Confusion is much more prevalent between pairs of back vowels. The confusion structure suggests that the lax back vowels are insufficiently differentiated from the adjacent tense back vowels. Typically, half close /ʊ/ is poorly contrasted with /u, o/, while half open /ʌ/ is poorly distinguished from open /ɑ/. The asymmetries between members of confused pairs suggest that /ʌ/ is pronounced too far back, while /u/ should be raised relative to /o/. Since /ɔ/ is identified as /o/ in half of the cases, the best advice is to teach PA learners of American English that /ɔ/
does not exist and that all low back vowels should be pronounced as /ɑ/.
/o, ɔ, ɑ, ʌ/, and /ʊ, u/). Mainly, mid-back and mid-front vowels, which are NEW vowels for PA EFL learners, were involved in the confusions. Such overlapping configurations between adjacent new and similar vowels (relative to the PA inventory) are expected to negatively affect the intelligibility of the PA EFL speech production. The PA production results in this relation show a tendency in the learners’ responses to keep the AE tense-lax vowels separated, which suggests that EFL learners were able to specify sufficient weight to spectral qualities as they do for temporal ones. Combining the spectral and temporal qualities for PA vowel production yielded a better still configuration of the AE vowels as learners cue duration as a factor related to the tense-lax contrast among AE vowels.
My first question was concerned with the PA learners’ production of the AE vowels compared to the AE native speaker results. The informal and short answer is that nonnative vowels differ greatly from the native norm. To answer this question formally, two types of classifications (i.e., LDA and MLRA) were performed twice: once for spectral data (LDA 63.8% and MLRA 65.8% correct identification) and a second time with duration as an added predictor (LDA 74.7% and MLRA 77.4% correct). These results show that the MLRA results for spectral data and when adding duration were slightly better than LDA classifications. The addition of duration as a third predictor increased the percentage of correct vowel identification in the PA results in both types of classification: +11% using LDA and +12% using MLRA.
These initial results only show how well PA EFL learners separate their AE vowel monophthongs. When training the automatic classification algorithms with the native production data results, the nonnative results can be classified by the native model, which allows me to see how the nonnative AE vowels would be classified by native listeners. Here, again, MLRA classification showed an overall better classification for PA results than using LDA. The overall nonnative results in this classification showed a drop in correct vowel identification, but better results were still yielded when adding the (normalized) duration predictor, which makes sense for speakers of L1 with a vowel duration contrast. This confirms my first hypothesis that PA learners have an incorrect (confused) conception of the AE vowel space. My results show that PA EFL speakers are aware of the spectral and temporal features of L2 AE, but they assign more weight to the latter than the former, which partially confirms the second part of my first hypothesis.
My second question asked which AE monophthongs are difficult for PA EFL learners to produce. To answer this question, I first presented the native results and combined the three predictors (F1, F2, Duration). The native speakers produced their vowels in such a way that they were hardly ever confused by the automatic classification algorithms. All vowel types were
correctly identified on acoustic grounds at 85% correct or better, with the notable exception, of course, of the cot-caught vowel pair. The nonnative results paint a different picture. PA EFL vowel production is confused, especially in the back area of the vowel space with AE NEW vowels. Combining the three predictors for PA results revealed the difficulties in correctly discriminating the vowels yielding confusions such as /u/→/o, ʊ/, /ʊ/→/o, u/, /o/↔/ʊ, ɑ/ (but not /o/↔/u/), /ɑ/↔/ʌ, ɔ, ʊ/ (but not /ɑ/↔/o/, /ɔ/↔/o/, /ʌ/↔/ɑ/, /ɪ/↔/ɛ/, and finally and most importantly, /ɛ/↔/ɪ/, which was confused in 68% of the decisions. It is noted that all these confusions involve NEW AE vowels for PA EFL learners, and confusions with similar vowels between the two languages can be better produced (and automatically identified) in terms of temporal differences. Additionally, there are asymmetrical relationships between confused pairs. The percentage of confusion for members of each confusing pair is not equal and not always reciprocal. For example, /u/ is confused as /o/ 33% of the times but not vice versa (/o/→/u/ never reached the 10% confusion criterion to be considered problematic for PA EFL learners), same for /o/→/ɑ/ 14%, but /ɑ/→/o/ never reached the 10% criterion. This suggests an asymmetry such that high vowels are pronounced too low, causing high (close) target vowels to be identified as lower (more open) vowels. Confused pairs were not limited to contrasting members of tense~lax pairs but also occurred within tense vowels (e.g., /u/↔/o/, /o/↔/ɑ/), or within only one lax vowel pair in the front area (/ɪ/↔/ɛ/). Confusions were also noted in the AE native results, in asymmetries favoring high vowels (no misclassifications of high vowels as lower vowels ever reached the 10% criterion needed to be considered problematic). Here, I exclude the more or less symmetrical confusion between the two low-back vowels /ɑ/↔/ɔ/, which is a single merged vowel category in the speech of Californians today. The asymmetrical confusion of /e/→/ɪ/ (15%) is because the diphthongization of /e/ was not measured in the acoustic analysis. A solution would have been to represent the (semi-)diphthongs /e/ and /o/ not by the mean formant values over the entire vowel duration but to extract the formants at 25%
and at 75% of the vowel duration and compute the difference between the early and late point in time as a measure of diphthongization (see, for instance, Van Bezooijen & Van Heuven, 2010: 362-363). Interestingly, the /ʌ/→/ʊ/ confusion (10%) coincides with another vowel merger in progress, whereby Californian speakers no longer distinguish between the two (Van Heuven et al., 2020: 124). This confusion did not occur in the vowels of the PA learners; instead /ʌ/ was incorrectly identified as /ɑ/ in the PA results (39%), while the reverse confusion occurred in 24% of the cases. It seems that PA EFL learners spot the differences between new AE back vowels but still confused their constriction place and tend to raise their tongues and
open their mouths more when trying to produce NEW back-low vowels in addition to confusing the vowels’ duration.
