Glossa Psycholinguistics

1. Introduction

The English writing system is difficult to learn and use, in part because a letter may have one pronunciation in some words and other pronunciations in other words. This difficulty, sometimes referred to as orthographic depth, is especially prominent for vowels. For example, the letter ‹a› corresponds to different phonemes in chapter and paper. How do readers cope with such variability? Studies suggest that the context in which a letter appears often provides useful information about its pronunciation and that readers benefit from this information (e.g., Kessler & Treiman, 2001; Siegelman, Kearns, et al., 2020). In the research reported here, we study one type of context effect in spelling-to-sound translation: how the number and type of consonants after a vowel influence its pronunciation. We focus on vowels that are followed by consonant sequences beginning with ‹s›, because, as we will explain, this is a good test case for examining context effects in the English writing system and how they influence readers’ behavior.

Many studies of the nature and use of spelling–sound relations in English (e.g., Kessler & Treiman, 2001; Siegelman, Kearns, et al., 2020) have been limited to monosyllables. Our study extends previous work by examining items with more than one syllable. Although some researchers have studied such items (e.g., Jared & Seidenberg, 1990), more work is needed, given the frequency of longer words in English, the importance of these words for understanding complex texts, and the fact that only some types of linguistic structures have been examined.

Learners of English are taught that each vowel letter has two main pronunciations: what educators call a short one and a long one. The five main vowels have their short sound in cat, bed, big, body, and run, and their long sound in cake, betel, bike, bony, and rude and cute. (Note that these designations are terms of art in phonics; the vowels in the two categories differ in quality, and the former are not always longer in acoustic duration than the latter.) In phonics instruction, students are often taught to consider the number of consonant letters after a vowel in order to choose between the long and the short pronunciations (Kearns, 2020). Specifically, students are directed to first try the short pronunciation of a vowel letter that is followed by two consonant letters, as in chapter and pepper. If the vowel is followed by a single consonant letter and then another vowel letter, as in paper, their first choice should be the long pronunciation. Thus, the view that underlies typical phonics instruction is that the number of consonant letters after the vowel influences its pronunciation.

Kearns’s (2020) analyses of the spelling-to-sound correspondences in English disyllabic words confirm the idea from phonics instruction that the number of consonant letters after a vowel is associated with its pronunciation. Specifically, according to his analyses, the first vowel of a VCCV (C = consonant, V = vowel) sequence is more likely to be pronounced as short than is the first vowel of a VCV sequence. However, the results of Treiman et al. (2023) suggest that the ratio of short to long vowels varies not only with the number of following consonant letters, but also with the phonological properties of the consonants for which the letters stand. According to the vocabulary analysis of Treiman and colleagues, long vowels are reasonably common in disyllabic words like apron, in which the medial CC sequence corresponds to an obstruent phoneme (e.g., a stop or a fricative) followed by a liquid phoneme. Long vowels are virtually nonexistent when a medial CC sequence corresponds to an obstruent followed by another obstruent.

Treiman et al. (2023) suggested that the higher proportion of long vowels in words such as apron than in words such as captain is related to the fact that the consonant sequence for which the ‹pr› of apron stands may be an onset (initial consonant or consonant cluster of a syllable). If a written word is broken up so as to maximize the onsets of its syllables (Chee et al., 2020; Wiley et al., 2024), then ‹a› is the last letter of apron’s first syllable. This syllable does not have a coda (a final consonant), promoting the long pronunciation of the vowel. The consonant sequence to which the ‹pt› of captain corresponds cannot be an onset in English. Thus, the first element of this sequence must be in the coda of the first syllable and the second element must be the onset of the second syllable. The first vowel of captain may be short because its syllable is closed by a consonant.

The vocabulary analyses of Treiman et al. (2023) show that the ratio of long to short vowels differs not just according to the number of following consonants, but also according to the consonants’ phonological properties, but does this affect readers’ behavior? Short vowels outnumber long vowels in disyllabic words with medial obstruent + sonorant sequences (e.g., ‹pr›) and words with medial obstruent + obstruent sequences (e.g., ‹pt›), albeit by a smaller margin for the former, and no program of reading instruction that we know of teaches children to consider the type of consonant sequence in selecting between long and short vowels. Thus, we might expect readers to be influenced by the number of medial consonant letters, but not the phonemes to which they correspond. To find out, Treiman et al. (2023) asked university students to pronounce novel disyllabic items with different types of medials. The proportion of long vowels relative to the total number of long or short vowels was significantly higher before CC sequences that correspond to obstruent + sonorant sequences, such as ‹pr›, than before CC sequences that correspond to obstruent + obstruent sequences, such as ‹pt› (26% vs. 6%). For example, participants were more likely to use a long first-syllable vowel in ‹baproc› than in ‹baptoc›. This result suggests that skilled readers have picked up, without explicit instruction, the knowledge that the nature as well as the number of consonants after a vowel influences the vowel’s pronunciation. This knowledge may reflect statistical learning: implicit learning about the links between spellings and pronunciations. Statistical learning plays an important role in the learning of reading and spelling, as in the learning of other skills (Siegelman, Rueckl, et al., 2020; Treiman & Kessler, 2022).

