|Discussion Topic 7:
Individual differences in L2 acquisition
Most studies of L2 speech learning have revealed important differences between
individual participants (Ss), especially late leaners who began learning their L2 in
adolescence or adulthood. No one has as yet managed to provide a convincing account
for the substantial inter-subject variability that is a hallmark of late L2 learners. To my
mind, doing so should be considered the Holy Grail of L2 acquisition research in general,
and of L2 speech learning in particular.
Most of the research carried out in my lab in Birmingham employed a randomized block design. This approach involves
recruiting Ss who differ according to one or more variables of interest such as AOA (an index of age of L2 learning), LOR
(a rather poor estimator of amount of L2 input) and self-estimated language use (another rather imprecise index of L2
input). We used ANOVAs to determine which groups differed significantly.
I think that these conclusions are unwarranted. Examining between-group differences does not imply that the
performance of individuals is unimportant. Comparing groups has made it possible to identify variables of general
importance for L2 speech learning, and has created the necessary context in which later research can attempt to
account for individual differences.
To illustrate the importance of input in L2 speedh learning, let's consider some upublished data I presented in a 2012
chapter entitled The role of input in L2 speech learning. We tested native Spanish Ss living in Birmingham at 6-month
intervals. At each session our native Spanish Ss (and native English controls) recorded English sentences later rated for
overall degree of foreign accent. The native Spanish Ss also responded to questions pertaining to their use of Spanish
Fig. 2 presents individual subject data. Of the 15 Ss who completed the study the three labelled "best" had a
substantially better pronunciation of English than did the remaining 12 native Spanish Ss, both at the beginning of the
study and at its conclusion. Why? I think it was due, at least in part, to differences in English language input.
Fig. 4. Mean foreign accent ratings obtained for male and female
native speakers of Italian differing according to age of arrival in
Canada. Flege et al. (1995), Journal of the Acoustical Society of
America, 97, p. 3129
Another study pointing to the potential importance of L2 input on individual importance can be seen in Flege et
al.(1995). The existing literature suggested that female L2 learners will have a better pronunciation than will male L2
learners. As shown above, this is indeed what was found for the native Italian Ss who arrived in Canada up to about
the age of 12 years. For those who arrived after about the age of 17 years, on the other hand, the reverse held
true. An important gender difference existed among the native Italian subjects who arrived in Canada as young
adults, At the beginning of their new life in Canada, the women tended to work in the home whereas nearly all of the
men began working outside the home almost immediately. I speculate that most of the English input the women
received during their fist few years in Canada was the foreign-accented pronunciation of English provided by their
short-term memory) be evaluated at Time 1 using L1 speech materials.The results of MacKay et al. (1991)
suggest that PSTM may help account for inter-subject variability in L2 learning. It would also be valuable to
determine if the FA ratings obtained from one group of listeners at Time 1 will account for variance in the FA
ratings obtained from another group of listeners at Time 2. If such a finding is obtained, it might point to
In his 1997 Lund University dissertation, Duncan Markham suggested that an examination of between-group differences
has tended to "obscure the reasons for the difficulties experienced by most non-child language learners than to explain
them" and that theories such as the SLM while providing some "useful information" explaining "some learning behaviour"
are nevertheless are "insufficient" inasmuch as "they assume that age and experience are the only important
determiners of achievement."
Fig. 3 shows the average percent English use estimates obtained for the three Ss having the best pronunciation of
English and the three Ss having the strongest foreign accents in English ("worst" for short).The "best" Ss reported using
English more often, on average, than did the "worst" Ss. In particular, the "best" Ss reported using English with friends far
more often than did the "worst" Ss. The difference between the two subgroups of three in self-reported use of English
with friends diminished over time. I speculate that the "best" Ss worked hard to become integrated into the
English-speaking community when they arrived in Birmingham, forming friendships with native English speakers. As they
began mastering English and getting settled into their careers, the "best" Ss may have begun socializing more other with
other native speakers of Spanish while at the same time maintaining their existing relationships with native English
speakers. This interpretation will need to be evaluated in a prospective study in which Ss are selected on the basis of
both quantity and quality of L2 input received.
Most of the research published by members of my research group in Birmingham included summaries of the performance
of individual Ss. We often consider the number of individuals per group who met some standard of performance (e.g.,
obtaining a scores that falls within +/- 2 SDs of the mean value obtained for a native speaker comparison group).
