The first report of a connection between vocabulary learning and phonological short-term memory was published in 1988 (Baddeley, Papagno, & Vallar, 1988). At that time, both Susan Gathercole and I were involved in longitudinal studies, investigating the relation between nonword repetition and language learning. We both found a connection. Now, almost 20 years later, in her Keynote Gathercole (2006) reviews a multitude of data bearing on the interpretation of this often replicated connection. Her main conclusions are three. First, both nonword repetition and word learning are constrained by the quality of temporary storage. She sees this storage as multiply determined, that is, affected by factors like perceptual analysis, phonological awareness (ability to identify and reflect on the speech sounds that make up words). Second, both nonword repetition and word learning are also affected by sensory, cognitive, and motor processes. Third, an impairment of phonological storage is typically associated with specific language impairment (SLI) but this may not be a sole causal factor.

The conclusions are closely related to the hypothesis of the phonological store in the Baddeley and Hitch framework as a language-learning device (Baddeley, Gathercole, & Papagno, 1998). According to this hypothesis, temporary storage in the phonological loop determines how well new material can be permanently presented in long-term memory (LTM). Two alternative explanations for the relation have been put forward. According to one, proposed by Snowling, Chiat, and Hulme (1991), nonword repetition reflects already existing language knowledge. Children with greater vocabularies are more likely to find similarities between the words they already know and the nonwords they are asked to repeat, which is likely to help them in the repetition task. Thus, nonword repetition is sensitive to past language acquisition speed, which is likely to be a good predictor of future word learning. A second major competitor to the temporal phonological storage hypothesis is that nonword repetition performance reflects a general phonological processing factor, affecting temporary phonological storage but also phonological awareness and the ability to rapidly name objects and symbols (Wagner & Torgesen, 1987). This view has been promoted by, for instance, Metsala (1999) and Bowey (1996). Various phonological processing tasks have also been found to be correlated with word learning in typically developing populations, children with SLI, and second-language learners (Gajar, 1987).

Teasing out causal relations between highly intercorrelated variables, like phonological awareness, phonological memory, and vocabulary, in correlational studies has proved difficult. None of our measures seem to be pure. Moreover, they may measure different cognitive processes in different populations. For instance, digit span may be a measure of learning of digit names in 3-year-olds, phonological memory without rehearsal in 5-year-olds, working memory capacity for rehearsal in 7-year-olds, phonological memory for efficiently rehearsed material, but also for skill with elaborative strategies, in university students. However, we may not need to select between the temporary phonological storage, the language knowledge, and the general phonological processing hypotheses, described above. I would like to propose a variant of the temporal storage view that incorporates both the competing hypotheses as well.

The main proposition is that what Baddeley has termed the phonological store and Gathercole calls temporary storage of phonological representations gives us a snapshot of a phonological learning device. What we see in nonword repetition, or other short-term storage tasks, is a measure of the efficiency of this learning device. I propose that this device is a phonological network that primarily creates representations as the result of top-down influences on incoming speech material. The quality of the representations is multiply determined, both by the signal to noise ratio of the perceptual signal and by the top-down mapping processes that associate it with knowledge in LTM. Unlike Gathercole (2006), I do not think that the top-down effects seen at encoding and/or storage (Luotoniemi, Service, & Maury, in press; Thorn, Gathercole, & Frankish, 2005) as well as redintegration are irrelevant to learning. On the contrary, more wordlike material is both easier to represent in this network and to remember later (Service & Craik, 1993). Top-down effects may, however, not be a source of significant variation between individuals with vocabularies of by and large similar sizes, for instance, children of the same age. Moreover, top-down effects may compensate for a generally noisier network, which would be why repetition of wordlike items is not predictive of individual differences in word learning.

The feature that makes temporary storage tasks so sensitive to the quality of representations in this learning network may be their requirement for simultaneous representation. For instance, in nonword repetition, stimuli with unpredictable syllabic structures, consisting of multiple phonemes, are presented at the rate of normal adult speech. The more phonological material needs to be active at the same time, the more competing information is there for each feature of the stimulus (e.g., each phoneme, each phonetic feature, each stress value). Noise can be controlled by top-down means (association to preexisting units with strong representations). In the case of common real words, such associations are well established; so well established that single-word repetition to a healthy person is a trivial task and says little about the efficiency of phonological storage. The same demand for discriminability between representations is also critical for their later recall. Noisier representations will be harder to retrieve among similar neighbors and more prone to proactive interference from material encountered earlier. It is also possible that noisy representations in short-term memory are a sign of inadequate top-down mapping of some of the individual phonemes. In this case, the representations will not necessarily add to the strength of previous representations for the same item, for instance at the second encounter of a new word. Instead a new, competing, representation will be created, which will add noise to the top-down process when the item is heard a third time. This type of mechanism would lead to a double problem in a noisy system. The system would be inadequate in representing an incoming item and top-down help would be hampered with competition between several imperfect candidates.

We have not studied SLI, but our studies of adults with developmental dyslexia suggest that they have a phonological memory deficit, detectable with digit span, nonword span, and nonword repetition (Laasonen et al., 2006; Laasonen, Service, & Virsu, 2001, 2002). Their short-term recognition memory deficit extends to binary sequences of tones (two frequencies), light flashes (in two locations), and touch stimuli (pricks of two fingers). They also have separate problems with tasks that require fine time discrimination between two streams of stimuli. We found sensory memory to be correlated with phonological memory in both our control group and the dyslexic group (although their performance on individual measures of a single construct was more variable). We interpret our findings to mean that developmental dyslexia, in at least a subgroup of individuals, is related to the ability to simultaneously represent multiple stimuli that are encoded with respect to a limited set of stimulus features. This problem is accentuated in the language domain where repeated exposure does lead to learning of phonological categories and patterns, however, at a clearly impaired rate.

We believe that noisy traces in representational networks lead to slower processing, which is at the core of the difficulties our dyslexic group had with temporal acuity tasks. This could also explain why the SLI group in the studies by Gathercole and her colleagues were more impaired in nonword repetition than temporal storage of the same syllables in a paced sequence. The syllables in nonword repetition tasks are delivered fast, and little time is available for a noisy network to settle and clean up a representation. Single syllables, one at a time, allow for compensatory time between items. At least, this hypothesis would be worth testing.

To summarize, a learning device for phonological material is sketched. Its efficiency can best be seen in tasks requiring the simultaneous storage of multiple phonological units. The quality of the created phonological representations can also be seen in long-term learning as well as in various tasks of phonological processing.