According to some views, experienced readers have internalized the statistics of their language’s spelling-to-sound relationships and mirror them in their behavior (Brown, 1998; Siegelman, Kearns, et al., 2020). In the above-described study of Treiman et al. (2023), however, the behavioral data did not match the vocabulary statistics perfectly. For readers, for example, the proportion of long vowels relative to the total number of long or short vowels was 26% for disyllabic nonwords with medial obstruent + sonorant sequences. In an analysis of disyllabic English words with medial obstruent + sonorant sequences, the proportion of words with long vowels relative to the total number of words with long or short vowels was substantially higher: 44%.

In the present research, we sought to expand our knowledge about the role of context in spelling-to-sound relationships and the degree to which experienced readers’ behavior mirrors the statistics of the vocabulary. We did so by examining medials that have not featured in previous studies: those with ‹s› as their first letter. Sequences such as ‹sl› and ‹sp› are like the sequences such as ‹pr› that were studied by Treiman et al. (2023), in that they stand for phoneme sequences that can occur at the beginnings of English words. Reading researchers have, therefore, assumed that sequences like ‹sl› and ‹sp› are onsets (e.g., Chee et al., 2020; Kinoshita, 2000; Wiley et al., 2024). If so, we would expect the vocabulary statistics to show the same proportion of long first vowels in words with medial sequences such as ‹sl› and ‹sp› as in words with medial sequences such as ‹pr›. We would also expect readers to use long vowels as often before sequences such as ‹sl› and ‹sp› as before sequences such as ‹pr›.

However, linguists have noted that consonant clusters beginning with /s/ have some unusual properties in English and other languages, such as Italian (Kaye et al., 1990), raising questions about the behavior of the corresponding orthographic units. One difference between /s/ clusters and other clusters is that, for clusters appearing at the beginnings of words, only those with initial /s/ may have three elements (e.g., /spl/). Another difference is that, in initial clusters that begin with a consonant other than /s/, the second element (e.g., /l/ of /pl/), is always more sonorant or vowel-like than the first element (/p/). Thus, the cluster has a rising sonority profile. But /sp/, /st/, and /sk/ do not have such a profile. Indeed, in some views of sonority, the sonority profile falls from the fricative /s/ to the following stop consonant. According to some linguistic views reviewed by Goad (2011, 2012), /sp/, /st/, and /sk/ are not onsets, whereas clusters such as /sl/ and /sn/, with their rising sonority profiles, are onsets. In other linguistic views, all clusters with initial /s/, regardless of their second element, are represented differently from clusters like /pr/ and /bl/ (Kaye, 1992).

In the research reported here, we ask whether medial sequences beginning with ‹s› are similar to, or different from, obstruent + sonorant sequences with respect to their influence on the pronunciation of a preceding vowel. Our first study considered ‹s›-initial medial sequences as a group, comparing them to medial obstruent + sonorant sequences. The second study distinguished between different types of sequences beginning with ‹s›, including those whose usual pronunciation has a rising sonority profile (e.g., ‹sl›) and those without rising sonority profiles (e.g., ‹sp›).

The first component of each study was a behavioral experiment that examined university students’ pronunciations of vowels in disyllabic nonwords. For example, Study 1 included nonwords with medial ‹s› clusters, such as ‹kestep›, and nonwords with medial obstruent + sonorant clusters, such as ‹ketrep›. The study also included, for purposes of comparison, items with single medial consonants, such as ‹kesep› and ‹ketep›. Participants were asked to pronounce each item with emphasis on the first syllable and then select, from two options, the one that better matched their pronunciation of the first vowel. Thus, participants decided whether their pronunciation of the first vowel of ‹kestep› was “ē as in beet” or “ĕ as in bet.” The experiment also included well-known English disyllabic words with long and short vowels in the stressed first syllable. Inclusion of these filler items allowed us to check whether participants understood the task and were paying attention throughout the experimental session.

