In section 9.4.2, I concluded that it was not possible to satisfactorily model perceived soci- olinguistic variation. Only a low percentage of the variation found for Advancement could be accounted for systematically. A plausible explanation for this finding is that there was not enough systematic sociolinguistic variation present73.
It seems justified to assume that the acoustic data did not contain enough systematic sociolinguistic variation, despite the finding reported in section 8.5 that the sociolinguistic variation was judged reliably for a subset of three vowels (although this effect was relatively small; the mean value for Cochran’sαacross all three perceptual variables was 0.46 for /E/, 0.477 for /O/, 0.423 for /Y/, 0.423 for /a/, 0.299 for /A/, and 0.199 /u/ see Table 8.16). Given the results for these three vowels, it was concluded in Chapter 8 that the perceptual data of these three vowels contains systematic sociolinguistic variation and that the listeners were able to systematically judge sociolinguistic variation. It could be expected that it would be easier
73An alternative explanation for the results of sociolinguistic modeling could be that the behavior of the listeners
to fit the models for Height, Advancement, and Rounding using the acoustic representations for these three vowels than for the other six vowels. Nevertheless, in section 9.4.2, it was found that no complete model (i.e., for Height, Advancement, and Rounding) could be fit for any of the vowels. It must therefore be concluded that these three vowels did not contain enough variation that could be reliably judged and subsequently modeled using the acoustic representations; the values of Cochran’sαfor /E/, /O/, /Y/, /a/, /A/, and /u/ were too low.
Three additional factors can be put forward that may have attributed to the low reliabil- ity scores for the individual vowels, in addition to the aforementioned low sociolinguistic variability displayed in the vowel tokens. First, I concluded in Chapter 8 that the listeners appeared to be heavily influenced by the response vowel category. It seems plausible that the listeners used a vowel-specific reference point. It can be hypothesized that they judged a vowel token’s Height and Advancement by determining the vowel category for that vowel token, and subsequently by placing that vowel token in the (designated) area in the quadrilat- eral for that vowel category. They placed the judgment closer or less close to their reference point for the specific response vowel category. In addition to being influenced primarily by the response vowel category, the listeners apparently chose not to use the entire area possible for their Height and Advancement judgments. A possible explanation for this behavior is that they judged the differences in the vowel tokens relative to the entire range of pronunciation variation possible. This means that they judged the vowel tokens from speakers from the N-R region, while somehow keeping in mind the entire variation spectrum that is possible for Dutch. In other words, the listeners judged the relatively subtle differences as if they were going to be presented with more variation during the experiment. However, it is not possible to test this hypothesis using the results of the experiment presented in the present research, it would take another experiment in which the N-R data was also presented as vowel tokens that display more sociolinguistic variation. Such a judgment strategy may have caused lower variability in the judgments.
A second factor that may have contributed to the low reliability scores is that the listeners did not receive information about their previous judgments in the experiment. Perhaps the listeners were unsure about the exact location of their reference point (for each response vowel) across experimental trials. This insecurity possibly introduced noise in the judgments and caused lower scores for Cochran’sα74.
A third factor that may have affected the reliability scores is that the judgment task was
74It should be noted that the explanation for the low reliability scores refers exclusively to the judgments of
Height and Advancement. For Rounding, the judgment behavior can be argued to be different. First, when judging Rounding, the listeners seemed to be influenced less by the response vowel category than when judging Height or Advancement. This was found in Chapter 8, when possible differences between the standard deviations across the three experimental conditions were studied. It seemed that, when listeners were presented with a judgment area that did not allow them to reserve a specific area for a specific vowel (as was the case for Rounding), then their judgments display less phonemic variation. This hypothesis can explain why it was more difficult to fit models for Rounding; all models for Rounding showed a lower percentage of explained variance than those for Height and Advancement.
intensive and perceived as fatiguing by the majority of the listeners. This may have led to noise in the judgments which in turn may have decreased the reliability scores.
In sum, the results for the Height and Advancement judgments can be interpreted as that the listeners preferred to maximize phonemic variation, at the cost of sociolinguistic variation in their judgments. In addition, while it was concluded in Chapter 8 that the listeners did reliably judge the sociolinguistic variation, the reliability scores were too low for the perceived sociolinguistic variation to be modeled through linear regression analysis. The present results do not allow me to conclude which one of the normalization procedures is most suitable for representing sociolinguistic differences in agreement with phonetically- trained listeners.
9.6
Conclusions
In this chapter, a comparison was carried out between the acoustic data described in Chapter 7 and the articulatory perceptual data described in Chapter 8. The 12 normalization procedures that are evaluated in the present research were compared on how well they produced data that could be used to model phonemic and sociolinguistic variation perceived by phonetically- trained listeners. The comparison was carried out for phonemic variation and for sociolin- guistic variation. First, it was found that phonemic perceived articulatory differences could be modeled very effectively using (transformed) acoustic data. The vowel-extrinsic/for- mant-intrinsic procedures performed best, followed by the vowel-intrinsic/formant-intrinsic, vowel-extrinsic/formant-extrinsic, and the vowel-intrinsic/formant-extrinsic procedures, re- spectively. The success of the vowel-extrinsic/formant-intrinsic procedures was attributed to the fact that all three of these procedures make use of speaker-specific information, such as the mean across a speaker’s vowels, to re-scale the acoustic data. The poor performance of the vowel-intrinsic/formant-extrinsic and (two of the) vowel-extrinsic/formant- extrinsic procedures was attributed to the fact that these procedures use F3 to model their second dimensions. For the sociolinguistic variation, it was concluded that it was not possible to model perceived sociolinguistic variation in the perceptual data satisfactorily, because the stimulus vowel tokens did not reflect enough variation to be judged reliably enough by the listeners. Because no models could be fit, it was not possible to establish which one of the normalization procedures was most suitable for representing sociolinguistic differences in agreement with phonetically-trained listeners.
General discussion and
conclusions
10.1
Introduction
The research described in this thesis aimed to establish which procedure for acoustic vowel normalization meets the following criterion best. The procedure must preserve the phone- mic variation and the sociolinguistic speaker-related variation, while minimizing the ana- tomical/physiological speaker-related variation in the transformed acoustic representation of vowel tokens. As was argued in Chapter 1, this criterion must be met so that the normalization procedure is considered suitable for use in sociolinguistics.
The present chapter discusses the performance of the normalization procedures, given the results of three sets of comparisons. The first are the acoustic comparisons (Chapter 7), the second are the perceptual comparisons, which involved a comparison of the performance of phonetically-trained experts (Chapter 8), and the third are the perceptual-acoustic compar- isons of the acoustic normalization procedures to the experts’ judgments (Chapter 9). This chapter aims further to provide a discussion of the results and to present the conclusions of the present research.
This chapter is set up as follows. Section 10.2 evaluates how well each of the 12 proce- dures performed at the three types of comparisons and discusses the core research ques- tion. Section 10.3 sums up the implications of the findings of the present research for sociolinguistics, while section 10.4 discusses the implications of the present research for phonetics. Finally, in section 10.5, a short discussion is provided of the limitations of the present research, suggestion for further research are given, and some concluding remarks are made.