Tag Archives: dual lexicon

Menn and Matthei (1992) The “two-lexicon” account of child phonology (Part 2)

In the previous post, I described Menn and Matthei’s assessment of progress on the two-lexicon model. They highlight several advantages of the model, but also note problems, including the apparent competition between children’s “selection rules” (or rules specific to the output lexicon), as well as non-deterministic cross-word patterns. To combat these and other problems, MM suggest that the formalism of the two-lexicon model migrate from a generative perspective to a more connectionist one. At this point, they make a very handy list of the key generalizations they would like to capture with a revised, connectionist two-lexicon model, or with any model of child speech production for that matter. I have restated them here, while keeping MM’s original groupings.

Reduction of Information

  1. Children recognize more words than they can say
  2. Children recognize more phonemic contrasts than they can realize in speech
  3. Early productions tend to cluster together in terms of phonetic properties
  4. Early productions also tend to contain a limited set of phonetic elements


  1. Children’s productions appear to be simplified (compared to adult forms) and often appear systematic (many words share a pattern)

Inertia of the System

  1. Early, frequently produced words may retain a high level of fidelity, resulting in “phonological idioms” compared to more recently acquired production forms
  2. Changes in systematic productions tend to happen to newly acquired words; more established words are more resistant to change

We could also add to this list MM’s frequent observation that imitated production forms tend to be much more like adult forms.

To provide a general feel for a connectionist model of early speech production, MM lay out the “initial settings” for such a model. With respect to connections, MM posit simultaneous and sequential connections. Simultaneous connections link the speech modalities of motor commands, auditory percepts, and kinesthetic sensation (of one’s own productions). The three modalities, motor/auditory/kinesthetic or MAK, must be wired together efficiently by learning. Sequential connections are within-modality connections that represent change over time. So, a simultaneous connection might link together the feeling, action plan, and acoustic record of a [b], while sequential acoustic connections might link together the [b] burst to the following formants of an [a] vowel in the syllable [ba]. Although MM do not make this explicit, it appears that sequences of connections also represent stored forms, or words.

Next, MM lay out a series of what I will call linking mechanisms. First, sequential auditory patterns can be stored and learned by attention to adult speech. Second, there is an internal feedback loop, which MM relate to babbling, which has a basic predictive property that allows the model to guess how a sequential motor pattern might sound and thereby modify it to observe whether the result is the same or different (essentially a supervised learning component provided by the stored, “correct” adult forms). Third, imitation will result in links between stored adult-produced auditory sequences and the child’s own MAK sequences. Fourth, stored adult sequences will be associated with real-word states (meanings), which then leads associations between the child’s own MAK sequences and real-world states.

MM give a fair amount of attention to the idea that adults might assist in the development of a child’s MAK sequences. The basic idea is that an adult mimics the phonetic properties of a child’s utterance (absolute pitch, formant values, etc.). Here’s an explanatory quote: “A purely sound-based imitation of the child by the adult…will produce links between the child’s internal MAK associations and the sound of the adult’s voice, the child’s innate normalization abilities should be enhanced.”

Once normalization is established (although I’m not sure why it needs to be established first in this proposal), the child might seek to produce words in a more adult-like fashion. MM propose that social factors like semantically contingent responding by parents (Snow, 1977) could provide such a mechanism. MM conclude by saying that their connectionist model is not fully developed, and that many attractive qualities of the old two-lexicon model, like the selection rules, have been replaced by vaguer concepts. However, they believe that the absolute boundaries of the input and output lexicons in the original model simply do not serve us, and we should abandon them.

My primary concern with the connectionist model that MM propose is that it seems to completely abandon the original problem that the two-lexicon model addresses. Looking back at their list of key generalizations, I would single out two, but the connectionist model does not clearly address either. First, how is it that children can recognize more words/sounds than they can produce? Second, why are children’s early productions both simplified and systematic?

It’s difficult to see how the proposed connectionist model makes headway on these problems. In fact, it seems as if they have been replaced with several other problems in the study of child speech. The discussion of speech normalization is a perfect example. Given general agreement that toddlers have a good understanding of the perceptual form of their native language, this problem could be assumed to be solved at the time that production begins. For example, I know of no evidence that children ever attempt to imitate the absolute values of any acoustic property of adult forms, which seems to be a major problem if we want to address normalization.