The second component of each study was a vocabulary analysis. Here, we analyzed the spelling-to-sound relationships in comprehensive lists of English words with the same medials as the items in the behavioral experiment. For example, the vocabulary analysis of Study 1 examined long and short vowels in the first syllables of disyllables with medial ‹s› sequences (e.g., whisper, postal), medial obstruent + sonorant sequences (e.g., progress, apron), single medial ‹s› (e.g., closet, resource), and single medial obstruents (e.g., widow, opal). Given that people’s reading of vowels in disyllables may be affected by their experience with words of more than two syllables as well as by their experience with disyllabic words (Treiman et al., 2023), we also included words with three or more syllables (e.g., testify, nucleus, prosecutor, stabilizes) in some of the analyses. Our vocabulary analyses go beyond those of Kearns (2020), who did not distinguish among CC sequences based on the phonemes to which the letters correspond, and Treiman et al. (2023), who did not include words with medial sequences beginning with ‹s›.

To summarize, the goal of the present research was to shed light on the role of context in spelling-to-sound translation by examining use of long and short vowels before medial sequences that begin with ‹s›. This structure provides a good case study for examining the role of context on vowel pronunciation in English and the degree to which skilled readers have internalized statistical patterns in the vocabulary.

2. Study 1

3. Study 2

4. General discussion

The English writing system is complex, to the point that some have suggested that “it is almost impossible to predict the actual pronunciation” of a letter string (Domahs et al., 2014, p. 68). However, studies with English monosyllables paint a more positive view, showing that the context in which a letter occurs can help to specify its pronunciation and that skilled readers take advantage of these effects (e.g., Kessler & Treiman, 2001; Siegelman, Kearns, et al., 2020). The goal of the present research was to expand our knowledge of context effects for items of more than one syllable, which have been underrepresented in previous research. We focused on two-syllable items with different types of medial CC sequences – a good test case for learning about the English writing system itself, how readers translate from spelling to sound, and the degree to which skilled readers have picked up patterns in the vocabulary without explicit instruction.

Educators who follow a phonics approach often assume that a vowel is generally pronounced as long if it is followed by a single consonant and then another vowel letter and that it is generally pronounced as short if it is followed by two consonant letters (Kearns, 2020). The results of our vocabulary analysis confirm that the ratio of long to short vowels in English words is affected by the number of following consonants. For example, long vowels are more common than short vowels when they occur before ‹s› followed by another vowel letter, but long vowels virtually never occur before ‹ss›. However, our results also show that the number of medial consonant letters is not the only systematic influence on the use of long and short vowels. The ratio of long to short vowels, rather than being the same for all types of medial consonant sequences, is affected by the characteristics of the phonemes that the consonant letters represent. Long vowels are fairly common, although not in the majority, before sequences that correspond to an obstruent followed by a sonorant (e.g., ‹pl›). Long vowels are much less common before medial sequences that begin with ‹s›, regardless of whether the following consonant corresponds to a sonorant (e.g., ‹sl›) or a stop (e.g., ‹sp›). Thus, English spelling-to-sound relations are more predictable than is often believed.

Our behavioral results are not consistent with the idea that readers rely on the rules of phonics, either because they were taught to do so as children or because they have induced them on their own. Users of phonics rules would not be expected to produce more long vowels before sequences such as ‹sp› and ‹sl› than before sequences such as ‹pr›, but our participants used fewer long vowels in the former case than the latter case. Users of phonics rules would be expected to produce primarily long vowels before single medial consonants, but our participants did not.

Another view within reading research is that readers break up written words so as to maximize the codas of their syllables (Taft, 1979; Taft & Álvarez, 2014). If so, the first orthographic syllable of all the items in our behavioral studies would end with at least one consonant letter. If readers tend to use a long pronunciation for a vowel in a syllable that is closed by a consonant, there is no reason to expect long vowel use to vary across types of items, as it did in our studies. Thus, the idea that readers form syllables via coda maximization is not supported.

A more common view in reading research is that readers form syllables by maximizing onsets and that sequences such as ‹sp› and ‹sl› are onsets, just like sequences such as ‹pr› (e.g., Wiley et al., 2024). If so, readers should have shown the same pattern of responses for items with medial sequences such as ‹sp› and ‹sl› as for items with medial sequences such as ‹pl›. For both types of items, they should place the consonant cluster at the beginning of the second syllable and pronounce the first vowel, which is in an open syllable, as long. However, our participants were much less likely to use long vowels before ‹s› clusters than before clusters such as ‹pl›.