Moreover we normally used regression techniques in an attempt to explain the performance of all Ss sampled,
irrespective of group membership.
The ultimate goal of behavioral research is to account for variance in performance. To illustrate this point in the domain
of L2 speech research, let's consider two studies which examined overall degree of perceived foreign accent (FA) in
English sentences spoken by 240 Koreans and 240 Italians who had immigrated to North America and had lived in
primarily English-speaking communities for decades.
As reviewed in a 2009 chapter entitled "Give input a chance!" I carried out principal components analyses of the FA
ratings obtained from the Koreans and Italians. The analyses for both immigrant groups yielded very similar results,
accounting for 72% of the variance (see Table 9.5, p. 186). I interpreted these findings to mean that the two primary
factors that determine how well the participants eventually pronounced English was the extent to which they maintained
their L1 (Korean or Italian) and the quality and quantity of English language input they had received.
When people think of individual differences in L2 speech (or language) learning ability
they normally think about some known or unknown characteristics of the learner that
influences overall success or some specific measure of performance. This is perfectly
reasonable. However, a trend that seems to be emerging in the literature that does not
seem reasonable to me is to ignore factors that, while likely to be highly predictive of
speech learning outcomes, do not reflect inherent characteristics (or apitudes) in
individuals. I'll illustrate via a "thought experiment".
Let's imagine that an ANOVA reveals that both the quantity and quality of instruction
affect the geometry proficiency scores and that a modest correlation of 0.30 exists
between geometry aptitude and proficiency. The question of interest is whether the
aptitude scores predict proficiency independently of the quantity and quality of instruction
There are two problems with this approach. First, the Ss -- at least in our research -- have not been randomly selected
from a population of L2 learner. Second, the groups constituted in this way normally differ in some factor(s) other than
the nominal selection variable. To help mitigate the second problem, we have used a subgroup matching technique to
control for confounded variables. (See, for example, our 1999 JML study, Age constraints on second language learning.)
Accounting for variance
Prof. Jones believes that she has identified a cognitive factor that might account for
individual differences in the ability to learn geometry. To test her hypothesis she recruits
1000 high school students who are assigned to receive varying amounts of geometry
instruction (1, 2, 3, 4 or 5 hours per week) from five teachers known to differ in their skill
in teaching geometry (1=worst, 5=best). The 40 students in each of the 25 resulting
subgroups (amount of instruction x quality of instruction) are balanced for general
Intelligence and socioeconomic status. Prof. Jones administers her aptitude test on the
first day of the geometry class and obtains a measure of geometry proficiency on the last
day of class.
As can be seen in Fig. 1, there was no change in the native Spanish Ss' overall pronunciation of English over the 5-year
study interval, but substantial differences in degree of foreign accent (FA) existed among the 15 late learners. I attribute
the lack of change in FA to the fact that Ss improvements in pronunciation had already occurred prior to the Ss
enrollment in the study.
In a recent talk at the 2016 New Sounds conference in Aarhus, I explained why individual subjects, not groups, must
be considered to be primary unit of analysis in research carried out within the SLM framework (see slide #16). For the
SLM, the perceived relationship between sounds in the L1 and L2 is of primary importance. Given that the phonetic
categories of individuals in their L1 may vary, perceived L1-L2 dissimilarity may in turn vary, leading in some cases to
differing L1-L2 "mappings".
This question can be evaluated by entering the aptitude scores in a stepwise multiple regression analysis following
amount and quality of instruction. Will the aptitude scores account for a significant amount of variance in proficiency after
the effects of the two "obvious" predictor variables have been entered into the model? If so, the new cognitive factor
identified by Prof. Jones is of interest even if accounts for substantially less variance than the two "obvious" predictor
variables. An especially impressive finding would be to find that aptitude continues to account for a significant amount of
variance in proficiency if entered following the measures of General Intelligence and SES. It would be a mistake,
however, to ignore the predictive power of amount and quality of instruction simply because these variables are so
Only 15 of the 135 Ss originally recruited completed all 11 sessions. These Ss had arrived in the US at an average age
of 29 years (SD 6, range 19-39 years). At the beginning of the study they had a mean age of 34 years (SD 6, range
24-41) and had lived in the US for an average of 4 years (SD 1, range 2-6 years). Importantly, the 15 native Spanish Ss
differed widely according to self-reported use of English (M = 60% at Time 1, SD = 21, range = 20-100%),.