LENITION AND THE ASYMMETRY IN MUTUAL INTELLIGIBILITY OF DANISH AND SWEDISH
Master´s Thesis of Research Master Linguistics Rijksuniversiteit Groningen
Student:
Sandrien van Ommen
Table of contents Acknowledgments 1 0. Abstract 3 1. Background 5 1.1 Mutual intelligibility 5 1.2 Lenition 8
1.3 Danish intervocalic lenition 11
2. Questions and Hypotheses 17
2.1 Main research question 17
2.2 Production 17 2.3 Perception 18 3. Production 19 3.1 Introduction 19 3.2 Method 19 3.2.1 Material 19 3.2.2 Task 21 3.2.3 Acoustic correlates 22 3.3 Hypotheses 26 3.3.1 Questions 27
3.3.2 Hypotheses regarding prevocalic position 27
3.3.3 Hypotheses regarding intervocalic position 28
3.4 Results 28
3.4.1 Prevocalic position 28
3.4.2 Intervocalic position 30
3.5 Discussion phonetic analysis 31
4.2.1 Subjects 41 4.2.2 Material 42 4.2.3 Task 43 4.3 Results 44 4.4 Discussion experiment 51 4.5 Classification tree 52 4.6 Discussion perception 54 5. General Discussion 59 5.1 Research questions 59 5.2 Production 60 5.3 Perception 60 5.4 Main findings 62
6. Conclusion & Recommendations 65
7. References 67
1 Acknowledgments
Writing this thesis would not have been possible without help. I would like to thank my supervisors Dicky Gilbers, Charlotte Gooskens and Petra Hendriks for their help during the whole process. Furthermore, the research group of the mutual intelligibility project at the University of Groningen for their help during our many meetings: Charlotte Gooskens, Anja Schüppert, Renée van Bezooijen, Erik Vedder, Jan Vanhove and special thanks to Vincent van Heuven.
Finding participants for the perception experiment has been very difficult, and I cannot begin to thank everyone who contacted Danish and Swedish friends. I would like to thank my friends (Marten, Kirsten, Sabine and everyone else), family and family of friends for their help and support and, finally, Yvonne Dijkstra, Christina Däldborg and Linda Dahlen without whom I would not have had
3 0. Abstract
5 1. Background
1.1 Mutual Intelligibility
Danish, Swedish and Norwegian, the three mainland Scandinavian languages, are very similar. So similar even, that speakers can communicate with each other using their own language while understanding the other. Haugen (1966) calls this semicommunication, but this suggests
communication is unsuccessful, or only halfway. The term receptive multilingualism (Braunmüller, 2002) may be more appropriate.
The fact that these languages are very similar and mutually intelligible does not mean speakers understand each other to an equal extent. In previous studies (Maurud, 1976, B, 1978, Delsing & Lundin Åkesson, 2005, Gooskens, 2007) it has consistently been found that mutual intelligibility between mainland Scandinavian languages is asymmetrical: for example, Swedish is easier to understand for Danes than Danish is for Swedes. In fig. 1 the average scores in spoken intelligibility tasks of four studies are given. In Delsing & Lundin Åkesson (2005) the spoken intelligibility task consisted of spoken text which the participants had to answer five questions about. In Maurud (1976), too, participants were asked questions about spoken text, and besides that, they had to do an auditive translation task with words, embedded in sentences.
6 As the graph shows, Norwegian people score high on spoken language intelligibility of Danish and Swedish. Furthermore, Swedish is consistently understood better by Danish and Norwegian participants than Danish and Norwegian are understood by Swedish participants. Danish and Swedish people have most trouble understanding each other. In the current study this latter asymmetry, i.e. the asymmetrical mutual intelligibility of Danish and Swedish, is investigated. This asymmetry is particularly interesting because the mutual intelligibility is low and not facilitated by either lexical similarities (as in the Danish-Norwegian relation) or phonetic similarities (as in the Swedish-Norwegian relation) (Gooskens, 2007).
In intelligibility research, but also in second language acquisition research, the wish to quantify differences between languages is strong (see e.g. Chiswick & Miller, 2004). This so-called linguistic distance has had several forms. Chiswick & Miller (2004) for example, uses an index of proficiency, rendering the linguistic distance by inversion of linguistic score (1/linguistic score). However, this is circular reasoning when investigating the effect of linguistic distance on intelligibility or language acquisition. Another way is to measure linguistic distance by the relatedness of languages, using ‘family trees’ (Crystal, 1987). When investigating the influence of linguistic distance on mutual intelligibility between closely related languages, however, this method is not very useful either. Quantifying the various complex characteristics of languages in linguistic distances is laborious, but in previous studies on Scandinavian languages and varieties of Dutch (Gooskens 2007) some of these distances have been proven to correlate with mutual intelligibility. As Gooskens (2007) states, morphological and syntactic differences between the Scandinavian languages can be assumed to be of hardly any importance for mutual intelligibility, but lexical and phonetic distances, on the other hand, show a strong correlation with scores on spoken mutual intelligibility tasks. The lexical distance was in this case computed by counting the number of cognates1
1)
. The higher the percentage of non-cognates, the larger the lexical distance. The phonetic distance between closely related languages also makes use of cognates, but in this case the difference between two pronunciations of a cognate is computed, by use of the Levenshtein distance (see a.o. Heeringa, 2004, Nerbonne & Heeringa, 2009). Levenshtein distance is a string distance measure, computing the amount of phonetic differences between two words. In (1) I give a simple Levenshtein distance computation of English
star and Dutch ster (‘star’).
0 0 1 1 Levenshtein distance: 2
The phonetic distance between languages, then, is the mean of Levenshtein distances as computed over a substantial amount of cognates. The larger the amount, the more solid the result. This distance is used in dialectology, but also in intelligibility research (a.o. Heeringa, 2004, Gooskens, 2006) .
Linguistic distances are useful to quantify differences between languages, but an intrinsic feature of distance makes it an insufficient measure for the current study: distance is symmetric. Since mutual intelligibility in Scandinavian languages has been found to be asymmetric and as this asymmetry is
7 the subject of the current study, expressing differences between languages in distances will not suffice.
To account for the asymmetric mutual intelligibility in mainland Scandinavian languages some extra-linguistic factors have been investigated. It has been suggested an important extra-extra-linguistic factor is language contact. Language contact increases by increased geographical proximity, prevalence of a language in radio and television and the popularity of a country for touristic purposes. Maurud (1976) found that for his Danish, Swedish and Norwegian participants Sweden was the most frequently visited country, whereas Norwegians were the most widely travelled people. As Maurud (1976) states, previous surveys, e.g. by Haugen in 1952-53, show that Swedish radio and television is broadcast and watched more in Denmark and Norway than the other way around. Maurud replicates these findings, reporting that Danes and Norwegians watch Swedish television and listen to Swedish radio, whereas Swedes almost never listen to Danish or Norwegian radio or television. Furthermore, Swedish periodicals, magazines and newspapers are read by Danes and Norwegians, whereas Swedes read relatively little Danish and Norwegian. Again, Norwegians tend to read more in their
neighboring languages than Norwegian is being read by Danes or Swedes. These findings concur with the data as given in fig. 1.1: Swedish is understood relatively well, whereas Swedes have less
understanding of Danish and Norwegian. Norwegians, again, score high on intelligibility tasks of Danish and Swedish. Besides contact Maurud surveyed attitudinal factors. Swedes show little confidence in their ability to understand neighboring languages, while at the same time preferring their own language over other languages. This strongly suggests an important role for contact and attitude in mutual intelligibility. In the investigation of Maurud, however, the results on the surveys pertaining to contact may have been influenced by the fact that all subjects came from areas around the capitals in the three countries. This means the Danish and Norwegian subjects lived closer to the Swedish border than the Swedish subjects did to the Danish and Norwegian border.
