L2 Acquisition of Prosody: An Underresearched Area

Although the body of research on second language (L2) acquisition in the areas of phonetics and phonology has grown steadily in recent decades, its focus has been on the acquisition of speech segments (i.e., phonemes), whereas prosody has been largely neglected. A survey of research on the phonological and phonetic properties of nonnative speech by Gut (Reference Gut2009) revealed that, of the 172 empirical studies published in international journals between 1969 and September 2008, aspects of L2 intonation featured in only nine studies, and speech rhythm was investigated in only four (e.g., Mennen, Reference Mennen2004, for intonation; Altmann, Reference Altmann2006, and Kijak, Reference Kijak2009, for stress; Rasier & Hiligsmann, Reference Rasier and Hiligsmann2007, for accentuation; Lin & Wang, 2008, for rhythm). This is surprising because nontargetlike prosody is notoriously persistent even in advanced learners (e.g., Grosser, Reference Grosser, Kettemann and Wieden1993; Rasier, Reference Rasier, Mettouchi and Ferré2003), and it is often considered to be one of the main stumbling blocks for L2 learners (e.g., “rhythm presents the greatest difficulties” [Barry, Reference Barry, Trouvain and Gut2007, p. 98]).

Perhaps unsurprisingly, the existing theoretical models of L2 phonological acquisition ^{Footnote 1} have neglected to account for L2 prosodic acquisition, or they are poorly equipped to deal with different prosodic domains (e.g., rhythm and intonation). For instance, the speech learning model (Flege, Reference Flege and Strange1995) and the perceptual assimilation model (Best, Reference Best and Strange1995) focus on sound segments, whereas the Universal Grammar-based model of stress acquisition (Archibald, Reference Archibald1994) accounts for word stress parameters only. Because these models address very specific phenomena, they cannot straightforwardly be applied to other aspects of L2 speech. Moreover, in the absence of a unified theoretical framework for L2 phonology, it is difficult to relate the models to one another, unless they were developed to deal with the same phenomenon (see Gut, Reference Gut2009, for an overview of existing models).

The objective of this article is to contribute to the development of a unified model of L2 acquisition by investigating prosodic development in learners of typologically different first languages (L1s) at different levels of proficiency; the focus is on rhythm. We vary the L1 (Mandarin and German) to examine the possibility of prosodic transfer in the learner’s L2 language system (i.e., their interlanguage; the target here is L2 English), and we vary the proficiency level of the learner groups (B1 and C1 in the Common European Framework of Reference for Languages [CEFR]; Council of Europe, 2001) to evaluate commonalities and differences in the prosodic development between different L1 groups at different levels. We argue that the various prosodic properties involved in speech rhythm do not develop in a uniform manner because learners’ L2 performance for some properties is influenced by their L1, whereas other properties follow similar developmental paths despite differences in L1s. Existing SLA theories do not predict this. Whereas these theories make different claims about how factors such as L1 and universals influence L2 acquisition, they all appear to assume that the linguistic properties within their scopes would be influenced by such factors in the same manner (e.g., Larsen-Freeman & Long, Reference Larsen-Freeman and Long1991). As we will show, we instead have to assume that rhythm development is multisystemic (Post, Payne, Astruc, Prieto, & Vanrell, Reference Post, Payne, Astruc, Prieto and Vanrell2013, for L1 prosody).

In the following sections, we first review the different approaches to speech rhythm that are relevant to the present study. These include the earliest isochrony view of rhythm, the phonological account of rhythm, and the prosody-based view of rhythm. The prosody-based view of rhythm reevaluates the relative importance of different linguistic properties in speech rhythm and motivates the proposal of a multisystemic model of rhythm development; it forms the starting point of the current study. Then the various methods for quantifying rhythm will be introduced, including those that were applied in the current research. A review of previous work on L2 prosody and L2 rhythm will conclude the Background section.

