Phonological processes of vowels from orthographic to pronounced words in the Buckeye Corpus by sex and age groups^*

The Buckeye Corpus was built by researchers (Pitt et al., 2007) at Ohio University. It consisted of individual recordings by forty speakers in Columbus, Ohio, which occurred through a spontaneous conversation with one of the two interviewers. The speakers were ten participants representing each sex and age group so that patterns specific to a group could be examined. The corpus offers both audio and text files of the speech as well as orthographically transcribed and phonetically labeled productions.

Orthographic and pronounced phonetic symbols of spontaneous speech were collected to facilitate a comparison and analysis of the phonological processes between them. The procedure was as follows. First, 255 phonetically transcribed text files of the Buckeye Corpus were copied manually into four group folders by sex and age group (i.e., young males, old males, young females, and old females). Then, an R (R Core Team, 2018) script was written to create an integrated text file to which all the files within each folder were appended as character strings using the table writing function of the R library “reshape.” The first eleven lines of each text file and the final line contained file header and file end information. Hence, these lines were excluded in the appending procedure. After checking the integrated file in Microsoft Excel, the author corrected file format errors and unrelated labels (for example, B_TRANS, SIL, VOCNOISE, UNKNOWN) after repeatedly sorting the data by each column. Nasalized vowels with the phonetic symbol Vn were included after removing the nasal consonant. For example, the symbols aan and ihn were changed to aa and ih, respectively. All syllabic el’s, em’s, and en’s were discarded for purposes of analysis. The total number of words analyzed was 283,522.

The corrected Excel file was imported onto R Studio, and another R script was created to trace the phonological processes from the orthographic to the pronounced phonetic symbols and to extract factors for statistical analysis. The procedure was as follows. First, the script initialized the list of 33 vowels and created a temporary matrix of eighteen columns. Phonetic strings of the orthographic and pronounced words were extracted along with the number and position of vowels in the words. The vowel processes were divided into the same or different orthographic and pronounced vowels and placed in corresponding columns to determine their frequency distribution. If the number of vowels in a given orthographic word precisely matched that of the corresponding pronounced word, then each position of the vowel was assigned to the matrix column in the same order. The entire word was also assigned to one of those columns to facilitate later visual checks. The same number of vowels in both orthographic and pronounced words accounted for most vowel processes. The sex and age variables were also added to the matrix columns. After extracting all pertinent information from each word, the matrix was appended to a final output file. If the numbers were unequal, then all information except the pronounced symbols were saved onto the matrix for manual verification. Instances of syllables of some frequently used words were significantly reduced in the pronounced forms of the Buckeye Corpus. For example, the word definitely had an orthographic label of d eh f ih n ih t l iy but was pronounced as d eh f ih en l iy, d eh f ah l iy and d eh f l iy. The four vowels were reduced to two or three vowels. Here, the stressed first vowel retained its vowel qualities in the pronounced labels. It might be possible to create a script to match the vowels case by case if a script were meticulously organized. Instead, the author checked incongruent words manually because the size was manageable and because in so doing, general patterns could be observed. The steps included finding the corresponding syllables between orthographic and pronounced words and assigning them to appropriate columns in the Excel file. To test dependence between male and female groups and between young and old groups, Pearson’s Chi-square value and correlation statistics were generated in R using the frequency data.

The total number of 283,522 words in the Buckeye Corpus had 365,002 vowels in the orthographic labels and 354,635 vowels in the pronounced ones. Of the orthographic words, 269,291 (95.0%) had the same number of vowels as did the pronounced words, with only 14,231 (5.0%) words differing between the two sets of words. This result proves that vowels tend to retain their quality in spontaneous speech. Among the many vowel processes, deletion occurred 12,204 times (4.3%, See Johnson, 2003 for discussions on massive deletion), nearly six times as often as insertion, which occurred 2,027 times (0.7%). Thus, we can conclude that deletion processes are preferred in natural speech. The total number of vowels indicates that changes are not one-to-one. In other words, each count of the pronounced words for a given syllable number may involve both deletion and insertion processes. For example, the word elementary with five vowels was realized as eh l ah m eh n t r iy, and eh l m eh n t r iy, including the full pronunciation with five vowels. Thus, one vowel or two vowels of the word can be deleted, resulting in a three- or four-vowel word in spontaneous speech. Of course, the insertion of a vowel is involved with the word job, for instance, as jh aa aa, or jh aa b ah. Deletion mostly occurred in words containing more than two syllables; thus, the number of one-syllable words increased in spontaneous speech. Although the absolute number of deletions tends to decrease by vowel position, the proportion of total orthographic labels increased. For example, 586 of the fourth vowels were deleted, accounting for 17.3% of 3,389 orthographic labels. Additionally, 35.5% of the fifth vowels were not pronounced. The reason for this finding might be related to the fact that compared with insertion processes, deletion processes would be more effective in decreasing a speaker’s efforts and accelerating their rate of speech. For example, the word especially was labeled ih s p eh sh ah l iy in the orthographic form but was produced as s p eh sh in the pronounced label. Nearly three of its four vowels were deleted. Another example is the three-syllable word anything, which was pronounced as one syllable, i.e., th ih ng. The two-syllable word because was produced mostly with k z and sometimes, with one consonant, k. We will return to the process in the following section to generalize deletion in the case of unstressed syllables.

