Perception

Learners’ production is partially based on how they perceive the target sounds. Literature on L2 phonology learning was heavily focused on production until two decades ago. Since then, however, we have witnessed a surge in studies on the importance of perception in shaping interlanguage productions. The relationship between perception and production is a complicated one. Questions such as “What determines the perception of foreign sounds?” and “How do things change with the experience in L2?” have been the subject of several studies. In the following, we will briefly look at three widely discussed models that offer explanations in L2 sound perception.

Native Language Magnet theory (NLM)

This theory, developed by Kuhl (1991, 1993, 2000), aims at explaining the development of speech perception from infancy to adulthood. Its main focus is on the dependence of perception on a given representation, and its consequence for production. NLM proposes that native language categories are prototypes, which are sounds “that are identified by adult speakers of a given language as ideal representatives of a given phonemic category” (Kuhl et al. 1992). Each one of these occupies a specific location in a space defined by certain phonetic properties (e.g. vowels by formant frequencies). These prototypes act as perceptual magnets that warp the perceptual space. Once these language-specific magnets are developed, infants lose the ability to discriminate sounds that they previously could, because the magnets distort perceptual space, making certain phonetic boundaries disappear; hence the perceptual reorganization from language-general to language-specific patterns of perception.

This seems to account for the facts relating to the changing abilities of small children in sound discrimination. Very young infants are capable of hearing all differences among the sounds in human languages, whereas adults display a reduced discrimination sensitivity outside their native language. Exposure to language produces a change in perceived distances in the acoustic space underlying phonetic distinctions. Infants’ precocious adaptation to the native language’s sound categories and adults’ difficulty in discriminating non-native phonemic contrasts have been shown in several studies. Werker et al. (1981) found that English-speaking adults had difficulty discriminating two Hindi dental and retroflex stops (/t̪/ vs. /ṭ/) that Hindi-speaking adults predictably discriminated well. Yet English-learning infants at 6 to 8 months old discriminated both Hindi contrasts. Werker and Tees (1994) showed that English learners declined in their ability to discriminate between the Nthlakampx (Thompson Salish) velar and uvular ejectives (/k’i/ vs. /q’i/) as well as the above-mentioned Hindi contrasts. For both contrasts in the two foreign languages, 6- to 8-month-olds generally performed to criterion, while only about 60 percent of 8- to 10-month-olds succeeded, and very few 10- to 12-month-olds did so; thus, it was concluded that the decline was virtually complete by 10 to 12 months, except for infants learning those two languages.

Since perceptual mappings differ for speakers of different languages, the perception of one’s primary language is completely different from that required by other languages. For NLM, the presence of a L1 language-specific perceptual filter makes L2 learning difficult, as later learning is shaped by the initial mappings. Foreign sounds are drawn to the native prototypes as a function of their distance from them in the phonetic space. More distant foreign sounds either assimilate to another prototype if they are closer to it, or do not assimilate if there is no nearby prototype. Two foreign sounds that are the same distance in the phonetic space from a native prototype are predicted to assimilate to it equally so long as one is not closer to another prototype.

NLM makes certain testable predictions in the degree of discriminability of foreign sounds in relation to native prototypes: it is predicted that assimilations to more well-separated prototypes will be more successful than to less well separated ones (Kingston 2003). For example, the prototypes of high rounded vowels that contrast in backness, such as /y/ vs. /u/, are farther apart than mid rounded vowels, /ö/ vs. /o/, because vowels are more dispersed higher in the vowel space. As a result, foreign vowels that assimilate to the high vowels would be predicted to be more discriminable than those that assimilate to mid vowels.

Despite these interesting claims, several shortcomings of NLM have been pointed out in the literature. For example, as noted by Mack (2003), the perceptual magnet effect does not seem able to account for the fact that some early bilinguals have two distinct VOT systems, whereas others with apparently similar dual-language experience, and hence presumably similar amounts of exposure to prototypical and non-prototypical vowels, do not. Therefore, it might be necessary to posit the existence of learner-specific prototypes. In that case, one would need to identify which learner-specific variables determine how a prototype is formed, which obviously is a momentous task.

It has also been shown that the perceptual magnet effect may not be robust across listener groups (Frieda et al. 1999), as well as that discrimination of unfamiliar phonetic contrasts can be improved even in adults through extensive natural experience, intensive laboratory training, or experimental manipulations that reduce task memory demands (Logan et al. 1991; Lively et al. 1993; Pisoni et al. 1982). Finally, there are several cases of children older than 12 months moving to a new country and acquiring native phonology.

