by Thomas Berg

In the course of their language development, children change their initially incorrect pronunciation in the direction of the adult language norm. For example, the correct rendition of the word "good" is preceded by the production of [du:t]. This raises the question of which processes take place in the child's mental lexicon during the change from [d] to [g]. The analysis of a case study indicates that these and similar changes are not localised at the level of linguistic units such as features, sounds and words, but in the connections between these levels.

When a child grows into a linguistic community, it is faced with the alternative of either delaying its first linguistic utterances until it is able to reproduce them to a largely adult-like extent or producing words at an earlier stage when it does not yet fulfil the requirements for correct production. The result is an utterance that deviates from the adult language norm. Despite not inconsiderable individual differences, children generally favour the latter strategy. This enables them to train their cognitive and motor skills through speaking ("learning by doing") and to satisfy their need to communicate - provided that the discrepancy to the adults' expectations does not stand in the way of communication. However, this strategy of "jumping the gun" also brings problems with it. It forces the children to replace the wrong pronunciation with the correct one at some point. The difficulties arising from this task should not be underestimated, as the incorrect pronunciation over a longer period of time can solidify into a pronunciation habit that is not easy to overcome.

How should we visualise this process of self-correction? Let's take the word "good" as a starting point, which many children initially pronounce [du:t]. (Square brackets indicate the pronunciation, not the spelling of a word). We can assume here that the child intended to speak a [g], but did not achieve this goal and articulated the incorrect sound [d] instead of the target sound. For the sake of simplicity, we speak here of a substitution process [g] -> [d].

Now, language is structured hierarchically and the hierarchically superior units are composed of a limited number of hierarchically subordinate units. The sound [g] can therefore be found in different words (cf. "ganz", "gleich", "Hagel" etc.). The structuring of language in the form of levels means that a sound unit such as [g] is sufficient to "supply" all words containing a [g]. Just as words can be broken down into sounds, sounds can also be broken down into smaller units called features. These can have an articulatory basis. For example, both [g] and [k] share the feature [velar], because in both cases the place in the mouth that determines the sound is the soft palate (velum). The linguistic hierarchy and the principle of supplying larger units with smaller ones can be visualised with the help of features, sounds and words as well as connecting lines between them, the so-called lines, as shown on the left.

The maturing child now has (at least) two possibilities to initiate the correction process: It can carry out the repair at the sound level by replacing the substitute sound [d] with the correct [g]. This would solve his pronunciation problem with a single intervention, virtually in the blink of an eye. Alternatively, the child can replace the wrong sound in each word individually. The two approaches have clearly different empirical consequences. While the first means that the [g] is mastered equally well in all words after the repair, the second involves a lexical variation that leads us to expect the correct production of the [g] in one word alongside the simultaneous replacement of the [g] in another word.

To the adult observer, the first method would appear to be the more efficient, as it identifies the problem as a single problem, whereas the second method splits the problem into as many individual problems as there are words in the child's lexicon that require correction. This makes this method considerably more time-consuming and demands the child's attention again and again. However, it is anything but clear whether these arguments carry any weight for the learner. This is because it is implicitly assumed that both paths are available to the child and, above all, are equally easy to follow. This certainly casts doubt on an overly mechanistic view.

In order to clarify how children actually tackle the repair problem, extremely detailed documentation of the language acquisition process is required. All relevant words ("types") as well as their possibly different forms of realisation ("tokens") must be recorded. Such a maximum data density can only be achieved through extensive daily contact with the learner. This requirement not only led to the decision in favour of an individual case study, but also in favour of my daughter Melanie as an (involuntary) informant. Of course, it was not possible to record all of her statements. Therefore, a number of restrictions had to be made. Firstly, only the correction of [d] -> [g] and of [t] -> [k] at the beginning of words was analysed. Secondly, all tokens were only recorded once per day, so that the same word that was spoken correctly five times and incorrectly once on the same day was only recorded twice (once as the correct form and once as the incorrect form). The resulting distortion of the token frequencies could be compensated for to some extent by recording the tokens anew each day. However, this concession was unavoidable if one did not want to (or could not) keep track of the child all day with pencil and paper. The observation lasted two or more hours a day and resulted in an average of 30 utterances per day, each containing different words beginning with [k] or [g] in adult language.