In later studies (Delsing & Lundin Åkesson, 2005) this ‘geographical’ factor2
As Gooskens (2007) states, differences between Danish and Norwegian are mainly lexical, differences between Norwegian and Swedish are mainly phonetic, but differences between Danish and Swedish are both lexical and phonetic. Since phonetic distance in previous studies (Gooskens, 2006,
Gooskens, 2007, Gooskens et al., 2009) has been proven to correlate strongly with mutual
has been controlled for, showing the (young) subjects in general have limited contact with the neighboring languages, although Swedish still dominates. Gooskens (2006) investigated the factors outlined above, contact and attitude, on their role in the asymmetrical mutual intelligibility, but only weak correlations were found. Furthermore, it has been hard to prove a direct relationship. The only significant relationships found in Gooskens (2006) were phonetic distance (r=-.82, p=.00) and the ‘beautiful’ scale (r=.56, p=.02), meaning subjects were asked to judge the appeal of a language, which score was then correlated with their understanding of that language. As correlations are not directional, it cannot be said whether speakers make more effort to understand a language they find beautiful, or whether they find a language more beautiful when they understand it better. So even though both phonetic distance and attitude correlate significantly with intelligibility, the role of attitude (in this case beauty) in intelligibility is ambiguous. With this, the asymmetrical relationship in mutual intelligibility has not been explained yet.
8 intelligibility and this correlation is stronger than the correlation with lexical distance, it may be worth looking into the phonetic relation more closely. Phonetics clearly play a role in mutual
intelligibility, but phonetic distance is unable to explain the asymmetries found. I will now turn to the possibility that phonetic differences in production may not necessarily impede non-native perception symmetrically.
Several phonetic differences between Danish and Swedish may be impeding Swedish listeners more than Danish listeners, but most of these differences point to one main phenomenon: phonetic reduction. Danish, for example, has been proven to have a higher articulation rate than Swedish (Schüppert, Hilton, Gooskens, submitted). A higher articulation rate an sich could already cause an asymmetric mutual intelligibility relation, but there is reason to believe this result is a measure of phonetic reduction, or lenition. The study found a higher articulation rate for the amount of uttered words and assumed underlying syllables (phonological) per second, but a lower articulation rate for amount of phonetically realized syllables per second. This result suggests a higher amount of reduction processes in colloquial Danish. In Danish, namely, reduction and lenition in some cases causes apocope (elision or reduction of the last syllable), resulting in a lower amount of realized syllables than what would be expected based on underlying forms.
As state above, the written form of Danish words is not always a direct representation of the spoken form. When in a language the phonetic realization of words is not easily derivable from their
orthography this language is considered to have a deep orthography (as opposed to a shallow
orthography). Comparing Danish to Swedish, Danish is considered to have a deeper orthography than Swedish (Elbro, 2006). This aspect may facilitate the intelligibility of Swedish for Danes in cognates. As written language is more conservative than spoken language, and spoken Danish has diverged further from its North Germanic roots (Elbro, 2006), Danish orthography is closer to Swedish spoken language than Swedish orthography is to Danish spoken language. Hypothetically, this would make it possible for Danes to benefit from orthography when mapping Swedish phonology to Danish.
Both the articulation rate and deep orthography of Danish indicate a role for the type of phonological change Danish has undergone, diverging from Swedish, in the mutual intelligibility relations in
mainland Scandinavia. This factor is the phonetic reduction in Danish, leading to elision of syllables and a phonetic divergence from orthography. This phonetic reduction has other effects as well, such as consonantal lenition, which I will discuss in the next section.
1.2 Lenition
Lenition is a term used for many phonological processes. It is a common phonological change, often associated with articulatory strength. In fact, there is no consensus on what constitutes lenition, but, as Harris (1990) points out:
“( ..) it is fair to say that there exists a core of phonological process types which would informally be classed as consonantal lenition by the majority of phonologists.” (Harris, 1990)
9 speakers can choose to articulate relatively poorly (hypoarticulation), because the overall message they are trying to convey is lexically or contextually predictable (see Lindblom, 1990). In some cases, articulating poorly leads to a loss of sounds and even whole syllables (see Ernestus, 2000) such as the pronunciation of we have as we’ve. In less obvious cases, articulating poorly leads to a change of sounds, making them less easy to identify. For example: vowels are centralized, sometimes resulting in schwa. Ernestus (2000) for example, reports that even the closed last syllable in Dutch bijvoorbeeld (‘for instance’) can be reduced from bijvoorbijvoorThe apocope in Danish mentioned in section 1.1 is another example of reduction.
Besides the fact that words or parts of words may be contextually predictable, lenition also has to do with the fact that features of sounds are less audible in one context than in the other. For example, some features of stops are harder to perceive word-finally than word-initially, because much
information on sounds can be taken from the ‘borders’ of sounds, i.e. the transition of one phoneme to the other. As the word-final stop is not followed by any sound, the information which could otherwise be taken from the following vowel (formant transition values), to identify the stop, is now lost. Some of the most notable accounts of lenition combine this perceptual distinctiveness of sounds (an acoustic measure, c.f. Steriade, 1997) with the lack of articulatory effort of the speaker
(Flemming, 1995, Kirchner, 1998, Boersma, 1998 and 2007). This means speakers use information on production and perception simultaneously. The consensus on lenition can thus be summarized as a consonantal ‘weakening’, connected to acoustic distinctiveness and articulatory effort.
In many cases, weakening is defined as a loss of consonantal strength, based on the sonority hierarchy (c.f. Selkirk, 1984):
stops << affricates << fricatives << nasals << liquids << vowels
Lenition, in this view, makes a consonant more vowel-like. A consonant, then, is considered to be more vowel-like when constriction is weaker (e.g. affricates vs. stops) or when it has periodicity (voiced vs. voiceless). Some examples of lenition are then (the list is not exhaustive):
Table 1.1 Examples of lenition
Type of lenition Phonetic implications Example
spirantization stop -> fricative (or approximant) Sardinian [baka] (‘cow’)– [saaka] (‘the cow’)
debuccalization loss of constriction Italian buco (‘hole’) - Tuscan
buco [bu:ho] (‘hole’)
sonorization voiceless -> voiced Latin amicus – Spanish amigo (‘friend’) degemination geminate -> singleton Finnish rottaa (partitive) – rotan
(genitive) (‘rat’) …
10 phenomenon, such as the example of Latin amicus changing into Spanish amigo. For the current study this distinction is not very relevant as I will not compare diachronic with synchronic phenomena and the distinction, in reality, is a gradual process.