Perceptual Assimilation Model (PAM)

This model, developed by Best (1995), aims to explain learner behavior in acquiring L2 sounds by accounting for the perception of the relationship between L1 and L2 sounds. The central premise is that listeners tend to assimilate non-native sounds to the native sounds that they perceive as most similar. In defining ‘perceptual similarity’, PAM draws from articulatory phonology (Browman and Goldstein 1986, 1989, 1992) in that it suggests that what listeners detect in speech is information regarding the articulatory gestures that generated the signal. Gestures are defined by the articulatory organs, constriction degree, and constriction locations. Categorizable L2 phonetic segments are perceptually assimilated to L1 phonological categories on the basis of their gestural similarity to L1 phonetic segments, unless they are uncategorizable (assimilated as an unrecognizable speech sound that gives rise to a new category) or unassimilable (heard as a non-speech sound).

PAM places emphasis on the perception (assimilation) of L2 contrasts by L2 learners, rather than on the perception of single L2 sounds. When non-native contrasting sounds are both categorizable, perceptual assimilation to the native system is predicted to show different degrees of difficulty. The non-native sounds may be phonetically similar to two different native phonemes and perceptually assimilated to separate L1 categories, which is termed ‘two category assimilation’ (TC). In such cases, the discrimination is expected to be excellent. If both non-native sounds are assimilated to a single L1 category, this will create a ‘single category assimilation’ (SC), which is predicted to be difficult to discriminate. SC assimilations are further taxonomized into those in which both foreign sounds assimilate equally to the single native category and those in which one assimilates far more than the other. In the latter case, the two foreign sounds differ in ‘category goodness’ (CG) with respect to the native category, and they are predicted to be discriminable to the extent that they do so. The members of such CG assimilations are still less discriminable than the members of TC assimilations, because they both assimilate to just one native category. Thus, we have a continuum that predicts listeners’ success in distinguishing different foreign sounds: TC > CG > SC, with CG cases varying between TC and SC depending on whether the CG differences between the foreign sounds are larger or smaller.

Support for PAM’s predictions are frequently found in the literature. The American English /w/ and /j/ appear to be assimilated to the corresponding Japanese /w/ and /j/. This is a case of TC assimilation. The English /ɹ̣/ and /l/ assimilating to a single Japanese /r/ is a case of SC assimilation (Best and Strange 1992). Polka (1991) reported that English listeners tended to assimilate Farsi voiced velar versus uvular stops (/g/-/G/) as a CG contrast, and Nthlakampx velar versus uvular ejectives (/k’/ – /q’/) as an SC contrast, with a tendency toward better discrimination of the former distinction, which is in accordance with PAM’s predictions.

As pointed out by Mack (2003), however, it is not clear what predictions PAM would make about the formation of two phonetic systems when simultaneous acquisition of two languages or very early acquisition of an L2 occurs, as the cases the model has been primarily applied to are those in which exposure to an L2 system occurs when an L1 system has already been well established.

In addition, the model, which is primarily concerned with the role of L1 in the perception of foreign sounds, is essentially static, and it does not include any means by which an existing L1 phonemic system might be altered by exposure to non-native segmental contrasts.

Speech Learning Model (SLM)

Flege’s (1995) Speech Learning Model also treats phonological acquisition with a view of phonetic approximation and interference based on perceptual judgments. This model is concerned with ‘ultimate attainment’, and thus focuses on long-term bilinguals and not on beginning L2 learners. SLM is built on the ideas of categorical perception and equivalence classification in the determination of how a learner will react to and ultimately acquire sounds in an L2. The phonetic perception of an L2 sound involves a comparison of the L2 sound with all sounds in the learner’s L1 system. SLM claims that the two phonetic subsystems (L1 and L2) are cognitively represented in a single phonological space and mutually influence one another. Learners relate L2 sounds to L1 positional allophones, and L2 perceptual failure occurs when the L1 phonological system filters out the distinctive features of L2 sounds. L1 and L2 phonetic segments can be related along a continuum; sounds are classified as ‘new’, ‘similar’, or ‘identical’ on the basis of the difference between L2 sounds and existing L1 sounds, and the model predicts how the learner will react. The different categorizations are made in terms of acoustic similarity or perceived cross-language similarity. If L2 sounds are categorized as ‘similar’, their assimilation to the existing L1 phonetic categories will be through a process of equivalence classification, and will be produced as the L1 sound (never as an authentic L2 sound). New categories will be formed for less similar and ‘new’ L2 sounds.