Approaches to Rhythm

Quantifying Speech Rhythm: Rhythm Metrics

Since the introduction of the continuous model of rhythm (Dauer, Reference Dauer1987), much work has attempted to capture systematic rhythmic differences across languages in a quantitative way, and, at the turn of the century, a number of metrics were introduced that quantified variability in durational variation on a phrase-by-phrase or utterance-by-utterance basis (for a recent review of rhythm metrics, see Gut, Reference Gut, Gibbon, Hirst and Campbell2012, or White & Mattys, Reference White and Mattys2007). Ramus, Nespor, and Mehler (Reference Ramus, Nespor and Mehler1999) proposed several utterance-global measures that involve dividing speech into vocalic and consonantal parts and computing (a) the proportion of the vocalic to the consonantal material (expressed as a percentage: %V), (b) the standard deviation of vocalic intervals (∆V), and (c) the standard deviation of consonantal intervals (∆C). However, because ∆C and ∆V have been repeatedly shown to be inversely related to speech rate (Barry, Andreeva, Russo, Dimitrova, & Kostadinova, Reference Barry, Andreeva, Russo, Dimitrova, Kostadinova, Solé, Recasens and Romero2003; Frota & Vigário, Reference Frota and Vigário2001; White & Mattys, Reference White and Mattys2007), Frota and Vigário (Reference Frota and Vigário2001) introduced ∆%C and ∆%V, which measure the variability of interval duration relative to sentence duration (i.e., the standard deviations of the space occupied by consonantal or vocalic intervals, whereby the space occupied by each interval was calculated by dividing the interval duration by the sentence duration and multiplying by 100). Similarly, Dellwo (Reference Dellwo, Karnowski and Szigeti2006) developed the rate-normalized metric VarcoC (the standard deviation of consonantal interval duration divided by the mean consonantal duration and multiplied by 100) and its vocalic counterpart VarcoV.

Other rhythm metrics do not measure rhythm exponents globally across the utterance but rather attempt to capture rhythm more locally by focusing on the degree of durational difference between neighboring intervals. The rationale for measures that take into account local variation is that utterance-global measures are blind to the sequencing of units, whereas rhythm “requires temporal sequencing of units with different properties” (Nolan & Asu, Reference Nolan and Asu2009, p. 65). The pairwise variability index (PVI) metric calculates the mean of the durational differences between adjacent temporal intervals in an intonation phrase (first applied in Low, Reference Low1998). The temporal intervals are usually vocalic or consonantal intervals, but they can be calculated over any interval that defines a specific stretch of speech material (e.g., syllables or feet). This raw PVI (i.e., rPVI-C for consonantal intervals and rPVI-V for vocalic intervals) contrasts with its normalized counterpart that controls for the effect of speech rate (i.e., nPVI; Grabe & Low, Reference Grabe and Low2002; Low, Reference Low1998; Low, Grabe, & Nolan, Reference Low, Grabe and Nolan2000), which is the mean of the durational differences (d) between successive vocalic or consonantal intervals (k) summed over the total number of intervals in the utterance (m), divided by the sum of the same intervals, and then multiplied by 100. This is represented in the following equation:

$$nPVI = 100 \times {{\left( {\sum\limits_{k = 1}^{m - 1} {\left| {{{\left( {{d_k} - {d_{k + 1}}} \right)} \mathord{\left/ {\vphantom {{\left( {{d_k} - {d_{k + 1}}} \right)} {\left( {{{\left( {{d_k} + {d_{k + 1}}} \right)} \mathord{\left/ {\vphantom {{\left( {{d_k} + {d_{k + 1}}} \right)} 2}} \right. \kern-\nulldelimiterspace} 2}} \right)}}} \right. \kern-\nulldelimiterspace} {\left( {{{\left( {{d_k} + {d_{k + 1}}} \right)} \mathord{\left/ {\vphantom {{\left( {{d_k} + {d_{k + 1}}} \right)} 2}} \right. \kern-\nulldelimiterspace} 2}} \right)}}} \right|} } \right)} \mathord{\left/ {\vphantom {{\left( {\sum\limits_{k = 1}^{m - 1} {\left| {{{\left( {{d_k} - {d_{k + 1}}} \right)} \mathord{\left/ {\vphantom {{\left( {{d_k} - {d_{k + 1}}} \right)} {\left( {{{\left( {{d_k} + {d_{k + 1}}} \right)} \mathord{\left/ {\vphantom {{\left( {{d_k} + {d_{k + 1}}} \right)} 2}} \right. \kern-\nulldelimiterspace} 2}} \right)}}} \right. \kern-\nulldelimiterspace} {\left( {{{\left( {{d_k} + {d_{k + 1}}} \right)} \mathord{\left/ {\vphantom {{\left( {{d_k} + {d_{k + 1}}} \right)} 2}} \right. \kern-\nulldelimiterspace} 2}} \right)}}} \right|} } \right)} {\left( {m - 1} \right)}}} \right. \kern-\nulldelimiterspace} {\left( {m - 1} \right)}}.$$