Another way to view lenition is as a kind of phonetic reduction leading to a loss of segmental complexity (Harris 1990). This kind of view works in phonological theories in which segments are built up from univalent phonological atoms, as opposed to binary features (e.g. [-voice] vs. [+voice]). Simply put, lenition in this view means that elements for which a segment is specified are lost from the internal representation of a segment. The difference between both views (i.e. consonantal weakening vs. loss of segmental complexity) may seem a theoretic issue, but in some cases a loss of segmental complexity constitutes an increased consonantal ‘strength’. An example of this is final devoicing in Dutch (which is also a general phenomenon in Turkish, Russian, Polish and many other languages). As final devoicing is ‘hardening’, because voiced (or lenis) sounds become voiceless (fortis), it is seen as a fortition. At the same time, in e.g. Dutch (a voicing language), as a.o. Jansen (2004) and Keating (1984) elaborate, the lenis stop is considered to be actively voiced, whereas in aspirating languages the fortis stop is actively devoiced. A stop with short-lag Voice Onset Time3
In both views, lenition is seen as a kind of weakening, either of consonantal strength or of
phonological contrast. Both these definitions may apply to the Danish case. Danish and Swedish are
considered to be aspirating languages,
(VOT) (lenis in aspirating languages and fortis in voicing languages), then, is the neutral stop. Lenition is the change of a segment into a more neutral form (a feature is lost), meaning that in aspirating languages lenition is voicing, but in voicing languages devoicing may, in this view, well be lenition (Harris, 1990).
4
In Danish, more intense than in the other North Germanic languages, a prosodic change called
Akzentballung (Bandle, 1973, cited in Schrambke, 2007) has occurred. This is the reinforcement of
primary stress on the stem syllable, which has lead to a.o. apocope and some forms of consonantal lenition (as the examples in table 1.1). Schrambke (2007) gives the following results of this
concentration of stress in Danish:
as most Germanic languages (Dutch, Yiddish and Afrikaans excepted) (Iverson & Salmons, 1995). This means that both in Danish and Swedish, the laryngeal contrast is one of short-lag VOT vs. aspiration (long-lag VOT). Reverting to the discussion above, the neutral form in Danish and Swedish is the lenis form (as the fortis form is actively devoiced), hence lenition would in these languages be a weakening, both when it is considered a weakening of consonantal strength and when it is considered a weakening of phonological contrast.
1. The regressive assimilation of tn > nn, e.g. vatn ‘water’ > danish vand []. The assimilation of tl > ll;
2. The development of fricativized g > semi-vowels: w/u after velar vowels, > j/i after palatals; in some parts, g was deleted completely;
3. The lenition of p t k > b d g (voiced or unvoiced); in some areas further fricativization to w/u,
ð, or complete deletion occurred, especially of t;
3 A short-lag VOT means the vocal chords start vibrating at the point at which the constriction of a stop is released. In chapter 3 I will elaborate on the phonetics of the production of stops.
11 4. The weakening and in some parts deletion of the final vowel (apocope).
I previously mentioned the reduction of the unstressed syllable as a reason for the higher articulation rate as found by Schüppert et al. (submitted) and as a factor in the orthography-phonetics
discrepancies in Danish. This reduction is a kind of phonological change which may be an explanation for the asymmetrical mutual intelligibility relations, as it may cause phonological confusion.
Anecdotally, in Swedish perception the Danish speak ‘with their mouth full’. In the next section I will discuss the lenition of p t k > b d g (point (3) Schrambke (2007) mentions), which is the subject of this thesis. I will discuss why it may be part of the explanation for the asymmetrical mutual intelligibility relation between Danish and Swedish and how I investigate whether this is the case.
1.3 Danish intervocalic lenition
The current thesis focuses on the consonantal lenition in Danish intervocalic stops, a type of
degemination and sonorization. Danish and Swedish both used to have intervocalic geminate stops
(Schrambke, 2007). In Swedish these geminates are preserved, but in Danish the stops have degeminated. In this process, the voiceless stops have sonorized, leading to simple voiced stops in intervocalic position.
Table 1.2. The (original) lenis-fortis contrast in Danish and Swedish intervocalic (geminate) stops in non-words.
LENIS FORTIS
orthography Swedish Danish orthography Swedish Danish
abbe appe
adde atte agge acke / akke
This particular lenition phenomenon does not only lead to loss of consonantal strength, but more importantly: loss of contrast. In Danish, namely, the intervocalic stops have become ambiguous, whereas Swedish still has contrastive stops in intervocalic position. This may contribute to the asymmetrical intelligibility relation between Danish and Swedish, as fig. 1.2 shows.
Danish Swedish
12 One reason to hypothesize an advantage for Danish listeners in this case is a mathematical one. A richer system leads to higher uncertainty when mapping a lesser amount of foreign sounds to native sounds. In the example above this would mean mapping Swedish sounds to Danish sounds is easier than the other way around. Moberg et al. (2006) use conditional entropy to investigate this idea computationally. Conditional entropy is a measure of uncertainty, based on the probability that a phoneme is mapped to the right phoneme in another language. A high entropy thus means a higher uncertainty, hypothetically decreasing intelligibility. The formula for conditional entropy is given below (as given in Moberg et al., 2006).
H is the entropy of one language (Y) given the other (X). The probability (p) that a foreign sound (y) is mapped to the right native sound (x) is transformed to entropy. The non-words in table 1.2 may help to explain this formula. The non-words appe and abbe are pronounced
[and,respectively, in Swedish. In Danish both are pronounced Swedish [, then, is invariably mapped to Danish [(p=1), Swedish [ is invariably mapped to Danish [ (p=1), Swedish [] is invariably mapped to Danish [ (p=1) and Swedish [:] is invariably mapped to Danish [as well (p=1). As all Swedish sounds are uniquely mapped to Danish sounds, the entropy for Danish listeners is –log21=0 for all sounds, rendering a total of H(Danish|Swedish)=0. For Swedish listeners, on the other hand, Danish [is mapped to [] half of the time, (for simplicity I assume equal frequencies for both sounds in this example) and to [:] the other half of the time, rendering the chance of a correct mapping in these cases p=0.5. So, out of 6 phoneme-mappings, there are two contributions of p=0.5 to entropy. Since –log2(0,5)= 1, this means H(Swedish|Danish)=1/6 + 1/6= 0.333. Thus, there is a higher entropy for Swedish listeners than for Danish listeners, meaning a higher uncertainty.
Facilitating the perception task for Danish listeners even more, as can be seen in table 1.2, and as has been mentioned in section 1.1 and 1.2, Danish orthography still reflects the non-lenited
pronunciation (as in Swedish). As prevocalic stops are still contrastive in Danish, it may be possible for Danes to map the Swedish pronunciation to the right orthography through knowledge of the prevocalic contrast.