The investigation began at a time when Melanie, aged 3 years, 4 months and 9 days, pronounced her first [k] (in the word "children") correctly, after it had long been replaced by [t]. This memorable day deserves a closer look. A total of 45 tokens were recorded - 30 words that begin with [k] in adult language and 15 words that begin with [g]. Of the 30 words with [k], 20 were pronounced correctly and 10 incorrectly; of the 15 words with [g], 3 were pronounced correctly and 12 incorrectly. The [k] therefore had a success rate of 67%, the [g] a rate of 20%. Considering that not a single correct production was observed for the previous day, this is a considerable increase in learning.

Despite this high initial speed, 15 months passed before the [k] and [g] were fully mastered, of which only the first 12 were recorded as described. This already reveals an initial result. Melanie was clearly not in a position to solve her pronunciation problem "in the blink of an eye". In other words, she was unable to make a correction at the phonetic level. The graph on p. 26 outlines the course of the acquisition process, showing the percentage of correct [k] and [g] productions as a function of time (divided into weeks). The total number of children's utterances included in the analysis is 11,224.

The success curves for [k] and [g] do not run in a straight line from bottom to top. Roughly speaking, three phases can be distinguished. In the first 4 weeks, the percentage of correct productions in relation to the total number of productions increases rapidly (from 0% to 88%). This is followed by a very long consolidation phase (weeks 5-45), during which the level of correctness is roughly maintained despite numerous fluctuations. Finally, in the last 5 weeks of data collection, the curves rise again, whereby the time of 100% mastery of the two sounds lies outside the recorded period. It is not difficult to recognise that the [g] generally performs less well than the [k], but on the other hand there are also striking parallels in the course of the two curves.

The fact that the development from [d, t] to [g, k] does not take place in an instant implies the possibility that there are special conditions which favour this development and thus influence the speed of the sound change. Four main factors have been identified:

• the sound to be learnt itself. As already mentioned, Melanie's [k] productions show a higher degree of correctness than her [g] productions.
• the phonetic context. This factor proves to be the most important. Velars are more likely to be pronounced correctly before vowels than before consonants. For example, Melanie finds it easier to pronounce the [g] in the word "good" than in the word "same".
• the stress. In stressed syllables, [k] and [g] are articulated correctly more often than in unstressed syllables.
• the single word. How well one and the same sound is mastered also depends on the individual word in which it occurs. The words can even have a very similar sound structure. For example, Melanie had difficulties with the [g] in "ganz" for a relatively long time, while it was acquired without any problems in "gut".

If we now return to the question of the processing level at which the child applies its repair strategy, it is advisable to take another look at the diagram (see p. 24). There are five possible starting points in this model (and thus three more than initially assumed): the word level, the connections between the word level and the sound level, the sound level, the connections between the sound level and the feature level and finally the feature level. Changes at each level appear very differently in the empirical evidence: If a change occurs at the word level, all associated sounds must benefit from it to the same extent, since the word is a holistic entity and is equally connected to each of its sounds (case 1). If a connection between the word level and the sound level is optimised, only the sound at the end of this connection is affected by the learning process, as each sound has its own connection to the word above it (case 2). If the change occurs at the sound level, it is visible in all words in which this sound occurs (case 3). If the learning gain concerns a line between the sound and feature level, only the production of the respective sound or feature can be improved in this way (case 4). And if a feature is made more accessible, this favours all sounds associated with this feature (case 5).

Melanie's utterances allow us to demonstrate the level (or levels) at which her sound change took place. To do this, we can refer to two cornerstones of her development: the different changes in the velar sounds in different words and the events of the first day. Let's start with the latter. If the change is at the feature level, the two velar sounds must have a largely identical acquisition profile. However, this is not the case. The sounds [k] and [g] have a certain independence from each other and develop independently to a certain degree. This rules out the feature level as the site of change. It is equally unconvincing that the change took place at the word level. This would mean that Melanie repaired several dozen words individually within less than 24 hours. Although this cannot be completely ruled out, it is unlikely for the reason that this level only leads to individual changes, whereas Melanie's development has a more global character. Since the connections between the sound and word levels also have an individual character, this level cannot be considered for the same reasons.

