L2 Phonetic and Phonological Acquisition

Research on L2 phonetic and phonological acquisition (similar to research on L2 acquisition within other linguistic domains) has focused on searching for evidence of developmental sequences and on describing and explaining patterns observed in the acquisition of L2 sounds and sound systems. Much research in this area has examined the acquisition of new L2 contrasts, positing and supporting the general claims, represented by models of L2 speech perception and production (e.g., the Perceptual Assimilation Model [PAM], Best, Reference Best and Strange1995; PAM-L2, Best & Tyler, Reference Best, Tyler, Bohn and Munro2007; the Speech Learning Model [SLM], Flege, Reference Flege and Strange1995), that new L2 sounds are typically associated, at least initially, with their closest L1 counterpart and that the degree of similarity between L1 and L2 phones can help predict the degree of difficulty in the perception and production of a new L2 sound. Flege’s (Reference Flege and Strange1995) SLM, for example, posits that L2 learners relate L1 and L2 sounds to each other on a “position-sensitive allophonic level” (p. 239), such that positional allophones in the L2 are perceptually linked with the closest positional allophones in the L1. Flege’s model hypothesizes that learners are more likely to establish new categories for L2 sounds that are perceived as more dissimilar to the closest L1 sound and that, in cases in which phonetic dissimilarity between L1 and L2 sounds is not perceived, a single phonetic category is used to process both sounds—a phenomenon Flege refers to as “equivalence classification” (p. 239). In these cases, the sounds are likely to resemble one another in production. Finally, Flege recognizes that even when L2 learners establish a new phonetic category for a L2 sound, the properties of that category may differ from the native category for that sound, especially if nonnative speakers construct the new category around different features than those native speakers use to make distinctions. Although the SLM has generally been employed to model the acquisition of new L2 phonemic contrasts (e.g., Flege, Takagi, & Mann, Reference Flege, Takagi and Mann1995, Reference Flege, Takagi and Mann1996), it has also been used in previous studies of bilingual /l/ production (e.g., Barlow, Branson, & Nip, Reference Barlow, Branson and Nip2013; Simonet, Reference Simonet2010; reviewed in more detailed later in this section), and its hypotheses are relevant for the present study’s investigation of L2 allophonic acquisition.

L2 Allophonic Acquisition

In general, less is known about the L2 acquisition of noncontrastive sounds, or allophones, than about the acquisition of new L2 contrasts. Acquisition of L2 allophones involves acquiring both the phonetic features or acoustic properties of the particular variants used in the L2 as well as the L2-specific contextual patterns of use of these variants. Nevertheless, by definition, the nontargetlike employment of L2 allophones generally contributes not to any sort of potential meaning differences (as contrastive differences can) but rather to issues such as foreign-accented speech. One example of research on allophonic development is Shea and Curtin (Reference Shea and Curtin2011), which investigated the acquisition of the stop-approximant alternation of Spanish phonemes /b d ɡ/. In Spanish, the voiced plosives /b d ɡ/ are realized as stops [b d ɡ] after pauses, after nasal consonants, and, for /d/, after laterals; all other instances of these phonemes are realized as approximants [ß ð ɣ]. Measuring the deployment of two acoustic cues—consonant intensity and presence of a release burst—in different contexts (word-initial and word-medial positions, stressed and unstressed contexts), these authors demonstrated that, generally, learners with more language learning experience exhibited cue use that more closely approximated that of native speakers than did learners with less language learning experience. Additionally, Shea and Curtin found that phonological environment interacted with learner developmental level to affect the production of the phonetic detail of the L2 allophones: Learners with greater experience exhibited a more nuanced pattern of production in terms of the influence of position of the target sound in a word and lexical stress as well as the specific consonant being produced, as opposed to using a basic rule, such as word position alone, across all three consonants. Given these results, Shea and Curtin proposed that learners’ representations of these L2 phonetic categories become more detailed with greater experience with the target language: Their high intermediate learner group made use of more information (i.e., place of articulation details in addition to positional information) than the low intermediate learners (who demonstrated use of positional information but not place of articulation details), suggesting more robust and detailed representations of the L2 phonetic categories being acquired with greater experience.