As the preceding discussion shows, current theories of orthographic parsing and syllabification in reading do not account well for our results, including the difference in vowel pronunciation that we found between ‹s› clusters and obstruent + sonorant clusters. One linguistic interpretation of our results is consistent with Goad’s (2012) analysis of /s/-initial clusters. Goad showed that a variety of unique properties of such clusters, especially apparent violations of sonority sequencing, can be explained by theorizing that the /s/ belongs to the preceding syllable. That would mean that items like ‹feslum› and ‹feskum› syllabify as ‹fes.lum› and ‹fes.kum›. If readers are sensitive to that syllabification, and if they have a general tendency to choose short vowels more in closed syllables than in open syllables, it is understandable that they usually read these items with short first-syllable vowels. Another linguistic interpretation of our results is that /s/ before a consonant is especially susceptible to ambisyllabicity, being simultaneously a coda of the syllable to its left, thus favoring short vowels, and an onset of the syllable to its right, as expected by the maximal onset principle (Gussenhoven, 1986; Kahn, 1976). We thank a reviewer for the interesting suggestion that all consonants after stressed vowels could be ambisyllabic, but with a graded level of attraction to the two syllables. Differences in degree of ambisyllabicity would be compatible with the graded nature of our experimental data.

With regard to spelling-to-sound translation, our results support the idea that readers are sensitive to subtle statistical patterns, including those that are not explicitly taught. However, the behavior of experienced readers in this and other studies (e.g., Treiman et al., 2021) does not mirror the statistical patterns of the writing system to the extent expected by many statistical learning theories of reading (Brown, 1998; Siegelman, Kearns, et al., 2020). For example, most English disyllabic words with medial VCV sequences have a long pronunciation of the first vowel, meaning that people could maximize their chances of correctly pronouncing a new disyllabic word if they used a long vowel in this context. But our results show that people often use short vowels when pronouncing novel disyllabic items of this sort. There are several possible reasons, not mutually exclusive, for readers’ frequent use of short vowels. One reason may be that readers attend to both context-free and context-sensitive mappings and to different levels of contexts. For example, people may learn about the frequency of the ‹a› to /æ/ mapping irrespective of context and about its frequency in various contexts: before a consonant in all types of words and in words of specific lengths and before a CV in all types of words and in words of specific lengths. When trying to compute a pronunciation for the novel item ‹catab›, people could attend to any or all of those contexts, with competing or interacting results. A second reason for readers’ frequent use of short vowels in VCV contexts may be that they read some nonwords as compound words like login that have the root boundary and, therefore, a syllable boundary after the medial consonant. A third reason for adults’ frequent use of short vowels may lie in how they were taught the alphabet as children. U.S. children are typically taught that each letter has a name and a sound, the latter being how it is pronounced in words. The sound that is taught for vowel letters is the short pronunciation, and this early focus on short pronunciations may influence reading throughout life.

We turn now to models that are intended to simulate how readers translate spellings to sounds. Such models have played an important role in discussions of reading, because constructing a model forces us to consider in detail the workings of the reading process and because modeling can shed light on how reading develops and how it can go wrong. One way to test models is to ask whether they produce the same pronunciations for specific items that human readers do. Many models of spelling-to-sound translation work only for monosyllables (Chang et al., 2019; Coltheart et al., 2001; Harm & Seidenberg, 2004), but one model that produces pronunciations for disyllables is that of Rastle and Coltheart (2000), which we refer to as RC00. Built into this model are a set of all-or-none rules that determine whether the first vowel of a disyllable receives stress and, if stressed, whether it is short or long. Specifically, nonwords have first-syllable stress unless they have a defined stress-repelling prefix or stress-attracting suffix. When the first syllable is stressed, the pronunciation of its vowel is determined based on the vowel letter and specific characteristics of the following letters. For the stimuli in our behavioral studies, this amounts to pronouncing ‹a›, ‹o›, or ‹u› as long if the consonant that comes after the vowel is immediately followed by ‹e› or final ‹y› and as short otherwise. Table 3 shows that the RC00 model is compatible with our participants’ behavior in assigning fewer long vowels to items with two-letter medials than to items with single-letter medials (cf. Tables 1 and 2) and in treating the different types of ‹s› clusters in Study 2 the same. But the number of long vowels is much lower in all categories than the number produced by our participants – the root mean square difference between RC00 and the behavioral data across the eight categories is 26 percentage points in Study 1 and 21 in Study 2 – with most categories showing no long vowels at all. This model also did a poor job of accounting for behavioral data on vowel length in previous studies (Treiman & Kessler, 2023; Treiman et al., 2023).

