are clearly multiple forces operating together to create the observed patterns. Supplementary Materials Sections 13.3–4 give details.

8.5. PC5: “[s] low-frequency cutoff”

8.5.2. Observed relationship between PC5 and prefixedness

The binary regression model of PC5 is shown in Table S14.1.1, and the coefficient estimates are plotted in Figure 18. PrefixStatus is strongly significant, as are word Frequency and Stress. There are no significant interactions. The cutoff frequency is higher when the [s] is in a prefix compared with a nonprefix (p < .02), in higher frequency words (p < .01), and when Stress is late i.e., an unstressed syllable follows the [s] (p < .05).

Figure 18

Coefficients and 95% CIs from binary model of PC5. Formula: PC5 ~ Corpus + PrefixStatus + LogWordFreq + Stress + BoundaryType + MisDis +(1|Speaker). Significant predictors are Prefixedness (p < .02), LogWordFreq (p < .01) and Stress (p < .05). Numbers centered above the circles show each coefficient’s value: positive in blue, negative in red.

The Ratings model has a significant effect of prefixedness (p < .01, Table S14.2.1, Figure S6). Considered separately, there is a significant gradient effect within prefixes (p < .05, Table S14.2.2), but not within nonprefixes (p = .86, Table S14.2.3). The stronger the ‘felt prefixedness’ of a word, the higher its low-frequency cutoff in [s] (Figure 19).

Figure 19

PC5: Gradient effect of prefixedness rating, within prefixes only.

8.6. Summary of binary prefix effects across corpora

All five PCs showed a significant binary effect of PrefixStatus, either in isolation or in interaction with another factor. For PC1 and PC3, PrefixStatus interacted with Corpus: for these PCs, post hoc tests reported above show that prefixedness reaches significance within some corpora, but not others. Of relevance is whether each corpus marks the prefix phonetically. For example, PC2 and PC5 do not have a significant overall effect of Corpus, suggesting that the same general trend is likely to be present in all corpora. But all corpora containing the same trend is somewhat different from the effect being robust enough to reach significance within each corpus: the overall effect could be carried by some corpora while others are not significant. We tested this by adding a Prefix × Corpus interaction into those models for which that interaction is not significant (PC2,4,5), and conducting post hoc tests to assess for a prefixedness effect within each corpus. Table 4 summarises the results across PCs and Corpora, both in the overall binary model, and in the post hoc within-corpus testing (details in Supplementary Materials Section 15).

We see different patterns of significance across the corpora. It makes sense that the two larger corpora have more significant effects. The smaller number of effects for Glasgow and Liverpool should not be taken as evidence that they mark the distinction more weakly, especially since only 2/5 of the PCs have a significant Corpus × PrefixStatus interaction. Table 4 does suggest, however, that the patterns may be somewhat different across the corpora, and that each corpus shows a significant effect for at least one PC.

9.1. Acoustic parameters: PCs 1–5

The first five PCs capture all the main dimensions of acoustic covariation in the data: PC1, PC2 and PC5 for [s], and PC3 and PC4 for the vowel. PCs 1, 2 and 5 all involve differences in [s] spectral shape, but in different ways. PC1 mainly reflects longer [s] for prefixes, significantly so only in the NZ corpus. We interpret PC1’s spectral difference—degree of mid-frequency peakedness—as a secondary effect of the much stronger influence of [s] duration. PC2 reflects overall balance in relative amplitude between mid versus high frequencies. It is affected by prefix status only insofar as nonprefixes that are not adjacent to the main lexical stress have lower-frequency spectral energy. The spectrum is also skewed towards mid frequencies when phoneme 4 is a stop, regardless of prefix status. PC2 is discussed further with PC5 below, and Supplementary Materials Section 11.3 discusses articulatory causes of PC2’s overall pattern. PC5 reflects the frequency at which the spectral envelope of [s] gains its highest amplitude within the range 3–7 kHz—a range which covers most vocal tract lengths for this property. So it too captures an aspect of spectral shape, but, unlike PC1, not one that reflects degree of peakedness.

PC3 captures vowel duration. In the binary model, prefixes have longer vowels than nonprefixes in all corpora except Glasgow. The difference is especially large for the US. Prefixes also have longer vowels for both boundary types, with a more robust effect for CodaPossible forms than CodaImpossible forms. That is, the difference is greater for forms whose fourth phoneme is /p t k/. PC3 Ratings show the same PrefixStatus × Corpus interaction as the binary model, but there is no evidence of a separate gradient effect that explains variation within either the prefixed or the unprefixed words: tested separately within prefixes and nonprefixes, neither ratings model reaches significance.