Additional analysis found that 165,078 (58.2%) orthographic words were pronounced with the same vowels in the orthographic labels, whereas 118,444 (41.8%) orthographic words were pronounced with different vowels.

In the table, most words from the spontaneous speech data consist of one-syllable words, and the number of vowels in the words decreases as the syllable becomes longer. Specifically, one-syllable words account for 78.3% of the data, followed by two-syllable words, which represent 15.6% of the data, and finally, three-syllable words, accounting for 3.5% of the pronounced labels. For example, one-syllable words such as like, know, my, and one have the same phonetic labels in both orthographic and pronounced words. This result might occur because speakers attempt to convey the meaning clearly to the interviewers by keeping the words intact without deleting or inserting vowels to facilitate their pronunciation. The reason that one syllable word remained intact may be related to the fact that its meaning may not be relayed clearly without it. Thus, we can infer that speakers tend to retain only one vowel in spontaneous speech. However, certain function words were reduced to one consonant (i.e., the word has was pronounced s or z), which might be easily identified based on the context. For another example, the function word was was pronounced as a reduced form s or as a two-syllable word, uw ah z. The number of vowels was reduced mostly in words with more than two syllables, as shown in <Table 1>. The longest eight-syllable word was socioeconomically. The seven-syllable word homosexuality occurred nine times.

In the case of the Seoul Corpus, Yang (2016b) reported that the most frequent syllable type in Korean was 3-syllable phrases (178,185), which accounted for 32.6% of the phrases, followed by two-, four-, and one-syllable phrases.

The Korean syllable phrases included particles indicating syntactical information such as subjects, objects, and others so that the number of syllables in the original word might decrease by at least one syllable. In addition, the Carnegie Mellon University Pronouncing Dictionary listed two-syllable words as the most frequent, followed by three- and one-syllable words (Yang, 2016a). Yang assumed that two-syllable words would better express different names for a greater number of objects and actions that are used daily, without overloading human memory space.

The contingency table of the orthographic symbols to pronounced vowels is given in <Table 2>. To ensure legibility of the crowded columns in the table, five cases of ao>ao for the example word bought in both the orthographic and pronounced forms were deleted. The rare vowel might be related to either the speakers’ dialect or inappropriate transcriptions. One ey and two ow’s were realized as ye or oa, respectively, in the phonetic labeling. These cases were deleted because of incorrect transcriptions, not in the list of the aligner’s phonetic alphabet (See Table 2, Kiesling et al., 2018). In addition, 1,062 er>r’s are not in the table, whose changed consonants will be examined in a future study. The syllables el, em, en, and eng and certain orthographic vowels were produced as the semivowels w and y; these cases were also removed from the list. If a nasalized vowel combined with a full phonetic vowel (i.e., aan or eyn), then those vowels were included in the frequency count after removing the nasal consonant. The number of removed symbols was 15,235. Nine monophthongs were chosen to display the frequency distribution of both the same and different vowels in <Figure 1>.

Figure 1. Frequency distribution of vowel processes in 3D format. The y-axis indicates the frequency of each vowel occurrence.