The second main model that computes pronunciations for disyllables is the CDP++ model of Perry et al. (2010). This model can be seen as an instantiation of a statistical learning view, in that it includes a network that learns about the connections between spellings and pronunciations. The model is first trained on the simple phonics rules (e.g., ‹l› to /l/, ‹a› to /æ/) that are often taught to young children, and then with a large set of monosyllabic and disyllabic English words and their pronunciations. When we ran our nonword stimuli through the CDP++ program for Windows, we found that this model fit the behavioral data quite a bit better than RC00 did. The overall root mean square difference was only 4 percentage points in Study 1 and 13 points in Study 2. Like RC00, CDP++ reflected the behavioral pattern of using fewer long vowels before two-letter medials than before one-letter medials; unlike RC00, it also agreed with participants in using fewer long vowels when a medial cluster began with ‹s›. However, the CDP++ model showed pronounced differences between the types of two-letter ‹s› medials in Study 2, using more long vowels before ‹s› + sonorant clusters. In contrast, our participants treated all the ‹s› clusters roughly the same.

The fact that the CDP++ model accounts reasonably well for our behavioral results is consistent with the idea that adults’ reading of vowels is shaped by how they are taught the alphabet as children and by the statistics that they pick up by relating the spellings that they see in texts to words’ pronunciations. However, the CDP++ model needs further development, so as to capture the details of adults’ reading behavior. In future research comparing the behavior of models to that of human readers, it will be important to match for dialect. The RC00 and CDP++ models were meant for speakers of dialects other than the General American dialect spoken by our participants. We do not expect that differences should be substantial for our sets of items, which were designed to avoid major dialect differences, such as divergent treatments of ‹r›, but future studies will need to address this issue.

In future research, it will also be important to examine other linguistic factors that may influence the choice between long and short vowels. These include the identity of the vowel itself and that of the word ending. These factors have been studied for words with single medial consonants (Berg, 2016; Treiman & Kessler, 2023), and research is needed to determine whether these effects extend to words with two medial consonants. Also, it will be important to compare, within the same experiment, a task in which participants judge whether their own pronunciation of a specified vowel is long or short and a task in which their readings of nonwords are scored by phonetically trained judges.

Research on the role of context in vowel pronunciation is important, not only because of its implications for our understanding of the English writing system and the processes involved in reading, but also because of its implications for teaching. Some educational approaches, such as the idea that children should memorize the spellings of at least some words as phonically unanalyzable wholes or that they should use the pictures accompanying a text to guess words’ pronunciations, are based on the idea that English is highly irregular in its spelling-to-sound relationships (see Castles et al., 2018). Our results suggest that English is more predictable than often thought, but that we cannot rely on people to pick up all its patterns on their own.

5. Conclusion

In English, a letter may have one pronunciation in some words and another pronunciation in other words. Are there clues in the writing system about when one to use one pronunciation over the other, or must readers guess at random? In the present study, we addressed this question by examining long and short pronunciations of vowels in words of more than one syllable. The results of our vocabulary analyses show that the choice between long and short pronunciations depends on whether the vowel comes before a single consonant or a consonant sequence and, in the case of a consonant sequence, whether the sequence begins with ‹s›. The difference between sequences like ‹sl› and ‹pl› supports linguists’ views that /s/ clusters have unique properties, but not psychologists’ views that clusters like ‹sl› are orthographic onsets in the same way as clusters like ‹pl›. In behavioral studies, we found that adult readers are sensitive to the statistical patterns uncovered in the vocabulary analyses, even those that are not taught as a part of phonics instruction. However, adults’ behavior does not perfectly mirror the statistics of the vocabulary.

Appendix

Data accessibility statement

Data and analysis scripts for the studies are available at https://osf.io/by5gn/.

Ethics and consent

Ethical approval for the behavioral experiments was granted by the Institutional Review Board at Washington University in St. Louis; Reference numbers 202102148 and 202101025).

Acknowledgments

This study was supported by NIH grant R01HD102346.

Competing interests

The authors have no competing interests to declare.

Author Contributions

Rebecca Treiman: Conceptualization, Methodology, Writing – original draft preparation, Funding acquisition. Brett Kessler: Conceptualization, Methodology, Software, Formal analysis, Data curation, Writing – reviewing and editing, Funding acquisition.

ORCiD IDs

Rebecca Treiman: https://orcid.org/0000-0002-6819-8338

Brett Kessler: https://orcid.org/0000-0001-5759-0193