The present study continues this line of research by investigating L2 allophonic acquisition, both in terms of phonetic detail and contextual patterning, but it investigates a learning situation that involves a different type of learning task than that examined by Shea and Curtin: one that involves ignoring or inhibiting L1 patterns (as opposed to acquiring new distinctions). As will be discussed in more detail in the next subsection, English and Spanish both contain lateral categories, but Spanish laterals are “lighter” or produced in a more fronted position overall than English laterals, which tend to be darker or more backed. In addition, English lateral allophonic production is conditioned by syllable position (i.e., a lighter or more fronted [l] is produced in onset positions; a darker or more velarized [ɫ] occurs in coda positions), whereas, in Spanish, a light lateral is deployed in both onset and coda positions. Thus, in addition to acquiring the phonetic features of Spanish [l] (i.e., the overall lighter realization), a primary challenge that English-speaking learners of Spanish face in this arena may actually be in inhibiting the native language’s contextual constraints on allophonic production. That is, English-speaking learners of Spanish must learn that Spanish /l/s are not conditioned by syllable position as English /l/s are. This may pose a special challenge because, similar to Escudero’s (Reference Escudero2005) discussion of subset scenarios—or learning situations in which learners must reduce the number of L1 categories to have an accurate representation of the new L2 sound system—in this case, learners must inhibit use of [ɫ] and instead deploy the same allophone, [l], in both onset and coda positions. The learning difficulty of subset scenarios arises from the fact that “the nonexistence of an L1 property in the L2 environment can be discovered on the basis of positive evidence alone” (Escudero, Reference Escudero2005, p. 204). ^{Footnote 1}

Relevant to the present discussion is Maye and Gerken’s (Reference Maye, Gerken, Howell, Fish and Keith-Lucas2000) account of the acquisition of phonetic categories based on the statistical distribution of exemplars in the input even in the absence of minimal pairs. In that study, two groups of native English-speaking participants were trained on a new contrast in an artificial language simply by hearing tokens along a continuum of /d/ (e.g., “day”) to unaspirated /t/ (e.g., “stay”), with the manipulated difference being formant onset frequencies of the following vowel. Although both of these sounds exist in English, these two sounds do not constitute a phonemic contrast for English speakers because they never occur in the same context. Both groups of participants heard the same stimuli but with differing frequencies of tokens along the continuum: One group received input with a bimodal distribution (i.e., hearing tokens at both ends of the continuum more frequently than those in the middle); the other group received input with a monomodal distribution (i.e., hearing tokens in the middle of the continuum more frequently than those at either end). During the test phase decision task, participants from the bimodal group correctly recognized /d/∼/t/ pairs as different significantly more often than those from the monomodal group, suggesting that the former group of participants was able to learn a new contrast using distributional properties that indicated the existence of two phonetic categories in the new language.

The present study asks, then, what happens when the L1 distribution of a particular category is bimodal and the L2 distribution of this same category is monomodal? Can learners acquire L2 distributional patterns and phonetic categories that involve ignoring a distinction made in the L1? This study explores these questions through an examination of the production of the Spanish /l/ by English-speaking learners and follows Shea and Curtin’s focus on contextual knowledge to explore the acquisition of the acoustic and distributional properties of this phone. The next section describes the phone under examination and its properties in the L1 and L2 of the present participants and reviews previous studies that have examined its realization in bilingual Spanish.

Laterals in L1 English and Spanish

The /l/ segment in English is generally described as a voiced alveolar lateral approximant that has two variants in American English: a light variant that occurs in onset or prevocalic positions, as in the word lamps [læmps], and a dark, or velarized, variant that occurs in postvocalic or coda positions as in feel [fí:ɫ] (Bronstein, Reference Bronstein1960; Olive, Greenwood, & Coleman, Reference Olive, Greenwood and Coleman1993). The light [l] is typically articulated with the tip of the tongue touching the alveolar ridge, whereas the dark [ɫ] is generally produced through a primary articulation including contact between the tongue tip and the alveolar ridge and a secondary articulation of the dorsum of the tongue arching upward at the velum (Ladefoged, Reference Ladefoged2006; Whitley, Reference Whitley2002). It is this secondary articulation that results in what is referred to as velarization or darkness (Ladefoged, Reference Ladefoged2006). It should be noted that, in American English, all /l/s are relatively velarized (Ladefoged, Reference Ladefoged2006; Sproat & Fujimura, Reference Sproat and Fujimura1993), although the degree of velarization varies on the basis of several factors including position of /l/ in the syllable and surrounding vowels (i.e., with back vowels conditioning darker /l/ productions; e.g., Lehiste, Reference Lehiste1964; Sproat & Fujimura, Reference Sproat and Fujimura1993). Acoustically, the difference is most often characterized by differences in the second formant (F2; a measure often used to determine tongue frontness/backness during vowel production), in which a lower F2 indicates a darker or more velarized production (e.g., Recasens, Reference Recasens2012; Thomas, Reference Thomas2011).

The /l/ in Spanish is also characterized as a lateral approximant whose standard allophone is also apicoalveolar (Hualde, Reference Hualde2005). Unlike in English, /l/ in most varieties of Spanish is pronounced as a light variant in all environments (Hualde, Reference Hualde2005). Spanish does contain several /l/ allophones that arise in specific contexts, typically as a result of assimilation of the place of articulation to that of the following consonant (e.g., a dental allophone [l̪] as in falda [fál̪.da] “skirt”; Hualde, Reference Hualde2005; Quilis, Reference Quilis1993), but all of these allophones are generally considered to be light.

