Phonological processes of consonants from orthographic to pronounced words in the Buckeye Corpus^*

The Buckeye Corpus consisted of forty speakers (see Pitt et al., 2007 for details). They were divided into two sex groups, 20 men and 20 women, and two age groups, a younger group under 30 and an older group over 40. All participants were speakers of American English born in Columbus, Ohio. They passed a dialect screening test through a telephone interview. The recruited participants joined one-hour-long individual sessions with a conversation on general topics, including personal backgrounds and views on miscellaneous issues.

The phonetically transcribed text files of the Buckeye Corpus were divided into sex and age group folders. Then, an R (R Core Team, 2019) script was created to collect an integrated text file to which all the files within each folder were appended as character strings (see Yang, 2018 for details). The author removed format errors and unrelated labels from the integrated file in Microsoft Excel. The total number of words in the file was 283,522. The number of consonants for the orthographic transcriptions alone amounts to 1,580,547 in 30 phonetic categories.

The integrated Excel file was imported onto the R Studio, and another R script was created to trace the phonological processes from the orthographic to the pronounced phonetic symbols of English consonants. The procedure was based on the syllabification rules prescribed in Noyer (2016) and the list of English onset clusters in Duanmu (2002) and Williamson (2014). In brief, the script initialized the Buckeye list of 33 vowels and 27 single consonants along with 54 two-symbol onset clusters and 9 three-symbol onset clusters. Then, a template matrix was created with one row and forty columns. Phonetic strings of the orthographic and pronounced words were extracted along with the number and position of vowels in the words. The syllabification rules were applied to find the onsets and codas of the words. When the number of vowels in a given orthographic word matched that of the pronounced one, one of the columns in the matrix was recorded as “same”; otherwise, it was recorded as “different”, which indicates variants. Some variants were syllabic deletions, such as the word “probably”. Its orthographic phonetic form was transcribed as [p r aa b ah b l iy], while one of the pronounced ones was noted as [p r aa l iy], in which the stressed first syllable was pronounced but the second syllable and one of the onset cluster of the third syllable were deleted. The different rows were manually checked to assign appropriate consonantal processes according to the corresponding components in the given syllable structure later. The number of vowels in a word was used to set a looping frequency in the script. Generally, the syllabification rules were applied to obtain the onset (cluster) first, and then the remaining consonants between adjacent syllables were assigned to the current or previous syllable as the coda. The same procedure was applied to the pronounced words, and the list for each syllable was appended to the result file. In addition, 14,231 (5% of all words) different rows of the file were manually checked to match the phonological processes from the orthographic to the corresponding pronounced syllables appropriately.

A third R script was created to combine rows of onsets or codas using the R function “rbind” for all speakers or for the groups of men and women and to write the table data counting the number of phonological processes. In addition, chi-squared tests and correlation analyses were conducted to determine the association of the frequency distribution of the phonological processes of consonants by sex and the degree of correlation between the two groups.

Table 1 lists the frequency distribution of the phonological processes of single onsets in the Buckeye Corpus. The list includes only cases with more than 50 variants to make the table simpler. In addition, consonant insertions were excluded. The same screening criteria were applied to create the following tables in this paper. Consonant insertion almost never occurred compared with the abundant deletions. For example, there were 3,246 /dh/ deletions and 2,751 /hh/ deletions for the onsets. On the other hand, there were 131 /l/ insertions followed by 114 /n/, 61 /r/ and 61 /w/ deletions for the onsets, while there were 822 /r/ insertions followed by 809 /l/, 152 /n/, 110 /s/, and 73 /ng/ insertions for the codas.

Generally, the table shows that the majority of the onset consonants were pronounced as the same orthographic symbols and with the same voicing quality in the conversation. Specifically, 19,670 /w/s were pronounced as the same consonants, followed by 19,291 /dh/s and 14,251 /s/s and 14,060 /l/s. The speakers in the Buckeye Corpus must have pursued this as a strategy to ensure that they would be clearly understood. We will discuss the important role of onsets for word identification in the following section with additional coda data.

The alveolar stops /d, t/ have relatively diverse variants. For example, /t/ lists 11 variants followed by the voiced fricative /dh/ and its voiced counterpart /d/. The palatalized /ch/ occurs only 159 instances, while its voiced counterpart /jh/ occurs 23 instances. These results might be related to the semiformal mode of the interview, which we will discuss below. On the other hand, the voiced stops in the orthographic forms tend to keep the voice quality except for the dental stop /t/; there are 1,442 instances of voiced /d/. The fricatives also seem to maintain the same sound quality and voicing quality. The occurrence of 996 instances of /dh/ to /th/ might be related to the high occurrence of the fricative itself. For example, the proportion of the “same” consonants in both the orthographic and pronounced forms (dh<dh) to the different consonants (dh<th) was approximately 5%. The voiceless counterpart /th/ accounts for approximately 6%. This paper uses the percentage of the number of the same consonant processes as that of the given variants to estimate the relative weight of variants. The deletion of the fricative /dh/ topped the list, followed by the fricative /hh/. The missing fricatives are mostly in the middle unstressed syllables, as can be seen in vowel studies (Yang, 2018). The deletion of the /l/ sound also accounts for approximately 10% of the same pronounced forms.