PC4 captures spectral variation in vowel quality, as does, to a lesser extent, PC1. With the association between duration and F2-F1 separation controlled (in PC1), prefixedness and formant spacing are negatively associated. In the CodaPossible set, but not the CodaImpossible set, prefixes have smaller F2-F1 spacing than nonprefixes. The directions of change in PC4 differ because the dialects’ KIT vowels differ in quality, and these nonprefixes carry a weaker beat than prefixes. Though both prefixed and unprefixed mis-dis- are weak syllables, nonprefixes are more reduced, and so have a less canonical KIT vowel than prefixes. For NZ and Glasgow this means a shift towards [ɪ] with increasing syllable reduction. For Ohio speakers, it means a shift towards [ə]. Though only, as noted, within CodaPossible words in our data. We suggest PC3 and PC4 combine to give CodaPossible prefixes a heavier beat than nonprefixes.

We also see a gradient effect within prefixes, with higher levels of felt ‘prefixedness’ having high PC4 (and thus smaller unexplained F2-F1 differences). More data are required to fully disentangle these effects. What can be said, however, is that the present data support interpretations in the literature (cited in the Introduction) that prefixed cases like discolour differ in their syllabification from unprefixed cases like discover, in ways that simple phonotactic rules cannot explain: an analysis reflecting morphological structure is required.

Turning now to the broad picture, prefixes are marked by a tendency for NZ, US and Liverpool [s] to be slightly longer and peakier (only significantly so for NZ), and a relatively long vowel, whose F2-F1 spacing represents a more canonical KIT vowel for the specific dialect, but only when phoneme 4 is a voiceless stop. Vowel duration is the most robust differentiator of prefix status.

PCs 1, 3 and 4 show an effect of binary PrefixStatus, although always in interaction with one or more other predictor variable. PrefixStatus interacts with Corpus for PC1 and PC3, and with BoundaryType for PC3 and PC4. Within these generalizations, there is dialectal variation as already described.

PC5 stands out as the only component showing an effect of binary PrefixStatus independent of other predictor variables. PC5 also has one of the two significant gradient effects in our Rating data, the other being in PC4: the stronger the ‘felt prefixedness’ of a word, the more canonical the vowel for its dialect (PC4) and the higher its low-frequency cutoff in [s] (PC5). These two properties could both reflect speakers’ intuitive sense that greater syllable stress accompanies stronger prefixedness.

PC5’s pattern is compatible with that for PC2. Though independent of each other, both reflect effects on [s]’s spectral shape. Moreover, Stress is implicated in both. In PC5, Late Stress (in which syllable 2 is unstressed [disillúsion]) raises s-freqM, as does prefixedness. In PC2, Late Stress skews spectral amplitudes of nonprefixes towards lower frequencies. Thus these nuanced spectral measures of [s] potentially capture the degree of stress the segment carries, broadly defined as encompassing prefixedness, lexical stress, and presumably utterance prosody, along with related influences like word frequency. Together PC5 and PC2 may contribute to subtle gradations observable in English utterance and lexical stress: prefixes vary more than nonprefixes in comparable environments, including nuclear and postnuclear positions (Smith et al., 2012), with consecutive unstressed syllables tending to be somewhat differentiated from one another. More broadly, these tentative observations suggest how multiple parameters combine to determine subtle gradations in degree of syllable stress, from fully stressed to highly reduced. Confirmation would require work with very large corpora, or highly controlled parameters in standard experiments.

These results are largely compatible with the findings of Smith et al. (2012), but go beyond them not just by offering more nuanced observations about [s], but also allowing us to examine differences across dialects and BoundaryType, to seek gradient as well as binary distinctions, and to identify any independent influences of Stress and word Frequency, as well as patterns of covariation amongst the acoustic parameters.

9.2. General Discussion

We set out to answer four interlinked research questions: (1) To what extent is the prefixed-unprefixed distinction described by Smith et al. (2012) marked in spontaneous speech and different phonological contexts? (2) Is it observable across different dialects of English? (3) Can it be definitively distinguished from lexical frequency? (4) Is it best described as a binary phonetic distinction, or is there evidence of morphological gradience which is reflected in gradient phonetic realization?

For question (1), the distinction is clearly present in spontaneous speech. While the results are broadly consistent with Smith’s (2012) findings for SSBE, there are also differences. These likely stem from combinations of three influences: the methodology, allowing us to decouple different aspects of covariation and consider them independently from each other, albeit with much less control over segmental, prosodic and other contextual variation; the words analysed, which encompass a much broader range of following phonological environments; and perhaps the absence of SSBE amongst the varieties analysed here. (SSBE is phonetically very different from all four dialects analysed here, though in the relevant acoustical aspects it behaves rather like our US corpus.)