Acoustically, the articulatory and gestural differences between English and Spanish /l/s result in formant differences between Spanish and English laterals such that Spanish laterals exhibit substantially higher F2 values than do English laterals (e.g., Recasens, Reference Recasens2012; see Figures 1 [Spanish] and 2 [English]). Additionally, Spanish laterals are not thought to exhibit F2 variation based on the position of /l/ in the syllable as English laterals do (see Figure 2).

Figure 1.

Production of Spanish word sol “sun” by native Spanish speaker participant (P2). Coda /-l/ has F2 of 1560 Hz.

Figure 2.

Production of English word lull by native English-speaking first-year learner of Spanish (P121). Onset /l-/ has F2 of 1136 Hz; coda /-l/ has F2 of 826 Hz.

Laterals in Bilingual Spanish

The lateral phoneme has previously been utilized to examine phonetic and phonological systems in bilingual populations. Simonet (Reference Simonet2010) analyzed Spanish and Catalan alveolar laterals as produced by two groups of adult Majorcan Catalan-Spanish sequential bilinguals: Catalan-dominant bilinguals (i.e., Catalan L1) and Spanish-dominant bilinguals (i.e., Spanish L1). Catalan, like English, contains a dark /l/. Simonet found that Catalan-dominant speakers speaking Catalan produced significantly darker /l/s than Spanish-dominant speakers speaking Spanish, corroborating the difference between Catalan and Spanish /l/s described in the literature. Catalan-dominant speakers also produced significantly darker /l/s in Spanish than Spanish-dominant speakers, and Spanish-dominant speakers produced significantly lighter /l/s in Catalan than Catalan-dominant speakers, suggesting influence of speakers’ dominant language on their nondominant language. Additionally, within-speaker comparisons revealed that the majority of the Catalan-dominant speakers demonstrated intermediate formant values in their nondominant language (Spanish) that were significantly different from their dominant /l/ formant values but were also significantly different from Spanish-dominant speakers’ formant values. This statistical difference between L1 and L2 /l/ categories ^{Footnote 2} suggests that these speakers had developed a “new” category for their nondominant language, although the new category was not identical to the category produced by native speakers of that language.

Barlow et al. (Reference Barlow, Branson and Nip2013) analyzed pre- and postvocalic lateral productions by Spanish-English bilingual, Spanish monolingual, and English monolingual children to explore whether Spanish-English bilinguals demonstrate productive allophonic differences in lateral realization in English (i.e., darker /l/ in postvocalic positions). Results indicated that, prevocalically, lateral productions, in both Spanish and English, of the bilingual children were realized with relatively high F2 values and large differences between F2 and the first formant (F1) and were comparable to the productions of the Spanish monolingual children. However, postvocalically, bilingual children produced /l/s with lower F2 values and smaller F2-F1 differences in English only. The authors concluded that the bilingual children have a merged phonetic category for English and Spanish [l] in prevocalic positions but maintain distinct phonetic categories for English and Spanish /l/ in postvocalic positions.

The present study aims to investigate similar questions with adult English-speaking L2 learners of Spanish, who are acquiring Spanish later in life than Barlow et al.’s (Reference Barlow, Branson and Nip2013) bilingual children and Simonet’s (Reference Simonet2010) early sequential bilinguals and who must contend not just with acquiring new phonetic properties (i.e., along the lightness/darkness continuum) as Simonet’s Catalan-Spanish bilinguals do, but must also learn the contextual patterning of allophones particular to Spanish, which involves inhibiting a distinction made in L1 English.

Acoustic Analysis

All target words were isolated and the lateral segments were analyzed using Praat (Boersma & Weenink, Reference Boersma and Weenink2014). This study follows others in considering the primary cue to the vowel-lateral boundary location to be “spectral discontinuity” (Sproat & Fujimura, Reference Sproat and Fujimura1993), most clearly observed as a shift in F2 frequency and a decrease in the intensity of F2 (e.g., Simonet, Reference Simonet2010; Sproat & Fujimura, Reference Sproat and Fujimura1993). The lateral onset was marked at the first pitch period that preceded a break in formant structure and/or a change in the darkness of F2, usually coupled with a drop in the amplitude of the waveform (e.g., White & Mattys, Reference White and Mattys2007). For word-initial laterals, the onset was marked at the beginning of the first pitch period in which F2 could be observed (following Simonet, Reference Simonet2010). Similarly, lateral-vowel boundaries were placed at the beginning of the pitch period “consistent with the body of the vowel (e.g., unfricated and of comparable amplitude)” (White & Mattys, Reference White and Mattys2007, p. 507).