On a second look at the results, tense vowels do not have reciprocal confusions except for the /ɑ/↔/ɔ/ pair, which also occurred in the native results (see above). The PA learners, however, showed smaller confusion percentages for this pair of vowels than the AE natives.
Actually, this confusion is the only one in which PA EFL learners showed better performance.
Interestingly, PA production of lax vowels is always reciprocal in any confused (acoustically overlapping) pair, except for asymmetrical /ɑ→/ʊ/ (but not /ʊ/→/ɑ/). So, whenever a lax vowel is wrongly identified as a tense vowel, the confusion is reciprocal, but not necessarily so when a tense vowel is confused with a lax vowel. This can be attributed to the fact that duration is the main feature that is phonemically salient for PA learners. These confusions are better identified based on the nonnative duration results in the production task.
The results show a good separation of similar vowels between PA and AE front vowels.
Additionally, two pairs of confusions found in AE were not materialized in PA results; those are the /e→ɪ/ and /ʌ→ʊ/ pairs.38 PA learners also showed better separation for the reciprocal confusion for the pair /ɑ→ɔ/ than the AE natives did. PA EFL results show that the level of difficulty in producing pairwise AE vowel contrasts is: /ɛ→ɪ/, /ɔ→o/, /ʌ→ɑ/, /u→o/, /ʊ→u/, /ɑ→ʌ/, /ɔ→ɑ/, /u→ʊ/, /ʊ→o/, /o→ɑ/, /ɪ→ɛ/, /ɑ→ʊ/, /ɑ→ɔ/ and /o→ʊ/ in a descending order of incorrect identification by LDA trained on native AE vowel tokens (see Table 5.6 for details).
Table 5.6. Incorrect identification (%) of intended vowels produced by PA EFL learners by LDA trained on native AE vowel tokens. The symbol before the arrow denotes the intended vowel, and the symbol following the
arrow is the incorrect prediction by the algorithm.
# Confusion % # Confusion %
1. ɛ→ɪ 68 8. u→ʊ 15
2. ɔ→o 49 9. ʊ→o 15
3. ʌ→ɑ 39 10. o→ɑ 14
4. u→o 33 11. ɪ→ɛ 13
5. ʊ→u 28 12. ɑ→ʊ 13
6. ɑ→ʌ 24 13. ɑ→ɔ 13
7. ɔ→ɑ 18 14. o→ʊ 10
Only /e/, as a NEW AE vowel, did not take part in the confusions of the AE vowel production by PA learners, i.e., did not reach the 10% criterion. These results conform with my second hypothesis and consequently confirm the H2.1 sub-hypothesis concerning the mid vowels, and
38 By convention, ‘→’ means the left member of the pair was confused (incorrectly identified) as the right member;
in pair with ‘↔’ the confusion is reciprocal.
more particularly with the H2.2 sub-hypothesis on how learners may confuse the mid-back vowels. This also concurs with the perception results in Chapter 4.
The overall results on classifying the PA data in accord with AE native results show that similar sounds in the two languages posit less production difficulty for PA EFL learners in comparison with new vowels. Most of the difficulties were statistically found in the NEW vowels, which supports the L2LP model. The exception is the AE /u/ vowel as a near counterpart of PA /u:/ in this case. This finding is supported by SLM and is more likely to be separated by duration. This can be interpreted as learner-individual interlanguages that approximate the native ideal to different degrees.