Prefixes contain longer vowels (except for Glasgow) and longer [s] (though only significantly so for NZ). The vowel effect is more robust than the [s] effect, the net result being consistent with Smith et al.’s (2012) findings that prefixed forms have a smaller s:i ratio, albeit with the dialect-specific differences noted in Section 9.1. PC4 offers evidence that in spontaneous speech, prefixes can differ from nonprefixes in multiple, sometimes competing ways.

For question (2), all dialects mark prefix status acoustically in spontaneous speech. Duration (of vowel and [s] independently), formant frequencies in the vowel, and attributes of the spectral envelope of [s] can all contribute to the difference. Not all dialects use all these properties (see Table 4), but all use some, though sometimes only when other, more influential factors are statistically controlled, especially BoundaryType, Stress, and MisDis.

We do not wish to read too much into the specific differences observed in this dataset. The recordings are not completely comparable in terms of speech style, quantity of observations, or the specific words produced. To be sure of the exact nature of the dialectal differences, more controlled experiments are needed to replicate the observed differences, or a repeated analysis on other corpora. What does seem clear, though, is that not all the dialects behave in exactly the same way. This was expected given the number of acoustic properties involved and other differences between the four dialects. Nonetheless, for all of them, the vowel is longer in prefixes than nonprefixes, except for Glasgow, for which vowel duration is constrained by other morphophonological factors (Rathcke & Stuart-Smith, 2016).

The answer to question (3) is that the prefixedness effect is not an artefact of word frequency. Word frequency clearly affects the production of such affixes and nonaffixes: it has been reported in the literature (e.g., Pluymaekers et al., 2005), and significantly influenced vowel duration (PC3) in our own data. However, it did not interact with prefix status even for vowel duration. Importantly, for all PCs we controlled for frequency effects, and still found an effect of prefixedness over and above that of lexical frequency. This conclusion confirms Smith et al.’s (2012) more tentative observations.

For question (4), work on the cognitive representation of morphological structure has found evidence of gradient representation, with words containing the same affix seemingly affixed to different degrees. Factors that contribute to perceived degree of affixedness include, among other factors, semantic transparency (Gonnerman & Andersen, 2000; Wurm, 1997), the (relative) lexical frequency of the word form and the base (Hay, 2003), and the phonological transparency of the boundary (Gonnerman & Andersen, 2000). See Hay and Baayen (2005), Raffelsiefen (1999) and Stein (2023) for reviews, and Smith et al. (2012) Section 4.2 for brief further comments, including on historical change. Rather than quantify such factors separately, we attempted to get ratings reflecting an aggregate measure of subjective frequency that captured the combined effect of such factors.

As Figure 1 shows, participants responded bimodally to this rating task. Even though they were given a continuous scale and encouraged in the instructions to think in terms of gradience, they sorted words into two clear categories of ‘prefixed’ and ‘nonprefixed’. However, there was considerable variation within each category and they are not completely nonoverlapping (see misnomer, Table 12 [Appendix C] and its discussion in Section 4.1). So some words seem subjectively highly prefixed, while other prefixed words feel much more ambiguously prefixed. Indeed, we classed misnomer as unprefixed with some misgivings, which were confirmed by the collected ratings. The correlation plots in Figure 2 demonstrate substantial agreement across dialect areas in ranking the words, including the ambiguous words with intermediate ratings. This is not just an effect that emerges in the aggregate. Individuals used the scale gradiently. Just 11% of the individual ratings given were 0, 32% were 100, and the rest fell somewhere between the ends of the scale. No participant exclusively used the ends of the scale. Speakers, then, do have a sense of gradience within these categories. This perceived degree of membership of the category was associated with the same factors that significantly distinguished the two categories: lexical frequency, stress pattern, word length, and phonological context.

The statistical analysis of our acoustic-phonetic data likewise supports a largely binary prefixed-unprefixed distinction in phonetic realization. For all five PCs, a binary analysis shows some effect, and while subjective ratings of gradient prefixedness can sometimes also predict the acoustics well, for PCs 1–3 we found no evidence that the rating continuum adds significant explanatory power beyond that of the binary classification.

However, PC4 and PC5 are significantly higher in prefixed forms rated as more prefixed. This provides evidence of phonetic gradience in production within the prefixed category. Words can seem prefixed to different degrees and this affects their production. In sum, the weight of the evidence is that the prefix-nonprefix distinction is heavily bimodal, and reflected largely categorically in the acoustics, but there are also gradient phonetic effects in vowel quality and low-frequency cutoff of [s] within the prefixed category.

