CAPÍTULO I: CONTEXTUALIZACIÓN DE LA EXPERIENCIA
2.1. Marco Contextual
2.1.6. Incursión de grupos armados
The results of the comparison of the normalization procedures, summarized in section 10.2.1, show that the three vowel-extrinsic/formant-intrinsic procedures were most suitable for use in sociolinguistics. My results suggest that sociolinguistic differences in the pronunciation of vowels between groups of speakers can be investigated acoustically as follows.
First, the fundamental frequency and the frequencies of the first three formants should
75Although the vowel-intrinsic/formant-intrinsic procedures performed slightly better in the perceptual-acoustic
be measured. Although it was found in Chapter 9 that the models for the perceived artic- ulatory characteristics could be fit best using onlyF1 andF2, and that F0 andF3 did not improve the fit of the models, I do not recommend to excludeF0 and/orF3 a priori76. In tone languages or dialects such as, Mandarin Chinese or for Dutch dialects including tonal differences (e.g. some dialects spoken in the S-N region in the present research),F0can have a contrastive function for tones. F3helps listeners to distinguish between certain classes of front unrounded vowels in languages such as Swedish (cf. Fujimura, 1967) and for American English,F3is necessary to classify rhotacized vowels (Ladefoged, 2001).
Second, the measurements in HZ should be transformed to z-scores using LOBANOV. Nevertheless, it seems advisable to compare the mean values per vowel inLOBANOVwith the mean values in Hz, because the possibility that LOBANOV introduces results that are inherent consequences of the transformation itself cannot be excluded given the results found in Chapter 7. Any differences in the mean values found in the data transformed following LOBANOVand the mean values forHZshould be investigated further, because they may be indications of language variation.
I would like to emphasize that Disner’s (1980) recommendations are still valid; it is not advisable to carry out normalization procedures on data sets that are not fully phonologically comparable. Disner found that each procedure makes ‘implicit assumptions’ (i.e., the mean values and the standard deviations of both speaker groups have to be comparable) about the underlying vowel system when languages or dialects are compared on the basis of the normalized vowel frequencies. When the phonological vowel systems differ, the scale factors used in the vowel-extrinsic/formant-intrinsic procedure become biased. This may result in artificial differences between the data sets to be compared, or in the elimination of relevant differences.
One of the disadvantages of the three vowel-extrinsic/formant-intrinsic procedures is that they require information across all (monophthongal) vowels of a single speaker to estimate the scale factors (e.g., the mean formant frequency across all vowels for LOBANOV) that are used to normalize the raw measurements. Some sociolinguistic studies do not include all vowels from each informant. Nevertheless, the results in Chapter 7 indicate that it may suffice to obtain recordings and measurements of the three point vowels per speaker for the vowel system in question. Using only these three vowels per speaker, the scale factors can be estimated satisfactorily. I found that measurements obtained from scale factors estimated with the three point vowels per speaker correlate strongly with measurements obtained with scale factors estimated using all nine vowels per speaker. It should be noted that, because the correlations were not 100%, and because the results of the linear discriminant analysis showed a deterioration for the three-vowel case, measurements transformed using three vow- els per speaker are of lower quality than the measurements that were transformed using all vowels for a speaker77.
76It is advisable to also measure duration, for instance for distinguishing contrastive length differences. 77Although it should be noted that the measurements normalized using scale factors estimated using only three
In Chapter 9, it was concluded that it was not possible to model the sociolinguistic variation in the perceptual data. This means that it is not clear to what extent (transformed) formant frequencies must differ to be perceived by phonetically-trained listeners as a sociolinguistic difference. In addition, in Chapter 7 it was established that applying vowel-extrinsic/formant- intrinsic procedures to data inHZmay lead to differences between mean values that were not present in the data inHZ. However, some of these (additional) differences between the mean formant frequencies of the normalized data and the raw data can be excluded beforehand. This can be accomplished using the just noticeable difference for these formant frequencies. Kewley-Port & Watson (1994) measured (language-independent) discrimination difference limens for pairs of formant frequencies for isolated synthetic vowels (simulating a female voice). They stated that for the F1 region a constant difference of 14.9 Hz between two formant frequencies is just noticeable. For the F2 region, a linear difference of 1.5% is necessary to be noticeable. Kewley-Port & Watson suggested a ‘piecewise-linear’ function in which theF1region is defined as<800 Hz and theF2region is defined as>800 Hz. However, Kewley-Port & Watson’s formula is discontinuous in the region between 800 and 1000 Hz. I therefore suggest that theF1 region is set to<1000 Hz and theF2region is set to>1000 Hz, to avoid formant means with values between 800 and 1000 Hz showing difference limens that are too low.
By applying the modified version of the formula proposed by Kewley-Port & Watson, the differences between mean formant frequencies smaller than the difference limen can a priori be excluded as possible sociolinguistic differences, for the raw mean values in HZ. Before subjecting the differences between mean values to further analysis, it must be verified that these differences are larger than the difference limen. Differences that are smaller than the difference limen cannot reliably be perceived and are therefore not likely to represent a sociolinguistic difference. Note that this procedure is only to be applied to mean formant fre- quencies, and not to formant frequencies from two single vowel tokens. This issue deserves to be investigated further78.