With the development and increasing application of the various rhythm metrics, studies have been conducted to investigate their robustness and ability to discriminate between languages and speaker groups (e.g., Arvaniti, Reference Arvaniti2012b; White & Mattys, Reference White and Mattys2007). Several factors have been found to affect the rhythm values. As mentioned, these include speech rate, and rate-normalized metrics like Varco and nPVI have been claimed to be more reliable for that reason. However, the normalization for consonantal metrics has been argued to be undesirable because it is likely to eliminate linguistically interesting variation (e.g., languages that allow more complex consonant clusters may be associated with a greater consonantal variability; Grabe & Low, Reference Grabe and Low2002), a view supported by the results of White and Mattys (Reference White and Mattys2007). Additionally, consonantal variability metrics are sensitive to the nature of the materials used (White & Mattys, Reference White and Mattys2007); for instance, the syllable structures of the sentences that are being produced strongly affect consonantal metric scores (Arvaniti, Reference Arvaniti2012b; Prieto et al., Reference Prieto, Vanrell, Astruc, Payne and Post2012). Other factors include segmentation procedure, such as whether glides or syllabic consonants are treated as part of vocalic intervals or consonantal intervals, as well as speaking style (see Gut, Reference Gut, Gibbon, Hirst and Campbell2012).

Among the various metrics that have been proposed, %V and VarcoV and rPVI-C have been found to be the most robust and best able to reflect crosslinguistic differences in rhythm as well as rhythmic differences between native and nonnative speech (e.g., Grenon & White, Reference Grenon, White, Chan, Jacob and Kapia2008; Payne, Post, Prieto, Vanrell, & Astruc, Reference Payne, Post, Prieto, Vanrell and Astruc2012; White & Mattys, Reference White and Mattys2007). However, the usefulness of the metrics in rhythm classification has been called into question (Arvaniti, Reference Arvaniti2009, Reference Arvaniti2012b; Barry, Andreeva, & Koreman, Reference Barry, Andreeva and Koreman2009). For example, Arvaniti (Reference Arvaniti2009) argued that “the metrics are unreliable predictors of rhythm which provide no more than a crude measure of timing” (p. 46). By extension, the idea of rhythm classes has also been questioned (Arvaniti, Reference Arvaniti2009, Reference Arvaniti and Niebuhr2012a, Reference Arvaniti2012b). We agree that rhythm metrics provide only a very crude approximation of the perceptual differences that theorists have attempted to account for under the heading of rhythm class and that using metrics to assign individual languages to a rhythm class is not very fruitful. The categoricity implied by the term is at variance with the gradiency of the evidence that has been brought to bear as well as with the inherently multidimensional nature of the exponents of speech rhythm, which involves a wide range of parameters of which timing is only one (e.g., Cumming, Reference Cumming2010; Grabe & Low, Reference Grabe and Low2002; Nolan & Asu, Reference Nolan and Asu2009). Additionally, it is debatable that acoustically defined parameters like durational metrics truly represent the percept of speech rhythm.

Although the timing differences that are quantified in rhythm metrics are only crudely related to crosslinguistic differences in speech rhythm, and they do not allow us to predict or characterize the nature of the rhythm of speech samples, they have nevertheless been fruitfully applied to quantify differences in timing patterns produced by speakers with different language backgrounds at different stages of development (e.g., Grabe, Gut, Post, & Watson, Reference Grabe, Gut, Post, Watson, Barrière, Morgan, Chiat and Woll2001; Grabe, Watson, & Post, Reference Grabe, Watson, Post, Ohala, Hasegawa, Ohala, Granville and Bailey1999; Grenon & White, Reference Grenon, White, Chan, Jacob and Kapia2008; Payne et al., Reference Payne, Post, Prieto, Vanrell and Astruc2012; Prieto et al., Reference Prieto, Vanrell, Astruc, Payne and Post2012; White & Mattys, Reference White and Mattys2007). Because the objective in this study is to reliably distinguish between speaker groups rather than characterize or predict the nature of their rhythmic behavior, the metrics are a good tool to detect any systematic differences that may exist between them.