As Recasens (Reference Recasens2012) noted, determining the segmental boundaries of dark laterals can be particularly difficult. Additionally, as has been widely documented in sociolinguistic research on /l/ vocalization in English (e.g., Hall-Lew & Fix, Reference Hall-Lew and Fix2012), the distinction between a dark /l/ and a vocalized /l/ is especially difficult to determine. Thus, in cases in which the boundary between the lateral segment and the preceding or following context could not reliably be determined (as was common in English productions and in many of the learners’ Spanish productions), the lateral tokens were coded as such and were removed from the dataset prior to the formant analysis. Additionally, other production types that prevented the clear reading of formants (e.g., fricativized laterals and creaked or devoiced productions) were also excluded.

Second formant measurements were taken from the midpoint of the lateral segments using a script in Praat. The settings included a window length of 0.025 s (the Praat standard) for all tokens and five formants estimated under 5500 Hz for females and under 5000 Hz for males. A randomly selected 10% of each participant’s data were manually checked; this method both ensured consistency in coding from early through late stages and verified the formant readings taken by the script. All observed errors in formant tracking were corrected, and if more than one token per speaker demonstrated errors in formant tracking using the settings previously described, then an additional 10% of tokens for that speaker were manually checked. If more than one additional token showed problems, that speaker’s data were recoded manually. Of the 105 total participants, 14 showed more than one error in the first round of manual checking, and six required recoding of the entire set of tokens. Formant measurements were manually changed or tokens were excluded due to formant tracking problems; such changes or exclusions affected 176 Spanish tokens, or 1.4% of the entire Spanish dataset and five English tokens (i.e., 0.002% of the entire English dataset).

To account for physiological differences (i.e., differences in vocal tract size) that could affect sound production and resonance structure, all F2 values were normalized using a vowel centroid ratio (VCR) that was derived from Watt and Fabricius’s (Reference Watt and Fabricius2002) vowel S-centroid normalization procedure, which calibrates speakers’ vowel frequency values according to each speaker’s F1∼F2 center of gravity. The VCR utilizes measurements of the speaker’s front vowel /i/ and back vowel /u/ ^{Footnote 9} to determine a centroid value along only the F2 dimension. The VCR additionally incorporates a half-spread measure (i.e., a measure of the distance between the front and back extremes) to account for differences in vocal tract spaces that could be lost by using the centroid measure alone. Thus, productions that are VCR normalized are submitted to the following equation: (F2[token]–F2[centroid])/half-spread. ^{Footnote 10} In the present study, F2 values that are VCR normalized are reported along a continuum from +1 to −1 with speakers’ centroid always at 0. A value closer to +1 indicates a higher F2 and, thus, a more fronted tongue position; a value closer to −1 indicates a lower F2 and, thus, a darker or more backed production.

Statistical Analysis

There was a theoretical ceiling of 8,610 Spanish /l/ tokens (105 participants × 82 tokens) and 1,870 English tokens (85 participants × 22 tokens). Nevertheless, 733 Spanish tokens across all levels and speakers and 10 English tokens were either not produced at all, were produced incorrectly (e.g., Participant 39 said carason instead of caracol for “snail”), or were produced in a way that interfered with the identification of the segment (e.g., such as laughing through the production). Additionally, as previously mentioned, there were several phenomena that were observed in the production of laterals (especially in the learner corpus) that interfered with formant measurements and resulted in the exclusion of these tokens from the formant analysis (n = 2,470 in Spanish; n = 1,027 in English). ^{Footnote 11} The final corpus analyzed consisted of 5,407 Spanish tokens and 841 English tokens (a breakdown of Spanish token production by level is provided in the Results section).

The normalized F2 values for all Spanish data were submitted to a linear mixed model with restricted maximum likelihood in SAS (SAS Institute Inc., 2013). Visual inspection of residual plots revealed that the data were approximately normal and indicated the desired homoscedasticity. The model was fitted with participant and token as random effects and speaker level, task, syllable position, preceding vowel, stress, and cognate status as well as the interaction between level and all other independent factors as fixed effects. The model yielded no significant interactions between level and stress or level and cognate status nor any main effects of stress or cognate status; thus, these interactions (Level × Stress; Level × Cognate Status) and factors (stress and cognate status) were removed from the model in a stepwise fashion, and the model was refitted. The final model’s fit of the data did not differ significantly from that of the original model, χ²(17) = 19.70, p = .290.

A separate linear mixed model was run for learners’ Spanish and English /l/ data. The setup and results of this model will be mentioned briefly in the Discussion section and note 12 for comparative purposes but do not constitute a central component of this study.