Download Original Figure

The total number of vowels in the table is 347,793, which represents 95.3% of all the vowels in the orthographic labels (see <Table 1>). The most frequently pronounced vowels were ah (69,220), followed by ih (63,096), eh (36,435), iy (33,514), and ay (23,412). The diphthong ay ranks the sixth most frequent (12,384) vowel for the pronoun I. The pronoun was produced mostly as ay, along with a few variations such as ah, ao, and aw. The back vowel ah occurred slightly more often (68.6%) than did the back vowel aa (60.8%). The least frequently pronounced vowels were oy (649), aw (4,990), uh (5,917) and ao (8,452). The high back vowel uw (11,386) was the sixth least pronounced vowel.

The most common same vowel processes were oy (480 occurrences, which account for 97.4%), er (10,766, 84.7%), aw (4,826, 80.5%), ay (23,019, 78.6%), and uh (3,165, 76.7%), all of which are back vowels. The back vowel uw (10,838) had 44.8% of the same vowels. The average percentage of all back vowels with the same vowel processes is 72.9%. When the lowest uw is excluded, the average goes up to 76.4%. The average percentage for the five front vowels (iy, ih, ey, eh, and ae) is 68.5%. The lower rate of the front vowel processes might be related to the economy of production in spontaneous speech that relies on a smaller jaw movement. Wright (2003) reported that the vowels of words that people speak more frequently in their daily lives tend to be produced with reduced articulatory gestures; thus, vowel spaces decrease. Peripheral high vowels (iy, 76.6% and uh, 76.7%) show a higher percentage rate compared with peripheral low vowels (ae, 53.5% and aa, 60.5%). The speakers might have wanted to enunciate more clearly for the listeners by expanding their jaw opening. Interestingly, 44.8%, less than half the percentage of the high back rounded vowel uw, retained its vowel quality. This result might be explained by the fact that the speakers cannot further expand their movements upward to pronounce the higher vowel. In addition, the peripheral tense vowel uh showed a considerably lower rate of occurrence (3,165, 44.8%) than that of its lax counterpart uw (10,838, 76.7%). Some tense vowels were realized as ih (23.5%) and ah (16.6%). Despite the high percentage rate, lax vowels (11,386) occur nearly twice as often as tense vowels (5,917) do. In this case, the frequency must be considered along with the percentage in the description.

This trend may be applicable to the other tense and lax vowel pairs. The tense and lax vowel pair iy and ih exhibit an alternation of nearly double the ratio as the lax to tense ratio. Of the orthographic iy’s, 6.8% were produced as the lax counterpart ih, while 11.8% of ih’s became tense. Thus, one could surmise that in spontaneous speech, tense vowels tended to be pronounced as their lax counterparts.

Although the number of occurrences was low, the diphthong oy (493, 97.4%) was realized as the same one, followed by ay (23,019, 78.6%), ey (13,883, 64.9%) and ow (17,808, 63.8%). In addition, 6,489 vowels (1.9% of the total vowels) of the orthographic words were deleted in the pronounced forms (Null). A negligible number of vowels were added to the pronounced words. The number of null vowels and one-syllable words increased to 2,835 and 838, respectively, in the pronounced forms. Most null vowel words were interjections such as mm as em or hum-hum as hh em em. As described in the previous section, the syllabic em was not classified as a full vowel; thus, it was excluded from the table.

From detailed observations of the final data, we found that the vowels in a stressed syllable typically remained in the pronounced form. For example, the second vowels of the words police and remember were always pronounced in spontaneous speech. The function word about has many different forms such as ah, b aa, b aa dx, b ah, b ah dx, b ah t, b aw, b aw dx, b aw t, and b aw tq. Here, all the pronounced forms have the stressed second vowel, possibly because this stressed syllable is crucial to identifying the intended word. We also found that a suffix in the final syllable position in a word with more than three syllables tends to be preserved to facilitate clear communication. For example, the words definitely and especially were pronounced as d eh f l iy and s p eh sh l iy. Deleting the final suffixes in these cases may cause listeners to wait for a noun to follow. Overall, we can generalize that speakers tend to preserve those vowels that are important to convey their thoughts but to change or delete vowels that are relatively unimportant to communicate their intended message. Additional perceptual studies on spontaneous speech would be desirable, specifically studies involving a detailed analysis of the role of each vowel in a word.