It is interesting to see the flapped variants /dx/ of /t/, which amount to 4,052 instances along with 585 occurrences of /nx/. Because the Buckeye Corpus has relatively fewer reductions or variants, the speakers were presumed to talk semiformally. Zue & Laferriere (1979) listed six phonetic environments of the phoneme /t/ in the medial position of English words. More detailed analyses of the variants that consider the phonetic environments are desirable to complete the study of these phonological processes. Nonnative speakers such as Koreans may need to pay attention to those variants in pronouncing and listening exercises in order to carry out a casual conversation with Americans.

The nasal /m/ lists the same process and deletion, while /n/ lists an additional nasal flap /nx/. Interestingly, no /m/ variants occurring over 50 instances were reported here. On the other hand, the rate of the same consonant to the nasal flap of the orthographic /n/ is approximately 37%, which seems remarkable. In addition, 1,251 /n/ deletions occur in the spontaneous speech. The reason for the robust use of /m/ might be related to the distribution of English consonants in the place of articulation: English has fewer labial consonants and variants available than alveolar consonants and variants. Here, the labial /m/ occurs more in the onset than the alveolar /n/ does. The labial /w/ lists a few variants, and its deletion rate amounts to approximately 7.5%. The deletion rate of /r/ is 21.6%, while that for /l/ is 9.4%. We may say that the lateral /l/ is more robust to changes in the phonological process. The glide /y/ has many variants, which might be related to the fact that the glide tends to be combined with the preceding codas or the following vowels to form consonant clusters or diphthongs. On the other hand, we also note that the voiceless fricatives /f, s, sh/ prevail in the onset positions. Among them, the fricative /s/ is the most prominent. We will return to the frequency distribution of consonants by syllable components in the following coda section.

Table 2 lists the phonological processes of the onset clusters. The majority of the onset clusters have fewer variants than the single onsets and almost no clusters with more than 50 deletions except the cluster /bl/. That cluster lists four variants and 151 deletions. The cluster /tr/ has almost comparable variants of /tr/ and /chr/. The voiced counterpart /dr/ changed to /jhr/. The variants are palatalized alveolars. The fewer variants may account for the strong association of the two or three single consonants. If one of the onset clusters changes, listeners may have difficulty guessing the words. The cluster /pr/ lists 1,535 counts, while /pl/ lists 758 counts. A similar trend of a preferred combination of a consonant plus /r/ can be seen between /fr/ (987) and /fl/ (217) and between /gr/ (979) and /gl/ (76). However, the cluster /br/ lists 470 instances against 597 instances of /bl/. It seems very difficult to generalize the trend because the occurrence of clusters largely depends on both the phonological feature of the clusters themselves and the frequency of everyday expressions with the clusters. On the other hand, we note that the distribution of the onset triple clusters in the inventory varies. The cluster /str/ is most prevalent, with 234 instances, followed by /skr/; the 47 instances of /spr/ are not listed on the table. Whether the uneven distribution is related to the perceptual prominence of the clusters would be interesting to study in the future.

Table 3 presents the phonological processes of single codas. The deletion rate of the lateral /l/ amounts to 32.1% of the same consonant processes. Similarly, we can see large deletions of /n, v/. Moreover, the number of variants in the codas is smaller than the number in the onsets. The alveolar consonant /t/ lists nine variants followed by seven /ng/ variants. This result might be related to the inclusion of all the onsets which occur in the initial, middle, and final syllables. Moreover, the number of deletions in the onsets is relatively smaller than that of the codas. For example, 14.6% of the onset /t/ to the same consonant process was deleted, while 42.2% of the coda /t/ was deleted. This trend may be attributed to the information weight of the syllable components. Generally, the onsets seem more important in the delivery of ideas than the codas. Cutler (1982) described the beginning of a word as the most salient part for identification. In other words, speakers immediately guess words without waiting for their final coda. In addition, Yang (2018) mentioned that speakers tended to preserve vowels that were important to convey their thoughts but tended to change or delete vowels that were relatively unimportant to deliver their intended message. The same generalization may be applicable to the current results.

We described the phonological processes, but here let us briefly examine the consonant distribution in light of voicing and syllable components. Table 3 shows that the voiced /z/ prevails over the voiceless /s/, which occurs almost three times. The same trend can be seen between the fricatives /f, v/. It is noticeable that the onsets in Table 1 show the reverse trend: the voiceless consonants prevail. On the other hand, the voiceless stops in the coda still prevail. The stop /p/ occurs 5.4 times that of its voiced counterpart /b/. It seems difficult to generalize this conflicting distribution among the onsets and codas in the current corpus. Further analysis of more spontaneous speech corpora may be necessary to draw a conclusive statement on the phenomenon. We can also note that the voiceless stops /p, t, k/ occur much more frequently than their voiced counterparts. The voiced /b/ and /g/ occur 352 and 629 instances, respectively. Maddieson (1984) attempted to establish universals across world languages and reported that 80% or more languages had the consonants /p, t, k, m, n, s, j/. In addition, Eckman (2004) reviewed the Markedness Hypothesis and posited that the notion of typological markedness might be closely related to language universals (see the review in Yang, 2012). Voiceless consonants occur in languages more often than voiced consonants. Thus, voiceless consonants are considered simpler and more natural than voiced ones. However, we note that the distribution of consonants is different in onsets and codas. Thus, an observation of the total inventory distribution of consonants may have to be reexamined in light of the syllable components. In other words, if we combine the occurrence of the onsets and codas into one consonant category, the frequency of voiced and voiceless consonants may offset and obscure the individual characteristics of the given consonant category according to the syllable components.