13 of an effect of slight variation in phonetic realization of sounds on the perception of these sounds is not taken into account, since the computations are based on phonetic transcriptions. Best (1995) takes a direct realist approach to cross-language speech perception. In this view the notion of the ‘phoneme’ is an abstraction of the language-learner’s articulatory experience, or in the words of Best (1995):
“[ …] once a highly attuned perceiver has discovered the more optimal higher-order invariants, the more numerous degrees of freedom describing the lower-order details will generally be passed over as being irrelevant in and of themselves with respect to his or her increasingly abstract goals.”
Best (1995), here, refers to the ability of perceivers to generalize over the phonetic variation within a language to higher-order invariants, constituting the native phonemes. The Perceptual Assimilation Model (PAM) Best (1995) proposes, then, makes predictions for the categorization of non-native phonemes. She proposes non-native sounds (differing in phonetic realization from native sounds) (1) assimilate to a native category, (2) are perceived as an uncategorizable speech sound or (3) are not perceived as speech. These patterns differ according to the extent to which a sound assimilates native categories. With this model, then, articulatory differences between languages may impede non-native categorization.
From Pierrehumbert (2001), similar predictions could be made on an acoustic level. Exemplar dynamics, outlined a.o. in Pierrehumbert (2001) and (2003), is based on the idea that the acoustic identity of a phoneme is construed by an interplay of production and perception: all occurrences of a sound are stored and progressively either granulated (i.e. remembered, becoming a good exemplar) or forgotten. The sounds the speaker produces are based on these granulated ideals, but influenced by articulatory mechanisms. This produced item, again, is stored, and so on. Pierrehumbert (2001):
“The exemplars encoding frequent recent experiences have higher resting activation levels than exemplars encoding infrequent and temporally remote experiences.”
Classification of a non-native sound, then, is a positioning of this sound in the parameter space, judging the similarity of this sound to the existing exemplar ‘clouds’. When a sound is sufficiently similar to existing exemplars, classification is as in (1) of Best (1995): assimilation to a native category. If a sound is very dissimilar a new category may have to be developed (2), or the sound is not perceived as a speech sound (3). An important difference between the direct realist approach, as described in Best (1995) and exemplar dynamics as outlined in Pierrehumbert (2001) is the role of articulatory gestures and acoustics, respectively.
For the current study, I phonetically analyze produced sounds on an acoustic level, after which I link the classification of these sounds by non-native perceivers (Swedish participants perceive Danish and Danish participants perceive Swedish) to the results of the production study. The task for the
participants is the classification of two members of a contrast. Best (1995) makes detailed
14 1) Two-category assimilation type (TC): Each non-native segment is assimilated to a different
native category, and discrimination is expected to be excellent
2) Category-goodness difference (CG): Both non-native sounds are assimilated to the same native category, but they differ in discrepancy from native ‘ideal’ (e.g. one is acceptable, the other deviant). Discrimination is expected to be moderate to very good, depending on the magnitude of difference in category goodness for each of the non-native sounds
3) Single-category assimilation (SC): Both non-native sounds are assimilated to the same native category, but are equally discrepant from the native ‘ideal’; that is, both are equally
acceptable or both equally deviant. Discrimination is expected to be poor (although it may be somewhat above chance level).
4) Both uncategorizable (UU): Both non-native sounds fall within phonetic space but outside of any particular native category, and can vary in their discriminability as uncategorizable speech sounds. Discrimination is expected to range from poor to very good, depending upon their proximity to each other and to native categories within native phonological space. 5) Uncategorized versus categorized (UC): One non-native sound assimilated to a native
category, the other falls in phonetic space, outside native categories. Discrimination is expected to be very good.
6) Nonassimilable (NA): Both non-native categories fall outside of the speech domain being heard as nonspeech sounds, and the pair can vary in their discriminability as nonspeech sounds, discrimination is expected to be good to very good.
These predictions allow for an asymmetrical intelligibility relation, because the category goodness of non-native sounds is not necessarily the same for both languages. I will illustrate this in fig. 1.3, which is a schematic representation of two languages and two categories, for which the predictions differ when reasoning from one language than from the other.
b … language 1 _ language 2 a A B ^factor 1 >factor 2
Fig. 1.3 Schematic representation of two categories in two languages.
With the example of production data in fig. 1.3 (two fictitious languages with two fictitious categories on two fictitious factors), predictions can be made on the non-native perception of sounds.
15 to one native category (a). Category A contains better exemplars of category a than does B, pointing to a category goodness difference (CG), which is prediction (2): both non-native categories are assimilated to one native category. Judging from language 2, though, non-native category a is most similar to native category A, and non-native category b is discrepant from any native category. This leads to prediction (5) (UC): one non-native category is assimilated to a native category, the other falls inside phonetic space, but outside of native categories. Prediction 5 expects categorization to be very good, so language 2 would be very able to discriminate non-native categories a and b, but for language 1 discrimination may not be good. First of all, both categories are possibly assigned to the same native category, suggesting there is no discrimination whatsoever. Second of all, it is plausible that both non-native categories are sufficiently similar to each other and the native category that they are not discriminated on any phonetic dissimilarities either.
With the production study (chapter 3) the similarities and dissimilarities between Danish and
17 2. Questions and Hypotheses
2.1 Main research question
This thesis investigates the role of lenition in the asymmetrical mutual intelligibility relation of Danish and Swedish. The lenition of Danish intervocalic stops is hypothesized to impede Swedish
categorization of Danish sounds, whereas Danish listeners are expected to be able to categorize Swedish stops more accurately. The main research question is:
Is Danish intervocalic lenition a factor in the asymmetrical intelligibility found between Danes and Swedes?
As discussed in chapter 1, there are reasons to expect lenition to have a role in the asymmetrical intelligibility relation between Danish and Swedish. First of all, on a phonemic level, the entropy (insecurity measure) would in this particular case (intervocalic stops) be higher for Swedish listeners, meaning they would have more trouble mapping Danish intervocalic stops to Swedish intervocalic stops than Danish listeners would have mapping the corresponding Swedish stops to Danish stops. Second of all, Danish listeners may have an advantage in this mapping procedure as their
orthography reflects the (phonemic) realization of Swedish stops, while the Swedish orthography does not reflect the Danish pronunciation. In this thesis the expected asymmetry is tested from a phonetic point of view. Simply put: in production Danish intervocalic stops are expected not to be contrastive on any phonetic correlate, making them difficult to categorize in perception, but Swedish stops may be more easily categorized by Danish listeners, because the latter have knowledge of the prevocalic stop contrast in their native language.
2.2 Production
The research consists of two parts: production and perception. With the results of the production part (a phonetic analysis), I will make predictions on the cross-language perception (a classification task) of Danish and Swedish stops. The results will show to what extent cross-language perception is influenced by phonetic detail and whether the asymmetrical intelligibility relation between Danish and Swedish can be explained from this phonetic angle.
I will start the phonetic analysis by investigating what factors are important in the Danish and
Swedish fortis-lenis contrast. Which phonetic correlates contribute to the contrast and how does this differ between the languages? The research questions for production are given below.
- How is the fortis-lenis contrast phonetically expressed in pre- and intervocalic stops in
Danish and Swedish?