L2 Prosody and the Acquisition of Rhythm

Research in L2 prosody has tended to focus on the influence of the native language on the L2 (e.g., 58.7% of 172 studies on L2 phonology, according to Gut, Reference Gut2009, p. 42). Previous findings suggest that learners commonly transfer prosodic features from their L1 to the L2 (e.g., Backman, Reference Backman1979; Lepetit, Reference Lepetit1989), including pitch accent distribution and realization (e.g., Benet, Gabriel, Kireva, & Pešková, Reference Benet, Gabriel, Kireva, Pešková, Ma, Ding and Hirst2012; Lepetit, Reference Lepetit1989; Rasier & Hiligsmann, Reference Rasier and Hiligsmann2007), but universal development patterns also play a role in L2 prosodic acquisition (e.g., parallels in reduced pitch range and incorrect pitch accent assignment, Gut, Reference Gut, Trouvain and Gut2007; lexical stress, Archibald, Reference Archibald1994). However, crosslinguistic parallelism in L2 production does not necessarily imply that the observed patterns derive from a universal constraint that operates in the acquisition process. In many cases, the parallel patterns can in fact be attributed to L1 strategies that happen to be shared by different L2 learner groups, instead of a linguistic prosodic property that does not manifest itself in the L2 learner’s speech, even when it is present in the L1 (see Mennen, Reference Mennen2006, for the difficulty of isolating universal development from L1 transfer). Important variables in such studies are the presence versus absence of the key property under investigation in the different L1s in the study (e.g., as in Archibald, Reference Archibald1994) and the inclusion of different proficiency levels to shed light on the development of the property relative to the L1 source and the L2 target in the interlanguage (i.e., the language system that is developing in the learner and that is, to some extent, independent of source and target; it should not be considered as a defective or deviant form of either). Because only very few studies have addressed this issue in prosodic acquisition, the relation between L1 transfer and universals remains largely uncharted territory, and it has not been investigated for L2 speech rhythm at all.

Although various quantification methods have been used to capture differences between L1 and L2 rhythm, not much research has shown evidence for rhythmic differences between L2 learners of different proficiency levels or with different native language backgrounds. Guilbault (Reference Guilbault2002) compared the rhythmic properties of the spontaneous speech of more and less experienced learners of French with Canadian English as the L1, using the variability index (VI) metric, which calculates the mean durational differences between successive syllables. He found that the less experienced learners’ VI tended to be lower than that of the more experienced learners, but the difference between the two groups was not significant. Neither the L2 proficiency nor the time spent in the target language environment seemed to have any significant effect on the L2 rhythm of learners’ speech. Similar results were obtained by Gut (Reference Gut2009), who compared the rhythm of learners before and after a 9-month stay abroad in the target language country or a 6-month training course in pronunciation. She found no effect of either the stay abroad or the training on the learners in terms of rhythmic properties as measured by the syllable ratio (SR), the average ratio of adjacent stressed to unstressed syllables (i.e., the ratio of the syllable with a full vowel to the syllable with a reduced vowel, in this study). Sarmah, Gogio, and Wiltshire (Reference Sarmah, Gogio and Wiltshire2009) used %V and nPVI-V to evaluate rhythm in L1 Thai learners of American English. They found that learners who had spent less than 6 months in the United States exhibited a higher nPVI-V and a lower %V than learners who had spent more than 18 months in the United States. Surprisingly, the group that had spent more time in the United States showed values that were more distant from native English values. Stockmal, Markus, and Bond (Reference Stockmal, Markus and Bond2005) studied L1 Russian learners of Latvian using similar metrics, including variability in consonantal durations (∆C and rPVI-C). They found significant differences between the less proficient and the more proficient learners (proficiency based on learners’ self-reports) in consonantal variability but no differences in %V and nPVI-V.