To conclude, I generally see the box-and-arrow iteration of the two-lexicon model as being preferable, if only for specificity. Athough I agree with MM that the box-and-arrow model could be replaced advantageously by a connectionist model, the advantages are simply not clear enough here. In the future, I will present a more recent attempt at a connectionist network by Menn and colleagues, which may address the perception-production disparity more directly.



Snow, C. E. (1977). The development of conversation between mothers And babies. Journal of Child Language, 4, 1-13.

Menn and Matthei (1992) The “two-lexicon” account of child phonology (Part 1)

Menn and Matthei (hereafter MM) begin with some information about the historical development of the two-lexicon model. They quote a paper by Ferguson, Peizer, and Weeks (1973), who noted a general human tendency to know more words than are typically said. That is, both children and adults know words that they rarely or never say. Thus, there seems to be a set of lexical representations for which the details of production are either murky or nonexistent, and we might hypothesize a split between input and output representations (Ingram, 1974), in other words, two separate lexicons.

So long as there is a consistency in children’s pronunciations, however, separate lexicons are unnecessary. If there is a regular mapping between the input representation (presumed to be identical to the adult forms) and the output representation, then a set of rewrite rules that capture the mapping are sufficient, and no output lexicon is needed. However, children are rarely consistent, and MM provide the example of two words (“down” and “stone”) that move in and out of a nasal harmony rule: They start out with no harmony ([dawn] and [don], resp.); the harmony rule then applies to other words (/binz/ –> [minz] and /dæns/ –> [næns]); finally, the harmony rule overtakes “down” and “stone”. With inconsistent mapping across similar words, rewrite rules are not helpful, or at least require arbitrary exceptions. Granted, two-lexicon models must also have lexical exceptions, but there are other advantages.

One of these advantages is that arbitrary exceptions in a one-lexicon system lead to more serious problems. The example is from Smith (1973) as interpreted by Macken (1980). The data comes from the child, Amahl, who displayed a pattern of velar harmony (/tr^k/ –> [kr^k]). Eventually, the pattern gave way to accurate production of alveolars, but one word, “took”, persisted as a regressive idiom, [gUk].

Macken assumes that this is possible because Amahl must have learned /gUk/ as the underlying form. Thus, when the harmony rule disappeared, /gUk/ would still surface as if harmony applied. As MM point out, however, this assumes that the child perceives “took” as /gUk/, which would lead us to expect that Amahl would not understand “took” as produced correctly. This seems highly unlikely, especially given our present-day understanding of children’s perceptual abilities. Furthermore, the example above with “down” and “stone” resisting a nasal harmony rule does not make sense if we assume exceptions are cases where the child has learned his own productions as underlying forms. At the very least, it would suggest that the underlying forms of words where nasal harmony does apply are perceived as if they had initial nasals. That defeats the advantage of the one-lexicon model, however, where we assume child and adult underlying forms are the same.

An output lexicon is helpful in this case because it provides a space for pronunciation representations that may be linked by a rule that operates across words or by arbitrary connections between input and output forms. Just as importantly, the output lexicon still allows children to be able to accurately perceive those words. That is, the output lexicon provides a storage facility for consistent or variable output representations while allowing for stable and accurate perception.

Despite the advantages, MM detail several problems they see with the two-lexicon model. First, it appears that selection rules—or the rules that lead to childlike forms in the output lexicon—sometimes operate over two words. This is problematic, however, if we take up the very standard assumption that combining words is done by the syntax and word combinations do not exist in the lexicon.

Another problem is that selection rules may sometimes be in competition with one another for a given word. MM give the example of productions by the child Daniel (also discussed by Menn in previous papers, I believe) of “boot” and “boat”, which are variably produced as [bup-dut] and [bop-dot] respectively. Thus, there appear to be separate labial harmony and alveolar harmony rules that compete in terms of realization of the same word. MM point out that there isn’t any sort of formalism in the two-lexicon model that allows for rule competition.

Other problems are given through the examination of daily changes in a couple of diary studies. For example, a child Jacob exhibited something like a vowel convergence, where [i] was produced like [ε]. So “tea” is first produced as [di] and then as [dεi]. “Key” was produced first as [ki], then as [xiε], and finally as [xε]. At the same time words with a mid front vowel switched between a low and high specification: “tape” was produced with both [i] and [e]. Ultimately, MM conclude that these similar words must be influencing each other in terms of production, but in a very unruly way. Similar cases are given for stress placement on two-syllables words beginning with [k] and over-application of the plural/3rd singular/possessive morpheme.