- Does Danish undergo complete or partial phonological lenition of stops in intervocalic
position?
18 comparable to relevant correlates in the Swedish contrast, the Swedish categories may still be
(correctly) assimilated to different Danish categories. More detailed hypotheses will follow in chapter 3, after I presented the materials and the phonetic correlates I measured.
The reason why not only intervocalic, but also prevocalic stops are measured, has to do with the way the native abstract goal is constructed. I am referring to Best (1995) and Pierrehumbert (2001), in which it is theoreticized that lower-order variants, i.e. the variably expressed phonetic exemplars, together make up an abstract ideal of a sound (traditionally termed a phoneme). This abstract ideal can be seen as a label over a phonetic map, and it is thinkable that the highest-order label comprises characteristics of all occurrences of this sound, not only in intervocalic, but also prevocalic, word final or any other position in the word. Since in Danish a stop contrast still exists in prevocalic position, this may be an advantage for Danish listeners in cross-language perception, as they will have
contrastive categories to link non-native sounds to. Swedish listeners, on the other hand, will need to categorize an unfamiliar ambiguous sound (the Danish intervocalic stop).
2.3 Perception
The results of the production experiment will clarify the phonetic differences and similarities between Danish and Swedish stop-production, after which I will make predictions on the cross-language perception of these sounds. The questions to be answered are given below
- Can cross-language perception of Danish and Swedish pre- and intervocalic stops be
predicted on the basis of phonetic correlates only?
19 3. Production
3.1 Introduction
I will investigate the role of phonetic detail on mutual intelligibility. The phonetic detail to which the research pertains is the difference between fortis and lenis stops. The contrast between fortis and lenis plosives is not necessarily expressed in the same way between languages. A well-known difference found between languages is the difference in VOT, where some languages have a voicing contrast (e.g. Spanish, Dutch) and others an aspiration contrast (e.g. English, German). In a voicing contrast the fortis sound is voiceless (zero VOT, tenuis) and the lenis sound is prevoiced (e.g. Dutch [p] vs. [b]), in an aspiration language the fortis sound is aspirated and the lenis sound is not (e.g. English [p] vs. [b], where [b] is tenuis). There are also languages with a three-way contrast (e.g. Korean and Thai) and even a four-way contrast (e.g. Hindi and Marathi) (Lisker & Abramson, 1964). A three-way contrast typically has a prevoiced stop, a tenuis stop and an aspirated stop. In a four-way contrast the fourth member can be e.g. breathy-voiced (e.g. Hindi) or ejective (e.g. Lezgian, see Yu, 2004). VOT is not the only factor in the fortis-lenis contrast. Other articulatory and acoustic
differences, i.e. the ensemble of correlates of the ‘force of articulation’ (see a.o. Lisker & Abramson (1964) and Debrock (1977) for an overview), ensure that the contrast can still be perceived in, for example, whispered speech. The force of articulation is a term referring to a greater articulatory effort and build-up of air pressure due to a greater constriction at the place of articulation. According to Malécot (1970) sonority (voicing) is a redundancy of the force of articulation, meaning that
without sonority (i.e. in whispered speech) fortis and lenis sounds are still contrastive. Debrock (1977) though, states that, even though “the hypothesis is acceptable that a lenis, which is normally voiced, can lose this sonority while remaining lenis, it does not appear from known studies that a fortis, which is normally voiceless, can lose this voiceless character while remaining fortis, nor, and more importantly, does the opposite seem to be possible: nowhere is there any question of a voiceless consonant that could be produced or perceived as lenis, nor of a voiced one that could be a fortis.” Which characteristics of the opposition are redundant could plausibly differ between
languages, depending on which characteristics are grammatical (Cho & Ladefoged, 1999), higher order invariants, and which are lower order details (Best, 1995).
In this chapter I will describe the method used to phonetically analyze the fortis-lenis contrast in Danish and Swedish. First the material will be described, after which the correlates to be measured will be given and the phonetic analysis will be explained. The phonetic analysis consists of a statistical comparison of the data and a classification analysis. Finally I will discuss the results and formulate hypotheses for the perception experiment, which will be described in chapter four.
The research questions I search to answer with the production experiment are:
- How is the fortis-lenis contrasts phonetically expressed in pre- and intervocalic stops in
Danish and Swedish?
- Does Danish undergo complete phonological lenition of stops in intervocalic position?
20 The object of production experiment is the intervocalic (originally geminate) plosive in Swedish and Danish (Schrambke, 2007). A list of minimal pairs is made, given in table 3.1. The list compiled for the phonetic analysis and the perception experiment (chapter 4) had to meet two criteria:
- The list has to consist of words which are minimal pairs differing in the medial (geminate) plosive (/k:/,/g:/,/p:/,/b:/,/t:/,/d:/) only;
- the members of each minimal pair have to be existing words in both languages.
Although Danish and Swedish are very similar languages, there are not many exact cognates with this structure. A list of only six words (three minimal pairs in each language) could be made. To have more data for phonetic analysis, six more words were chosen in Swedish and six in Danish. All of these words, however, are minimal pairs in one language, but not across both languages. These additional words will be used in the acoustic analysis but not in the perception experiment, since they do not meet the strictest criteria. This will be further explained in chapter 4.
Next, six non-words were chosen, which are the same in both languages. These non-words contain all six geminates (fortis and lenis velar, alveolar and labial plosives), placed in an intervocalic context (e.g. /at:a/).
Next, six words with a prevocalic plosive were chosen. These words are minimal pairs in both Danish and Swedish. Investigating the prevocalic plosives is important, because in Danish word-initial plosives have not undergone lenition. The phonetic characteristics of the word-initial plosives may thus provide a more complete picture of how the fortis-lenis contrast in plosives is expressed in Danish. Also, if there is a difference between Danish and Swedish in how this contrast is expressed, it will become clear from the prevocalic position, as in this position fortis and lenis plosives are
contrastive in both languages.
table 3.1 minimal pairs used in production task. In Danish and Swedish, with English translation
contrast fortis/lenis Danish Swedish Translation
prevocalic fortis ’pakker’ ’packar’ take ’taler’ ’talar’ talk ’kalde’ ’kalla’ call lenis ’bakker’ ’backar’ back
’daler’ ’dalar’ go down ’galle’ ’galla’ gall intervocalic (word) fortis ‘lapper’ ‘lappar’ rags
‘retter’ ‘rättar’ straighten ‘lkker’ ‘läcker’ delicious lenis ‘labber’ ‘labbar’ paws
‘redder’ [ ‘räddar’ save ‘lgger’ ‘lägger’ lay intervocalic
(non-word)
fortis ’appe’ ’appe’ - ’atte’ ’atte’ - ’akke’ ’acke’ -
lenis ’abbe’ ’abbe’ [ -
21 ’agge’ ’agge’ -
intervocalic (non-cognate)
fortis ’ripper’ ’klippa’ rip cliff ’rytter’ ’smitta’ rider infection ’tikker’ [ ’snacka’ [ ticking talk lenis ’ribber’ ’klibba’ ribs cling
’rydder’ [ ’smidda’ [ clear forge ’tigger’ ’snagga’ beggar crop
The list of words eventually consists of 24 items, 12 of which are cognates.5
- intervocalic geminate plosives in existing words (which may be influenced by frequency- factors
With this list we thus investigate three different phenomena:
6
- intervocalic geminate plosives in non-words (which are considered a neutral context) , but are a good example of the natural language situation)
- Word-initial plosives (which function as a control condition)
3.2.2 Task
The lists were read out loud by three Danish and three Swedish speakers. All of the speakers are male native speakers of their respective languages. They all speak the standard language and originate from the capitals, i.e. Stockholm in Sweden and Copenhagen in Denmark.