Taken together, these studies give inconsistent evidence for rhythmic differences in L2 speech depending on level of proficiency or source language. One potential explanation is that the differences in experience or length of stay between the participant groups in these studies were not, in fact, reflected in a (sufficiently large) difference in L2 pronunciation proficiency. Because none of the studies directly compared learners with different L1s, the inconsistencies could also be due to L1 effects. Another possibility is that the quantification methods that were used are not a reliable or robust tool to capture rhythmic differences in L2 speech. Therefore, we ask the following questions:

1. 1. Does the acquisition of speech rhythm (as measured by rhythm metrics) in a L2 reflect different levels of proficiency, and does this differ among typologically different L1s?

2. 2. Does the L2 acquisition of two prosodic properties that contribute to speech rhythm (i.e., accentual lengthening and final lengthening) reflect different levels of proficiency, and does this differ among typologically different L1s?

3. 3. Is the acquisition of speech rhythm related to the acquisition of the prosodic properties?

4. 4. Can the findings be accounted for under a L1 transfer hypothesis, or do they point toward a universal developmental path independent of the properties of the L1? ^{Footnote 2}

L1 Transfer and Universals

The strong phrase-final lengthening effect that we observed in the native speaker data for English and German as opposed to Mandarin was not reflected in the L2 English data. The lack of lengthening in the lower level German group could be argued to indicate that a universal mechanism is at play by which lower level learners make prosodic boundaries less prominent (at least in durational terms), regardless of their L1 background. However, when syllable structure was controlled for, the crosslinguistic differences disappeared, in both the L1 and the L2 data. This suggests that the discrepancy is related to a difficulty with pronouncing English consonants in complex syllables rather than syllable complexity or boundary marking per se. This explanation would agree with the findings for consonantal variability, which, as we previously indicated, could also be accounted for in this way, but there is no evidence that L1 transfer or universal constraints on the acquisition of these properties played a role.

The accentual lengthening data shed more light on the question, showing values for lower level German speakers that are similar to their native values, whereas higher level learners are indistinguishable from native English speakers in this respect. This indicates that the lower level L1 German learners may be transferring their L1 settings for accentual lengthening into their L2 (either at the abstract phonological level or at the level of phonetic implementation during pronunciation). However, because the Mandarin learners, who do not have accent or accentual lengthening in their source language, exhibit the exact same lengthening effects in their L2 English as the Germans, an alternative explanation is possible: The L2 development of the accentual lengthening could be an instantiation of a universal process, with L2 learners making use of a default or unmarked value of the relevant phonological properties at the early stages of acquisition. This interpretation would also fit with the pattern of results in the vocalic variability data, in which all learners who were acquiring a L2 with high vocalic variability defaulted to relatively low vocalic variability in their interlanguage, regardless of L1, and progressed in the direction of the L2 as their proficiency increased.

Results of other studies support this interpretation. For example, Barry (Reference Barry, Trouvain and Gut2007) analyzed L2 German productions by learners whose L1s are English or Russian (both stressed timed) or Korean (assumed to be syllable timed) and found that “all tend to deviate from the average native-speaker pattern in a way which reflects too little differentiation of accented syllables and unstressed syllables” (p. 112). There is also evidence that the acquisition of lexical stress, which is a prerequisite for acquiring appropriate L2 accentual lengthening, has universal aspects (e.g., Archibald, Reference Archibald1994; Pater, Reference Pater, Hannahs and Young-Scholten1997). Additionally, Rasier and Hiligsmann (Reference Rasier and Hiligsmann2007) found that universal markedness is an important factor in predicting and explaining L2 accentual patterns for learners with different L1 strategies for marking information structure (i.e., plastic vs. nonplastic languages). The process of the L2 acquisition of lexical stress and accent distribution appears to follow a universal developmental path, with learners reverting to default (or unmarked) values at low proficiency levels and proceeding to L2 values at a later stage when necessary (e.g., Van der Pas, Wissing, & Zonneveld, Reference Van der Pas, Wissing and Zonneveld2000, for lexical stress). The hypothesis that a universal mechanism is at play here therefore seems warranted, but further research on L2 English learners with other typologically different L1 backgrounds will have to confirm that the shared pattern observed here is indeed truly universal in nature.

By contrast, the distribution of the vocalic proportions (%V) in the utterances across the different L1 and L2 speaker groups provides clear evidence of a straightforward transfer effect. Native Mandarin and German are both different from English, with much lower and much higher vocalic proportions, respectively. Second language learners show intermediate values on opposite trajectories toward the L2 target that mirror each other quite closely in terms of the rate of progression. For L2 prosody, evidence for a L1 influence on L2 production has also been observed for intonation patterns (Backman, Reference Backman1979; Mennen, Reference Mennen2004).