Other cases account for a small proportion of observations, but nevertheless merit attention. The orthographic vowel ey (3,947, 18.4%) was pronounced as ah. The word butt is transcribed as b ah t. Many speakers produced the indefinite article a (ey in the orthographic label) as ah including the contracted form they’ll as ah l, and dh ah. The article is generally produced as ey when the speaker is emphasizing the single count of a particular type. Otherwise, it is produced as schwa. The word schwa in Hebrew means nothing or no vowel, referring to its short and indistinct vowel quality (Bauman-Waengler, 2009). The aligner’s phonetic alphabet does not contain schwa. Schwa is the commonest vowel in unstressed syllables (Davenport & Hannahs, 1998). The only schwa label is used with the consonant r, i.e., er. That broad transcription might have limited the phonetic labels of the Buckeye Corpus. This limitation might owe to the high frequency (69,220) of ah in <Table 2>. A similar limitation is apparent in the orthographic vowel aa. Many aa’s (4,821, 26.9%) were pronounced as ao. Two hundred nineteen cases of the word call were transcribed as k aa l in the orthographic form and as either k ao or k ao l in the pronounced form. Moreover, the orthographic vowel uw (5,697, 23.5%) was pronounced as ih. The reason why this vowel has a higher frequency may be because many speakers produced the frequent word you (y uw in the orthographic labels). The vowel was pronounced as ih mostly in such combinations as you’ve, you’re, and you’d. Thirty-four yourself’s were represented as y ow r s eh l f, whereas six yours’s were represented as y ow r z. We also observed cases of inappropriate vowel labels in such frequently occurring words as eleven, election, expect, and experience, whose first vowel was labeled eh in the orthographic form but ih in the pronounced form. Thus, the proportion of vowel processes may require further comparison with additional varieties and rigorous checks before a general trend can be discerned.

The selected distribution of the vowel processes is provided in <Table 3>. Instead of listing all the cases, twenty-six topmost vowel processes in descending frequency list (thirteen vowel processes each from the same and different groups) were chosen to test the dependence of sex and age to facilitate a balanced comparison. The thirteen vowel processes account for 298,729 vowel processes, or 81.8% of the total vowel processes (365,002), while vowel processes from the young and old groups account for 298,738 vowel processes, or 81.9%. Notably, the sex group difference in the same vowel processes was nearly negligible, accounting for 6,374 vowel processes, 1.8%. A difference in the same vowel processes owing to age was observed in approximately 19,573 instances, or 5.4% of the vowel processes. A difference in those vowel processes could be attributed to sex in 2,869 instances, or 0.8%, and to age in 4,951 instances, or 1.4%.

To ascertain statistical dependence from the table, Pearson's Chi-square tests were conducted on the frequency distribution of various sex and age groups. First, the statistical result on the sex group from both the same and different vowel processes was significant at p<.05 level (χ²=1,127.1, df=25). This result indicates that the male and female groups were strongly dependent in the frequency distribution. Nearly the same significant result was obtained between young and old groups at p<.05 level (χ²=689.9, df=25), which also showed a strong dependence in the frequency distribution. The paired correlation between the two sex groups was significantly strong with r=0.989, as was that of the two age groups, whose r=0.992. Second, further detailed analyses separating the sex and age factors from the factor of same or different vowel processes were conducted to examine dependence. The statistical results for the same vowel processes were significant at p<.05 with respect to sex (χ²=449.2, df=12), r=0.987, and with respect to age (χ²=390.5, df=12), r=0.988. The statistical results for the different vowel processes were also significant at p<.05 with respect to sex (χ² =660.0, df=12), r=0.822, and age (χ²=298.9, df=12), r=0.926. The correlation coefficient of the sex group was slightly lower for different vowel processes than for other vowel processes, which might be caused partly by the relatively wider deviation of the male and female frequencies. The table indicates that ah>ah showed a visible difference between the sex groups. In the different vowel processes, the female processes ae>eh and ah>ih had an opposite distribution. The age group comparison showed a moderate difference in the distribution of the same vowel process ah>ah and the different vowel process uw>ih. Yang (2017) reported that the Korean female group in their forties changed the diphthong yo to yv at the end of the pronounced phrasal words more often than the male group did in the Seoul Corpus. If we pursue the sociological trait further after applying narrower transcriptions and dividing the four groups of the Buckeye Corpus by age, we may be able to identify a sociological trend in terms of either sex or age. Currently, however, the low frequency and transcription errors prevent further analysis. Thus, we tentatively offer the general conclusion that regardless of sex and age or whether the vowel processes are identical or different, the vowel processes occurred in nearly the same pattern in the natural spontaneous speech data. The results appear comparable because the corpus contained sufficient data. With less biased data, the vowel processes might have been mostly not significantly dependent.