The list each subject was asked to read consists of four parts. The first part is a list of 44 different words including 18 test items (the existing words) and 26 fillers in the subject’s mother tongue . The second part is a list of sentences, in which all words of the first part are incorporated. The sentences are of the following structure:
Danish Jeg siger nu pakker igen. Pakker. Swedish Jag säger nu packar igen. Packar. Translation I say now ‘pakker’ again. Pakker.
The third and the fourth part of the list are like the first and second part (word-list and sentence-list), but this time the lists comprise 24 non-words including six test items (the non-words) and 18 fillers. The fillers in the lists prevent the subjects from applying contrastive stress. Contrastive stress could cause the subjects to exaggerate the fortis-lenis contrast, negatively affecting natural pronunciation. Because of the way the lists are constructed, each item is read three times by every subject: once in a list, once in a sentence and once directly after the sentence. It was decided to use only one of each set of three tokens. The isolated words read in a list present the difficulty that they may be read with list-intonation. The embedded words in the sentence are subject to co-articulation (problematic for the analysis of word-initial plosives), which leaves the isolated word, read directly after the
22 embedding sentence (the underlined word in the example above) for the analysis. Besides the advantage that these words are not subject to list-intonation or co-articulation, another advantage of these words is that they are all combined with the same sentence. This makes it possible to relate intensities in the word to intensities in the sentence, to normalize for speaker differences and recording circumstances, like the distance to the microphone.
3.2.3 Acoustic correlates
The phonetic characteristics of the fortis and lenis plosives in prevocalic and intervocalic contexts that were investigated are presented in table 3.2. Although the point of interest is the geminate in intervocalic position, the initial plosive will also be measured, for reasons of comparison, as said in 3.2.1.
Eleven acoustic correlates were identified and are given in table 3.2. The grey area in the table shows three correlates which can be measured both in prevocalic and in intervocalic position. These
correlates can be used to compare the way in which the contrast is expressed in different positions in the word.
Table 3.2. Eleven acoustic correlates used to identify fortis and lenis plosives in Danish and Swedish
Prevocalic Intervocalic
1. Duration of release 5. Duration of release 2. Intensity of release 6. Intensity of release 3. Rise of following vowel 7. Rise of following vowel 4. Voice Onset Time 8. Decay of preceding vowel
9. Duration of preceding vowel 10. Closure duration
11. Closure Intensity
23
Fig. 3.1 Screenshot of phonetic measurements using Praat (Boersma & Weenink, 2009). Swedish
[(in figure written in XSAMPA)7used to illustrate measurement of phonetic correlates. The numbers refer to the numbers of the correlates in table 3.2. See appendix 1 for an enlargement of this figure.
Duration and intensity of release - correlates 1 and 2 (prevocalic) and 5 and 6 (intervocalic)
The difference between fortis and lenis sounds, besides difference in voicing, is often regarded as a difference in ‘force of articulation’, where fortis sounds are pronounced with more energy. The total amount of energy is the sum of duration and amplitude, i.e. the area under the intensity contour (see for example fig. 3.1, the intensity contour is the yellow line). It is thus expected that the duration of the release of the fortis sound is longer, or the intensity is higher, or both. Both correlates will be measured individually.
Fant (1973) identifies three stages of the release: the release transient, the fricative segment and the aspirative segment. In this study the duration of release that will be measured comprises two stages of the release: the release transient and the fricative segment. The aspirative segment, i.e. the voiceless initial part of the vowel, also called aspiration, is not considered as a part of the release in this study, because any frication in this stage is caused by a different articulator (glottal).
The intensity of the release is measured in decibels (dB), compared to the intensity of the word nu ‘now’ in the previous sentence (see 3.2.2). This word is chosen, because it is in a stress-neutral position in the sentence. Therefore, its intensity can be considered as a baseline, thus as a valid point of comparison.
24
Rise and Decay (rise of following vowel, decay of preceding vowel) – correlates 3 (prevocalic) and 7 and 8 (intervocalic)
Debrock (1977) suggests that rise and decay times should be measured as important correlates of the fortis-lenis contrast (or the force of articulation). Recordings by Smalley (1961-1964) were used to investigate the intensity curves of vowels after and before fortis and lenis stops and fricatives. The rise of a vowel is that part of the curve starting from the end of the previous consonant to the maximum intensity of the vowel (minus 10%, for reliability reasons). The decay of a vowel is the time from the maximum intensity (minus 10%) to the minimum (plus 10%). See figure 3.2 for a schematic example of the rise and decay of vowels.
^ intensity time > ^intensity time>
Fig. 3.2 Schematic representation of rise and decay time in the non-words ] and
In fig. 3.2 a schematic representation of the intensity contours of the non-words and are given on a time line. The curves represent the intensity contours, whereas the L-shapes represent the start and end of the rise and decay of the vowels. As can be seen from the figure the intervocalic fortis stop [t:] has steeper intensity contours, resulting in a shorter decay (left) and rise (right). When speaking of rise and decay as phonetic correlates in the current study, I refer to the horizontal dimension, i.e. the duration of the rise and decay of the vowels.
Rise and decay time was found to be shorter before and after fortis sounds than lenis sounds (Debrock, 1977), which can be seen in fig. 3.2 (in the left figure a fortis stop is given, in the right a lenis stop). The duration of the decay is always longer than the rise, meaning that vowels reach their peak intensity relatively fast, whereas decay of the vowel is slower.
Voice Onset Time – correlate 4 (prevocalic)
Voice Onset Time is a reliable acoustic correlate only in the comparison of fortis and lenis sounds in prevocalic position. In intervocalic position it is unreliable to measure VOT, because the voice of preceding sounds may continue through the closure of the plosive (causing the absence of VOT for lenis plosives).
VOT is the time measured from the start of the plosive (defined by the onset of the release noise burst) to the onset of voicing (i.e., the first glottal pulse after the closure), typically yielding negative VOT values for voiced sounds in voicing languages. In languages with an aspiration distinction instead
25 of VOT, aspiration time can be measured. This is the time interval between the end of the release burst and the onset of voicing. Because the voice always sets in at or after the release, the duration of the release is subtracted from the VOT, yielding aspiration time.
Danish and Swedish are generally considered aspiration languages, which calls for the use of aspiration time as a distinctive correlate. Helgason & Ringen (2008), though, finds that in Swedish prevoiced initial stops are used next to voiceless stops. In fig. 3.3 it can be seen that prevoicing (also called voice lead (Rietveld & Van Heuven, 2001) i.e. the onset of voice before the release burst) is a characteristic of the lenis stop in voicing languages. Because this type of voicing has been found in Swedish, in this study VOT will be measured as a distinctive correlate.