I’ll stop here for now. My next post will summarize what MM want to explain and then review the connectionist model that MM propose as a revised two-lexicon system.



Ferguson, C. A., Peizer, D. B., & Weeks, T. A. (1973). Model-and-replica phonological grammar of a child’s first words. Lingua, 31, 35-65.

Ingram, D. (1974). Phonological rules in young children. Journal of Child Language, 1, 49-64.

Macken, M. A. (1980). The child’s lexical representation: The ‘puzzle-puddle-pickle’ evidence. Journal of Linguistics, 16, 1-17.

Smith, N. V. (1973). The Acquisition of Phonology: A Case Study. Cambridge: Cambridge University Press.

N. Hewlett (1990) Processes of development and production (Part 2)

Hewlett begins discussion of dual lexicon models with basic premise that, if children have accurate perception but inaccurate production, then “there is not just a single, modality-independent lexicon in which phonological representations are stored.” (p. 28) Hewlett lists several advantages to this basic framework. First, lexical avoidance (Schwartz & Leonard, 1982) is easily explained. Second, the “rules” like fronting and gliding that apply to child speech do not need to occur in real time. In many ways, this is helpful for explaining why the rules apply to environments, rather than to particular words. Exceptions abound, however! These exceptions include regressive idioms, where a child produces a word incorrectly even though similar words are generally produced correctly; and progressive idioms, where a child produces one word correctly when similar words are produced incorrectly. The problem with idioms is where Hewlett strikes out on his own, proposing a revised dual lexicon model.

It seems likely that reproducing the box-and-arrow model from the chapter would be a violation of copyright, so I will do my best to provide verbal descriptions for now. There are four four key boxes in the model (clockwise from upper left): the input lexicon, the output lexicon, a motor processor, and a motor programmer. The input lexicon is where incoming acoustic signals are matched to stored lexical items. Hewlett states explicitly that, “The input lexicon contains perceptual representations in terms of auditory-perceptual features.”

Realization rules link the input lexicon to the output lexicon, which contains articulatory representations. From there, an articulatory representation can be sent to the motor processor, where a motor plan is assembled using syllabic units. There is an alternative route, however, going through the motor programmer. If a realization rule does not exist, or if there is cause to eschew the realization rule, then the perceptual representation is sent to the motor programmer, where a motor representations is built from scratch. From there, it can either go directly to the motor processing component for implementation, or it can go to the output lexicon for storage, or probably both. Additional levels of production mechanism follow motor processing, including a segmental level of motor processing (which is acquired after the onset of speech), a motor execution level where muscle contractions are planned, and finally the signal sent to the vocal tract, representing the actual articulations.

How well does Hewlett’s model handle the data discussed in my last post? First, lexical avoidance is explained by postulating an entry in the input lexicon that has no corresponding motor plan (Hewlett is unclear here, but I think he means there is no corresponding entry in the output lexicon). Realization rules in which sound contrasts are neutralized (fronting, gliding, etc.) are the result of multiple input lexicon entries being mapped to the same output entry. Improvement in speech accuracy over time is handled by various forms of feedback, including the revision of output lexicon forms by passing input forms through the motor programmer.

There are many positive aspects of Hewlett’s model, and it does improve on the model proposed by Kiparsky and Menn (1977). However, the empirical coverage of the model is still quite limited. Here are a few examples. First, although Hewlett is careful to point out how important phonology is for explaining paradigmatic phonological rules, his model does not include a robust phonological grammar. The input and output lexicons are connected by an arrow, but this obscures what a difficult relationship this must be. How, for example, are output lexical items merged when they remain distinct in the input lexicon (e.g., when the words ‘rock’ and ‘walk’ are pronounced identically, or when /r/ and /w/ are pronounced identically, in general)? What mechanism is responsible for the merger? Notice that previous generative approaches are not helpful here because part of the challenge is to show how the input lexicon–including words like ‘rock’ and ‘walk’–links to the output lexicon–where ‘rock’ and ‘walk’ become merged. Grammars which do not split the lexicon into input and output components are therefore shielded from this problem. Progressive and regressive idioms are also unexplained by the single arrow between the input and output lexicons. The model has no way of explaining why some words might not follow an otherwise consistent grammatical pattern.