While ‘rating’ does not significantly predict acoustic realizations better than prefix status for PCs 1–3, it is important to consider that our models of the acoustics control for many of the factors that statistically influenced these subjective ratings (stress, frequency, following environment). In our preliminary exploratory modelling, which did not contain all these controls, we did in fact find multiple cases where rating could predict realization within either prefixed or unprefixed forms. In many ways, the combined linguistic factors that we are controlling for are in large part the gradience (although other factors are involved too, most notably semantic transparency). The consistency of raters’ behaviour does suggest that gradience within the categories has some cognitive reality. It seems that the sense of affixedness and the linguistic properties are interrelated and likely work together in affecting the acoustic pattern.

Our data illustrate a wide set of available acoustic properties that can be harnessed to indicate prefix status, and how they work together, with different dialects using them in different combinations. If, as we and others have suggested, a broad overarching principle drives these results, then this raises interesting questions regarding how the dialect-specific differences we have found to reflect the prefix distinction can be generalized to how each dialect marks rhythmic prominence, and its opposite, syllable reduction. For example, speakers of Ohio English might vary vowel duration more substantially with the prominence of a syllable than speakers of Glaswegian English. For varieties that vary the vowel duration, if the overarching principle is governed by rules of syllable prominence and reduction, then we might expect the prefix status of the initial triphone to be less clear under conditions that favour extremely short or extremely long vowels: sometimes prefixes need to be spoken very fast, causing highly-reduced vowels, while at other times the talker may need to emphasize the first syllable of an unprefixed word, resulting in a heavier syllable and longer vowel. The implication is that prefix status may be most clearly conveyed when vowel durations have intermediate (non-extreme) values. Further work into the wider dialectal differences, then, may shed considerable light into the processes driving these acoustic consequences of prefixedness. If there are not comparable differences in the wider phonological systems of these dialects, then that would locate the observed differences very locally in the dis-mis- prefix, and suggest that the differences are instead located in phonetically-rich dialect-specific representations of the prefixes themselves.

What we have observed here goes beyond standard descriptions of differences between dialects. The bread and butter of the field of language variation and change is differences across dialects and speakers in terms of language usage. This might be grammatical variation (e.g., Tagliamonte & Baayen, 2012), phonological variation (e.g., Labov, 2001), phonetic variation (e.g., Kiesling, 1998), or—very occasionally—morphological variation (e.g., Nevalainen et al., 2011). Looking at the interface between these—as in, for example, how morphological structure affects interacting covariant phonetic cues—can potentially lead us away from a focus on superficial differences in isolated variables, toward an understanding of differences in the systemic organization of the wider communicative system.

Bybee (1985, pp. 89–90) cites Pagliuca’s (1976) unpublished study of the prefix pre, which the Shorter Oxford English Dictionary listed as having four different vowel qualities. She discusses the complex acoustic consequences for pronunciation dependent on frequency, transparency of meaning between putative prefix and its base, and the variety of meanings a word (or its base) can have: “there is a strong relationship between the vowel quality and the frequency of the word, and the vowel quality and the semantic predictability of the word.” (p. 89). And “the higher the frequency of a derived word, the more likely it is to occur in a variety of contexts, including some in which its related base word, and the semantic notions expressed by it, do not occur.” (p. 90). Together with our own results and those of others (e.g., Engemann & Plag, 2021), this illustrates that the pronunciation of prefixed words is likely to be subject to the multiple influences that affect all word production, ranging across phonological and prosodic word structure, phonotactics and neighborhood density, sequential and paradigmatic probability, and frequency of occurrence (of potentially the prefix, the stem and the whole word), as well as situation-specific influences stemming from the pragmatic and functional load on the prefix in the particular discourse.

Individual variation amongst talkers provides a further dimension of complexity, even within a single variety of the same language. Using ultrasound and acoustic measures, Strycharczuk and Scobbie (2016) found evidence of gradient differences between 20 SSBE speakers in realisation of vowels and /l/ in word sets like bimorphemic fooling versus monomorphemic hula (and also fool). While all speakers distinguished the bi- and monomorphemic forms, there was a fairly uniform distribution across speakers in the degree to which they made the distinction: some made a largely categorical difference, others made very little distinction, and most talkers were intermediate, resulting in gradiency overall, but not necessarily for everyone.