Aspiration language:
Voicing language:
Fig. 3.3 Schematic representation of prevocalic plosives in an aspiration language (above) and a voicing language (below). Above, in the aspiration language, the first glottal pulse of both fortis (red) and lenis (blue) start after the release burst (white arrow). Below, in the voicing language, the first glottal pulse for the lenis sound starts before the onset of the plosive(grey arrow), and the glottal pulse of the fortis sound starts shortly after the onset of the plosive. Because it is expected to find this latter kind of voicing (voice-lead) in Swedish, voicing will be measured from the start of the onset, i.e. VOT.
Duration of preceding vowel – correlate 9 (intervocalic)
The duration of the preceding vowel has been a generally accepted correlate of the intervocalic fortis versus lenis contrast since a long time. Malécot (1966) already mentions authors writing about it since 1940, Maddieson & Gandour (1976) refers to articles from 1904. The difference in duration of the preceding vowel has been considered to be due to an inherent property of the speech
mechanism, as it has been found in many different languages (see Cho & Ladefoged (1999) for a discussion on phonetic universals). Vowels before voiced consonants are longer than before voiceless ones. It is likely that the duration of the preceding vowel, the duration of the decay and the closure duration are interrelated: the vowel is elongated at the cost of the silent interval (see ‘Closure duration and closure intensity’), rendering a longer vowel and a shorter silent interval for voiced
C lenis fortis V
26 intervocalic plosives (Van Heuven, personal communication). Leaving aside the discussion on
whether the duration of the preceding vowel is a non-linguistic physiological phenomenon or a grammatical choice of the speaker, it cannot be discarded as an important phonetic correlate in the fortis-lenis contrast and it cannot be ruled out that a listener uses it as a cue in classification. As can be seen in fig. 3.1, the vowel (correlate 9) is the periodic signal containing formants between the aspiration of the prevocalic stop and either the preaspiration of the intervocalic stop or the silent interval.
Closure duration and closure intensity – correlate 10 and 11 (intervocalic)
As already stated by a.o. Lisker (1957) and Malécot (1970), closure duration is an important factor in the recognition of surd (voiceless, fortis) and sonant (voiced, lenis) sounds in intervocalic contexts. The closure duration is the (near) silence before the release of a plosive in intervocalic context. In fig. 3.1 it can be seen that this duration (correlate 10) is measured from the point after any occurring preaspiration to the point before the release of the plosive. The intensity of this closure referred to in the analyses is the intensity of the silent interval minus the intensity of the word nu ‘now’, to control for speaker- and recording variation. As stated above, fortis intervocalic plosives are expected to have a longer silent interval than lenis intervocalic plosives. The fact that in this case the duration of the release is taken as an individual factor, apart from closure duration (or the duration of the silent interval), allows the comparison of the word-initial plosive (of which only the release can be
measured) with the intervocalic plosive. The intensity of this silent interval is likely to differ between fortis and lenis sounds. Fortis sounds will not be actively voiced (or they will be actively devoiced) during this silent interval, causing the intensity to reach a lower level than it would in lenis sounds, which are actively voiced (or not actively devoiced) (Jansen, 2004) .
3.3 Hypothesis
3.3.1 Questions
To recapitulate, the questions I would like to answer by phonetically analyzing Danish and Swedish plosives are:
- How is the fortis-lenis contrast phonetically expressed in pre- and intervocalic stops in Danish and Swedish?
- Does Danish undergo complete or partial phonological lenition of stops in intervocalic position?
3.3.2 Hypotheses regarding prevocalic position
27 generally seen as the main contrastive correlate in the prevocalic stop contrast, for which divisions were made by Lisker & Abramson (1964). VOT is expected always to be positive, whereas Swedish may show negative VOTs for lenis plosives, i.e. voice-lead (Helgason & Ringen, 2008) and a positive VOT for fortis plosives.
3.3.3 Hypotheses regarding intervocalic position
In Swedish the fortis-lenis contrast is expected to be found intervocalically in all correlates measured. Most of the correlates measured are interdepent, e.g. when the duration of the preceding vowel is longer, the decay of the preceding vowel will be longer too. Especially the duration of the preceding vowel is expected to show large differences between fortis and lenis sounds, because this has been found to be a robust correlate in the intervocalic fortis-lenis opposition in many languages (Malécot, 1966). In Danish no contrast is predicted to be found. Instead, lenition of the (originally geminate) intervocalic fortis plosives is assumed to be complete (i.e. the original fortis-lenis contrast has disappeared). All plosives are expected to be pronounced like lenis plosives, but if a trace of a contrast should be found, it is expected in the rise and decay times. In a study by Debrock (1977) differences in rise and decay times in Dutch fricatives were found where voicing differences had already disappeared.
3.4 Results
3.4.1 Prevocalic position
In prevocalic position, four correlates were measured: 8 1. Release duration
2. Intensity of release
3. Rise of the following vowel 4. Voice Onset Time
Together, these correlates place the measured sounds in a multidimensional context, allowing for a view on the strength of a contrast between fortis and lenis sounds and the impact of a variable on the relative distinctiveness of these sounds. The question is which of these correlates significantly contributes to the fortis-lenis contrast and how this differs between Danish and Swedish.
In both Danish and Swedish, 9 values (3 speakers pronounced 3 minimal pairs) of each phonetic correlate in voiceless prevocalic stops are compared to 9 values of each correlate in voiced prevocalic stops. A Mann-Whitney U is performed, as there is not enough data for parametric testing.
Parametric tests make assumptions about the normality of the distribution of a dataset and they often assume there is no correlation between the mean and the standard deviation of the values. In a dataset in which each group exists of 9 values, as does the current set of prevocalic stops, these assumptions cannot be made.
28 The Mann Whitney U shows there is a significant difference in VOT for both Danish (Z=-2,874;
p=0,003) and Swedish (Z=-3,241; p=0,000). In Swedish none of the other measured correlates contributes to the contrast. In Danish there is a significant role for release duration (Z=-3,313; p=0,000), be it that the release duration is longer for the fortis stops than for the lenis stops. So within each language prevocalic stops are contrastive, and VOT is a main correlate in both languages. In figure 3.3 the data of both Danish and Swedish prevocalic stops are given in a 2D graph, seen as there are maximally two different correlates (one in Swedish, two in Danish) significantly
contributing to the contrast. The graph shows that the data are sparse. Danish voiced stops clearly form a separate cluster (apart from one outlier), but e.g. Swedish voiceless stops could easily be confounded with Danish voiced stops.
Fig. 3.4. Danish (circles) and Swedish (crosses) prevocalic voiced (blue) and voiceless (red) stops, compared by VOT and release duration.