If Melanie had made a change at the phonetic level, many, though not necessarily all, words would have benefited. In principle, this prediction is compatible with her learning growth being spread over many words. However, with this explanation one would also expect other sounds to remain unaffected by the change, since at the phonetic level each element is independent of its neighbours. However, the two velars do not exhibit such independence in their development. Both begin to change on the same day and do not develop completely independently of each other over the entire observation period. This tendency towards synchrony in the development of [k] and [g] cannot be reconciled with the hypothesis that the change took place at the phonetic level.

What remains are the connections between the phonetic and feature levels. In fact, there is much to suggest that the changes that took place on the first day are located at this intermediate level. In particular, this level can capture the apparent contradiction that the development of [k] shows both a certain dependence and a certain independence from the development of [g]. The independence is explained by the fact that two independent lines have been changed - the one from [velar] to [k] and the one from [velar] to [g]. The dependency, on the other hand, results from the fact that both lines converge at the same point on the feature level. This indicates that the initial changes in Melanie's mental lexicon are not to be localised at the level of linguistic size units, but in the connections between the phonetic and the feature level.

If, in the next step, we focus on the different development of [g] in "gut" and "ganz", it immediately becomes clear that the three lower levels (see diagram on p. 24) can be ruled out, as they predict effects that would have to be distributed over a large number of words, which is precisely what is not the case here. The two remaining levels differ in the local or global character of the sound change. If the word level is affected, each change must affect all sounds of the word equally. If, on the other hand, the connections between the word and the sound level are affected, the change remains limited to the respective individual connection. Melanie's data suggest that the change from [t] to [k] (or [d] to [g]) is an isolated process. There is no evidence that this change occurs in parallel with another change in the same word. For example, it is conceivable that in the word "komisch", which Melanie initially rendered as [to:mis], the "sch" would replace the [s] at the same time as the articulation of the [k]. The absence of such cases suggests that the change took place in the lines between the [g] and the word level. In this example, Melanie changed the line from [g] to "gut" faster than the line from [g] to "ganz".

As a result of the investigation, it can be stated that Melanie only tackled two of the five levels outlined. It is certainly no coincidence that these two levels are connecting levels. The sound change thus obviously only takes place between, but not on, linguistic levels of description. The question posed at the beginning, whether sound change is a phenomenon of the phonetic level or the word level, must therefore be answered with a "neither-nor".

How can this finding be explained? First of all, it must be repeated that the present study is an analysis of an individual case. In this respect, the extent to which the results can be generalised must remain open. There are at least the following possibilities for Melanie's language development: It is possible that the acquisition of the linguistic units of size was already complete when Melanie endeavoured to solve her problem with the velar sounds. One would then have to assume that linguistic units are generally acquired before the lines connecting them. This could be related to an asymmetry between perception and production, according to which only that which has already found its way into the memory via the perceptual route can be produced. For example, Melanie was able to hear the difference between "cellar" and "plate" even at a time when [k] and [t] still coincided in her language production. These linguistic units, which are already mentally represented ex hypothesi, would therefore be perceptual units that the child can use for production at a later point in time. The prerequisite for this hypothesis, however, is that fully functional circuits are not required for speech perception.

There is also the possibility that linguistic units are not acquisition-relevant to the same extent as conductors. This would mean that the dynamic aspects of language take place between the linguistic levels of description. This would be a new perspective for linguistics, which traditionally places linguistic units at the centre of its interest. Central areas of linguistics such as language variation and language change would thus appear in a new light.

The author
Prof. Dr Thomas Berg (39), an English linguist in the Department 11 of Literature and Linguistics, studied English and Romance Studies in Kiel, Braunschweig and Hull, completed his doctorate in 1986 on German and English slang and came to the University of Oldenburg in 1990. He habilitated in English linguistics in 1995 with a thesis on the relationship between language processing, language structure and language change. His research focus is psycholinguistics. On 1 April 1997, he was appointed to a professorship in linguistics at the University of Hamburg.