It is not surprising, therefore, that one can find gradient and sometimes contradictory effects on particular prefixes. But it does seem that the prefix status of our critical syllables exerts its own effect, which speakers of different dialects of English respect in one way or another, usually closely related to the way somewhat stronger rhythmic (yet still weak) beats are realized. This sensitivity to prefixedness is problematic for any model predicting that implementation should be blind to the morphology, and adds to a growing body of research which provides evidence of acoustic consequences of morphological structure (see Plag & Ben Hedia, 2018; Schmitz, 2022; Tang & Bennett, 2018 and references therein). A number of papers suggest that there is an impetus to avoid reduction of a morpheme to facilitate communication of affix meaning (Arnon & Cohen Priva, 2013; Hanique & Ernestus, 2012; Pluymaekers et al., 2010; Rose, 2017), and this interpretation accounts well for prefixes. Prefixes convey an important meaning. Many, including dis and mis, transform the meaning of the lexical stem to something close to its opposite. We know that listeners use the acoustic properties signalling prefixedness in real time to distinguish between lexical competitors (Clayards et al., 2021; Blazej & Cohen-Goldberg, 2015; see also Hawkins, 2011). Ensuring that the additional meaning is intelligible, and/or that listeners’ attention is drawn to the prefix, may explain why prefixes are typically more rhythmically/metrically prominent than nonprefixes.

How might these factors be conceptualised? Clayards et al. (2021) concluded that while listeners used the phonetic detail of both types of initial triphone to predict word identity, the rhythmic emphasis that comes with a true prefix probably allows faster and more accurate prediction of meaning than the absence of such a focus found in nonprefixes. They relate this observation to the centrality of metrical rhythm in a general, probabilistic and context-dependent theory of speech perception, drawing evidence from neuroscientific work that shows that the listening brain constructs metrical beats from an incoming rhythmic signal (which can be multimodal), and focusses attention onto them, thereby facilitating prediction and allowing meaning to be efficiently accessed. When such a metrical beat is lacking, attentional processes shift to more continuous monitoring. These systems can operate in tandem, with the rhythmicity of the stimulus determining their relative balance (Schroeder & Lakatos, 2008). Presumably, talkers normally adapt their speech to accommodate both types of listening process, such that the intelligibility of the utterance as a whole falls within acceptable limits for the dialect and communicative situation.

A theoretical prosodic account and communicative accounts such as those discussed by Clayards et al. (2021) need not be mutually exclusive. They have in common multiple interacting influences on behaviour. Advances in analytical methods and a wider variety of databases are beginning to indicate how these multiple influences may work within and across languages. See for example Strycharczuk (2019) for a number of European languages, Stein and Plag (2022) and Stein (2023) for English, and Tang and Bennett (2018) for Kaqchikel Mayan. A new type of statistical modelling Linear Discriminative Learning, LDL (Baayen, Chuang, Shafaei-Bajestan and Blevins, 2019; Heitmeier et al., 2021; Stein & Plag, 2021), uses the highly abstract concept of ‘lexome’ to map form to meaning in networks that preserve the integrity of words, whether morphologically complex or not (content lexomes), while recognising structural patterns that signal, for example, morphological relationships (derivational lexomes), and inflectional functions such as number and tense (inflectional lexomes). Thus form and meaning work together in an integrated system. What matters in the case of prefix status is that there is a structure, or form, some of whose details depend on whether the last segment abuts a prosodic boundary with the rest of the word, and this structure forms an identifiable pattern. Such patterns interact with other structural and statistical influences that may enhance or obscure the prefix-related pattern. In an implicit learning model like LDL, these patterns can be learned and will generalise to the rest of the language to form a largely (though not completely) coherent morphophonological system with predictable semantic consequences (e.g., Kuperman et al. (2007) and references cited above). They can be reflected elsewhere too, such as in the content-word versus function-word distinction (Bell et al., 2009).

One issue arising from our own work concerns whether the patterns of syllable reduction observed in the present study generalize to other syllables with similar phonetic structure but without potential prefix status, for example wisteria, hysterical, hysteresis. We hypothesize that they should, and will be subject to the same influences such as lexical stress, wider prosodic context and word frequency/predictability. Other challenges concern perceptual processing. Which combinations of acoustic properties convey the heavier beat said to characterize prefixes? For dialects like the US/Ohio, is the s:i ratio critical regardless of overall rhyme duration, as the SSBE data of Smith et al. (2012) suggest it might be, or does a relatively long triphone always convey prefixedness regardless of its internal acoustic structure? For CodaPossible dis- and mis- words, can a long VOT override acoustic-perceptual cues that would otherwise be heard as an unprefixed word?