When performing a Mann Whitney U comparing Danish and Swedish, Swedish and Danish differ in their values for VOT both in voiced (Z=-2,531; p=0,011) and voiceless stops (Z=-2,874; p=0,003). Again, as can also be judged from the graph, release duration is significantly different in Danish than Swedish for voiceless stops (Z=-3,403, p=0,000). This means the prevocalic contrast is not
29
3.4.2 Intervocalic position
In both Danish and Swedish 60 cases of intervocalic stops were pronounced, i.e. every speaker pronounced 20 words with intervocalic stops. In this position, 7 correlates were measured:9 5. Duration of release
6. Intensity of release 7. Rise of following vowel 8. Decay of preceding vowel 9. Duration of preceding vowel 10. Closure duration
11. Closure Intensity
An Anova shows that in Swedish three phonetic correlates have a significantly different mean for lenis and fortis stops. These correlates are the intensity of the occlusion (F(1)=792,840; p=0,000), being higher for lenis (M=-11,7831) than for fortis (M=-32,3091),10
In Danish, none of the correlates have significantly different values for fortis and lenis intervocalic stops. Figure 3.4 shows the values for Danish and Swedish in a three-dimensional graph, based on the correlates that show a significant contrast in Swedish.
the duration of the occlusion (F(1)=94,525; p=0,000) being longer for fortis (M=0,19110) than for lenis (M=0,13393) and the decay of the preconsonantal vowel (F(1)=7,650; p=0,008), being longer for fortis (M=0,07347) than for lenis (M=0,05763), which will be addressed in the discussion (paragraph 3.5).
9 The numbers of the correlates refer to the numbers as given in table 3.2
30
Fig. 3.5 Three-dimensional graph of Danish (circle) and Swedish (cross) fortis (red) and lenis (blue) intervocalic stops. All values on three correlates are given. The three correlates show a significant contrast in Swedish between fortis and lenis stops. The lines in the graph connect each value to the centroid of the group.
As the graph shows, the Swedish contrast is not only strong, but there is no overlap in values either. The centroids (central point of al measurements) of the groups are far apart. In Danish the two groups (lenis and fortis) completely overlap and all values point to the same centroid, indicating there is no difference between the groups. The cluster of Danish values is close to the Swedish lenis stops and the groups show an overlap.
31
3.5 Discussion phonetic analysis
In prevocalic position, a contrast between fortis and lenis stops was expected in both languages, assigning an important role to VOT as the main correlate contributing to this contrast. This expectation is partly confirmed: in both Danish and Swedish a strong contrast was found between prevocalic fortis and lenis stops. In Danish, however, this contrast was caused not only by VOT, but also by release duration. This correlate did not have a role in Swedish. Comparing the two languages, the VOT for Danish is longer than for Swedish, overall. It was hypothesized that Swedish would have voice-lead in some cases, but this was not the case. The fact that, overall, VOT tended to be shorter for Swedish does indicate the languages differ in this respect.
In intervocalic position no contrast was found between fortis and lenis stops in Danish. Lenition in Danish intervocalic stops, then, can be considered to be complete for these speakers, in the sense that intervocalic fortis and lenis stops do not differ significantly on any phonetic correlate. This is as expected. In Swedish three correlates contribute to the contrast, rendering two very separate clusters of measurements with no overlap whatsoever. An unexpected difference in means was found for the decay times, as vowels before Swedish lenis stops had a shorter decay time than vowels before Swedish fortis stops. A closer look at the data reveals a possible reason for this result. As Helgason and Ringen (2008) found previously and as discussed by e.g. Kortlandt (2003), some Swedish speakers tend to preaspirate word-medial or –final voiceless stops, meaning they utter a short -like sound before closure (i.e. after the vowel). The speakers who pronounced the items in the current study tend to do so as well, be it in a variable extent. The decay of the vowel in the current study is measured from the highest to the lowest point of intensity, which includes this preaspiration. As lenis (voiced) stops are not preaspirated, this latter decay is shorter.
The cluster of lenis intervocalic stops in Swedish shows some overlap with the Danish fortis/lenis cluster, where the mean values are significantly different on only one correlate. A significant
difference in means does not mean zero overlap, as can be seen in figure 3.4, meaning they could be confounded. This latter point is a point of interest, as the phonetic measurements have been made to be related to perception.
From the phonetic measurements I will make predictions on how prevocalic and intervocalic fortis and lenis stops will be perceived for Danish and Swedish listeners. First of all, native categorization of pre- and intervocalic stops is predicted rather straightforwardly: both Danish and Swedish prevocalic stops are contrastive, predicting an excellent classification. Furthermore, Swedish intervocalic stops are contrastive and very discrepant, predicting excellent native classification and Danish intervocalic stops are ambiguous (no contrast was found) predicting classification at chance level.
I would like to base the expectations on cross-language prevocalic stop recognition on the PAM-model of Best (1995). Below I give the six assimilation types for pairwise assimilation, as I did in chapter 1.
1) Two-category assimilation type (TC): Each non-native segment is assimilated to a different native category, and discrimination is expected to be excellent
32 other deviant). Discrimination is expected to be moderate to very good, depending on the magnitude of difference in category goodness for each of the non-native sounds
3) Single-category assimilation (SC): Both non-native sounds are assimilated to the same native category, but are equally discrepant from the native ‘ideal’; that is, both are equally
acceptable or both equally deviant. Discrimination is expected to be poor (although it may be somewhat above chance level).
4) Both uncategorizable (UU): Both non-native sounds fall within phonetic space but outside of any particular native category, and can vary in their discriminability as uncategorizable speech sounds. Discrimination is expected to range from poor to very good, depending upon their proximity to each other and to native categories within native phonological space. 5) Uncategorized versus categorized (UC): One non-native sound assimilated to a native
category, the other falls in phonetic space, outside native categories. Discrimination is expected to be very good.
6) Nonassimilable (NA): Both non-native categories fall outside of the speech domain being heard as nonspeech sounds, and the pair can vary in their discriminability as nonspeech sounds, discrimination is expected to be good to very good.
For the prevocalic stops the predictions made before the phonetic measurements have to be adjusted, judging by figure 3.4. Because a contrast was expected in both Danish and Swedish, it was predicted that these stops would not pose a problem in inter-lingual perception. As can be seen in figure 3.4 though, the Danish fortis stop is clearly distinguished from all other stops, whereas the Swedish lenis stops show overlap with the Danish fortis stops. Swedish lenis stops have a shorter VOT than the Danish lenis stops, which for Danish listeners may still be perceived as lenis. For Swedish listeners, Danish prevocalic fortis stops form a category inside phonetic space, but there is no overlap with a Swedish category. Danish lenis stops are close to both categories in Swedish, which makes it likely for Swedish listeners to show a guessing pattern in classification. It appears that Best (1995) does not give a clear prediction for this situation; if the categorizable sound would be assimilable to
one native category assimilation type (5) would be applicable, predicting excellent classification for
Swedish listeners. Since the categorizable sound is almost equally ideal for both Swedish categories I would expect the classification of Danish lenis sounds to be poor (at chance level), whereas fortis sounds may be categorized well.
Judging from fig. 3.4, both Swedish prevocalic categories are assimilable to the Danish lenis category, which leads to prediction (3): both non-native sounds are assimilated to the same native category. As both Swedish categories seem to be equally acceptable for Danish stops, classification is expected to be poor.