We conclude that the development of the various properties that are involved in speech rhythm is not affected in a uniform manner by the properties of the L1 or by universal constraints on acquisition. This implies that our findings cannot be successfully accommodated within existing models of L2 phonology.

Toward a Multisystemic Model of L2 Rhythm Acquisition

The results of our study support the view that L2 rhythm development is fundamentally multisystemic in nature. If the characteristic rhythm of a language is the product of a number of systemic properties like syllable structure, vowel reduction, and the marking of heads and edges (i.e., accentual and final lengthening in this study), the acquisition of rhythm must, of necessity, be closely intertwined with the acquisition of the individual systemic properties that contribute to it. Because these properties are acquired at different stages in the L2 developmental process and are variably subject to L1 transfer effects or follow more universal developmental paths, the acquisition of rhythm should emerge gradually as the learner acquires the relevant language-specific systemic properties as well as the language-specific regularities in phonetic implementation that differ between the L1 and L2. The acquisition of one systemic property could also have a knock-on effect on the development of other prosodic properties, accelerating or delaying prosodic development overall in specific L1 learner groups (just like nonprosodic or nonphonological systemic properties in the language, such as structural differences in syntactic phrasing between the L1 and L2).

Our results fit with a multisystemic model of prosodic acquisition. First, we have shown that the L2 learners’ development of some key prosodic properties (i.e., accentual lengthening and phrase-final lengthening) is reflected in L2 rhythm acquisition, whether or not the syllable structure of the sentence materials is controlled for. Second, we have found that the learners of typologically different L1s have a similar developmental path for L2 English vocalic variability, but they follow different acquisition paths in terms of the L2 English vocalic proportion. This difference is reflected in the L2 learners’ development of different systemic prosodic properties, which are influenced by various factors such as L1 properties and universal properties, as we have discussed in the previous section. Third, we have shown that learners perform better in terms of L2 prosodic lengthening when the syllable structure of the utterances they produce is of the CV type and that they are more off target when they produce more complex structures. This indicates that there is an interaction between the acquisition of the systemic property of syllable structure and the acquisition of the systemic properties that underlie prosodic lengthening.

Such a model could be couched within a dynamic systems approach to language development (de Bot & Larsen-Freeman, Reference de Bot, Larsen-Freeman, Verspoor, de Bot and Lowie2011; de Bot, Lowie, & Verspoor, Reference de Bot, Lowie and Verspoor2007). Dynamic systems are characterized by what is called complete interconnectedness: Subsystems or variables are nested and interrelated, and change in one subsystem will lead to changes in another as well as in the system as a whole. Speech rhythm could be seen as a dynamic system that consists of various subsystems, such as syllable structure and prosodic lengthening, and L2 acquisition of rhythm would also be dynamic in the sense that it involves the development of the subsystems and the connections between them, with change in one subsystem potentially triggering change in another. The interrelatedness of the development of the systemic properties of syllable structure and prosodic lengthening with rhythm in our study could be an example of such a connection.

Further research will have to explore to what extent prosodic development does indeed show the interconnectedness between subsystems in the L1 and L2 predicted by dynamic systems theory. Also, the multisystemic model referred to here has nothing to say about the role of variability as a source of development, which is a key aspect of dynamic systems theory. In dynamic systems theory, individual variation (even within a group) is accounted for from an emergentist point of view; that is, various internal and external factors interact in a complex manner, which can lead to unpredictable results. However, we controlled for the factors that might cause intragroup variation in our study because we are interested in what commonalities the members of a whole group share in terms of rhythmic and prosodic properties, rather than in intragroup variation.

We conclude that the various systemic properties that contribute to speech rhythm interact with one another in prosodic development and that their development in the interlanguage is directly reflected in the L2 rhythm acquisition process. In this process, different prosodic properties are acquired in a nonuniform manner across typologically different languages. Second language phonological theories, and SLA theories more generally, will need to be able to accommodate the multisystemic nature of L2 prosodic acquisition while taking into account the nonuniform manner in which the prosody of the interlanguage reflects L1 transfer effects as well as universal constraints on acquisition.