Applied to perception of L2 contrasts, SLM makes the following predictions: ‘identical’ sounds will present no problem for the learner, as all necessary knowledge is already available in the L1 (cf. PAM’s TC). If two contrasting sounds of L2 are designated ‘similar’ and both are assimilated to the same L1 category, discrimination will be difficult (cf. PAM’s SC or CG). For example, as we noted, the /ɑ/ – /Λ/ contrast of English (e.g. body [bɑdi] vs. buddy [bΛdi]) creates lots of problems for speakers of several languages (e.g. Spanish, Turkish, Greek, French, Arabic, and Russian, to name a few), because these sounds are perceptually assimilated to [a] in L1s, and result in discrimination difficulties of the contrast, as well as the accented production of both English vowels. If, on the other hand, there is great dissimilarity between L2 and L1 sounds, the sound will be judged ‘new’ (cf. ‘uncategoriz able’ in PAM), and it will not be assimilated to any L1 category. For example, English speakers learning French as a second language could produce French /y/ (a ‘new’ vowel for English speakers) more accurately than French /u/, because French /y/ is perceptually more distant from the closest English vowel than is French /u/, which has a near (but not identical) counterpart in English /u/.

SLM holds the view that there is no critical period after which the learner will be unable to acquire an L2 sound system; that is, adults can retain the capability for accurate perception of L2 contrasts. However, it is also stated that L2 development is constrained by age of learning. It is predicted that learners are more likely to have native-like perception with early age of learning (pre-puberty). The later the age of learning, the less likely a learner is to hear the differences between L1 and L2 sounds, because the learner’s L1 categories will be more developed and are likely to impede the formation of new categories for L2 sounds. The model also states that L2 development is further constrained by the amount of L1 use. It predicts an inverse relationship between frequent use of L1 and attainment of native-level L2 perception. That is, L2 learners who use their L1 frequently will be less likely to have native-level L2 perception.

Although SLM states that accurate L2 segmental production cannot occur unless there is accurate perception, Flege (1995) does not claim that all foreign accent is perceptually motivated. For example, it is acknowledged that the typical Spanish-accented English production of [εskul] for school can only be accounted for with reference to phonotactic constraints.

There have been several studies in the literature that showed considerable support for SLM (Bohn and Flege 1990; Flege et al. 1994, 1997; Fox et al. 1995; Rochet 1995). A more recent study (Aoyama et al. 2004) examined the role of the L1 and the perceived phonetic (dis)similarity between L1 (Japanese) and L2 (English) sounds in the production of English /ɹ̣/ and /l/ by Japanese speakers. Since English /l/ is perceptually more similar to Japanese /r/ than English /ɹ̣/, it was hypothesized that Japanese learners of English would have greater difficulty in acquiring the L2 (English) lateral liquid than the non-lateral liquid. The study looked at L2 perception and production of English liquids by Japanese children and adults at two intervals separated by one year. The results, in general, supported the hypothesis. While Japanese children’s perception of English /l–ɹ̣/ and /ɹ̣–w/ contrasts showed significant improvement after a year, the adults did not show any improvement over time. Also, the children showed greater improvement over time in the production of English /ɹ̣/ than English /l/. For /l/, neither the children nor the adults showed significant improvement. These findings support SLM’s predictions in that they show better acquisition of the more dissimilar L2 sound (English /ɹ̣/) than the similar one (English /l/). Also, age-related differences in the rate of acquisition were apparent, as only the children showed significant gains over time.

There are, however, several studies whose findings are at odds with SLM’s claims. For example, Zampini (1998; Zampini and Green 2001), examining VOTs of /p/ and /b/ in Spanish and English, found that students enrolled in an advanced undergraduate course in Spanish phonetics showed significant changes toward Spanish-like categories (toward shorter lag for English /p/) in both production and perception, but there was very little relationship between production and perception. Sheldon and Strange (1982) found that Japanese learners of English /ɹ̣/ and /l/ performed better in production than in perception, a finding that is certainly at odds with SLM’s claim that accurate L2 segmental production cannot occur unless there is accurate perception. De Jonge (1996) examined the production and perception of /ɹ̣/ and /l/ for Japanese speakers and of /b/ and /p/ for Arabic speakers. She found that while Arabic speakers mastered the contrast at early stages of proficiency, Japanese speakers did not even at high levels of proficiency. She used her results to evaluate SLM’s (and PAM’s) claims and found that neither of them could fully account for the data. There is also a problem with SLM’s claim on ‘new category’. For example, French front rounded vowel /y/, which is a ‘new category’, would be perceptually differentiated accurately from both back rounded and front unrounded French vowels, as well as from English vowels. However, some studies (Strange et al. 2004, 2005; Levy 2004) showed conflicting findings on perception of this vowel by American listeners.

The three models that have been looked at here (NLM, PAM, and SLM) capture the important insight that non-native contrasts are not uniformly poorly perceived. Instead, the difficulty with which a particular non-native contrast is perceived by listeners from a particular L1 background depends on the relationship between the sounds of the L1 and L2 in question. We also see that the perception–production relationship is a complicated one. Learners can have highly accurate perceptual abilities, but relatively inaccurate production ones. Alternatively, they may have more target-like production abilities than their perceptual ones. The different cues and skills used in perception and production are real challenges in understanding the learner’s knowledge.