Table 4 lists the phonological processes of coda clusters in the corpus. The coda clusters have many more variants than the onset clusters have. Additionally, relatively greater deletions occur in the clusters /nd, nt, st, kt/. The rate of processes from the clusters /nd/ and /nt/ to a single coda /n/ amounts to 26.9%. The first component of the cluster /st/ remains the same in 2,353 instances, while 1,621 /st/ instances remain the same in the pronounced forms. Many coda clusters keep their orthographic forms in the pronounced forms. Homorganic consonant clusters are expected to be deleted in the pronounced forms because it would be rather difficult for the participants to produce the clusters with a different manner of production almost simultaneously. For example, the majority of /nd, nt/ and /st/ become one /n/ and /s/, respectively. However, the similar homorganic clusters /ld/, /lt/, /lz/, /ls/, and /ts/ tend to remain the same.

Here, we discuss the frequency distribution of the clusters. Nasal, alveolar and lateral flaps are prevalent in the coda clusters. The majority of the clusters start with the consonant /n/ followed by /l/ or /r/. Among them, the consonant /n/ is most favored as the first component of the clusters. The frequency distribution of English consonants may have to consider these consonant clusters without separating them into single consonants. The frequency distribution of the cluster /ts/ is greater than that of /st/. In Table 2, we observed 2,873 onset /st/ clusters. Here, the mirror image of /ts/ has 3,346 instances. Clusters consisting of the fricative /s/ followed by other consonants are considered problematic for sonority analysis. The sonority sequencing principle states that “only sounds of higher sonority rank are permitted between any member of a syllable and the syllable peak (Clements, 1990:285)”. However, we have seen that the clusters /st/ and /ts/ do not follow the symmetry in syllable structure. In other words, the fricative /s/ is considered more prominent than the stop /t/ in a sequence. The statistical distribution of /st/ in the corpus might provide a gradient explanation of the asymmetry depending on the syllable components, which will examined in the future.

Tables 5 and 6 list the frequency counts of the phonological processes of men and women in the Buckeye Corpus. Instead of listing all the processes exhaustively, 94 instances of the onset and coda processes were chosen to conduct a statistical analysis of the association between the two sex groups. The usual approach would be to conduct the test over a large table of all phonological processes. However, some slots would have too few occurrences for the standard chi-squared test.

Table 7 lists the statistical results of the chi-squared tests and correlation coefficients between the frequency values of the onset and codas of men and those of women in the Buckeye Corpus.

The result of Pearson’s chi-squared test between the onsets of the men and women was statistically significant (χ²=1,200.5, df=92, p<.05). On the other hand, the result of the chi-squared test between the codas of the men and women was also statistically significant (χ² =1,863.3, df=93, p<.05). We can conclude that there is a significant association between the phonological processes for both the onsets and codas of men and women in the Buckeye Corpus. In addition, Pearson’s product-moment correlation between the paired data was conducted to obtain r=0.983 (p<.05) for the onset data and nearly the same or stronger positive correlation coefficient for the codas, r=0.995 (p<.05). The two statistical results indicate that the phonological processes of the two sex groups correlate strongly with each other. It is quite interesting that the two sex groups show a proportionate ratio in their frequency distribution. Another noticeable trend is the deletion processes of men and women depending on the consonants. With /hh/, /r/, and /y/, men tend to delete more consonants than to preserve them compared to women. For example, 634 instances of /y/ of the onsets of men were deleted compared with 223 instances of /y/ of women. On the other hand, women tend to practice the same consonant processes more often than men. This result may reflect women’s tendency to produce the sounds clearly as in the orthographic symbols. However, more cases of the same consonant processes for men than for women were also observed for the onsets /l, m, n, t, w/, a result that needs further study.

Here, we can extend the generalization on vowel processes (Yang, 2018) by including the current result on consonant processes: The distribution of the phonological processes of English vowels and consonants indicates a significantly strong association between men and women in the spontaneous speech corpus. As described in the previous study, this generalization may represent sex characteristics fairly well with sufficient data from the corpus.

Thus far, we have examined the general association between the data for men and women. If we go further and categorize those consonants by manner and place, we may find some specific association patterns, as in Kessler & Treiman (1997). They reported that the vowel-coda association was always stronger than the onset-vowel association after analyzing association patterns among the onsets, vowels, and codas of 2001 monomorphemic CVC words in the unabridged Random House Dictionary (Flexner, 1987). They added that the unequal associations indicate greater distinctiveness at the onset, which promotes more efficient and distinctive production and perception at the beginning of English words.