Second, how do articulatory representations develop? Consider who a child comes to produce their first word. Based on Hewlett’s model, we can reasonably assume that the child has an accurate perceptual representation of the word in their input lexicon. How is that word then matched up to any motor representation. Presumably, babbling plays some role in the developmental process, but this is not discussed outside of input from the motor programmer. We might look to work by Guenther to solve this problem (e.g., Guenther, 2006), but Hewlett leaves the process unspecified.

Finally, Lise Menn consistently mentions the important of explaining why speech accuracy improves during imitation, but Hewlett’s model is not specific enough to account for this fact.

Overall, Hewlett’s chapter provides an outstanding review of much of the work on child speech production and phonology up to 1990. His model offers several advances compared to similar models proposed by Menn (Kiparsky & Menn, 1977; Menn, 1983), but many facts about speech development remain unexplained.

N. Hewlett (1990) Processes of development and production (Part 1)

I’m following up on my review of Kiparsky and Menn (1977) with a review of Hewlett (1990), which extends the dual-lexicon model in several interesting ways, including a more detailed production component and an updated literature review. Unfortunately, the chapter is so long that it doesn’t really seem appropriate to review it all at once. In fact, this post will probably be too long. If you’d prefer shorter posts, let me know!



Hewlett reviews major findings in normal and disordered phonological/speech development, with the goal of motivating a model of early speech production building on previous work [1, 2]. The coverage in the manuscript is extensive, and the criticism is often very insightful. Below is a short description of the findings that Hewlett covers.

Hewlett begins his review with very early speech development, including babbling.* Babbled sounds are typically the same sounds in early words, and babbling usually overlaps with the first real word productions [3]. Relevant work not discussed by Hewlett include research from Boysson-Bardies and colleagues showing that babbling sounds are language dependent and even sounds that are common in babbling around the world often have language-specific phonetic characteristics [4, 5].

When word production begins in earnest, Hewlett argues that certain aspects of early speech are consistent. First, early ‘proto-words’ [6] are highly variable in their form. Thus, although the child’s production goal might be consistent—for example, they are always referring to ‘milk’—the form is entirely inconsistent. Second, early words are generally single words or unanalyzed phrases (the parts of the phrase don’t recombine).

Hewlett argues that a separate stage can be identified around 1;6 (years; months), which roughly corresponds to what is often called the ‘word spurt’. Hewlett further elaborates on phonological systematicity during early word production. Young children apply systematic patterns to their speech. These patterns might include consonant cluster reduction (‘snow’ is pronounced [no]), or application of a child-language-specific rewrite rule (/r/ à [w] word-initially and word-medially), or application of a prosodic template, such as a [CVjVC] template [7]. Hewlett writes, “The important implication of this is that the child’s pronunciation patterns exhibit regularities which yield to a systematic description within a phonological framework.” (p. 19) Thus, the enterprise of child phonology has been either to 1) describe the child’s phonological inventory, including contrasts and phonotactic restrictions, or 2) write rules that describe how children get from the adult form, which children are presumed to know based on their perceptual abilities. I will not go into great detail about these proposals, but Hewlett reviews well-known rules such as /r/ à [w]. Finally, although Hewlett discusses the issue later in the paper, this stage of phonological development includes many examples of ‘lexical avoidance’, or cases in which children avoid words with particular sounds [8].

At this point, Hewlett reviews models of phonological development, including proposals by Jakobson [9], Stampe [10], and Menn ([2]; the dual-lexicon model, also described in [1], which I reviewed in a previous posting). He then goes on to describe children’s perceptual abilities, which are generally agreed on to be quite good. And, of course, the explosion of the infant literature starting in the early 1990s confirms that infants are very good at learning linguistic/phonological patterns before they begin to speak.

As a sort of contrasting section to `phonological development’ as described above, Hewlett reviews `phonetic development’, in which he focuses on the measurement of speech production. Several findings are noteworthy. First, children’s speech is known to be more variable, including long durations for linguistic targets and greater variability. Regarding variability, recent work by my current mentor Lisa Goffman, and her collaborations with her mentor Anne Smith, have greatly added to our understanding of speech motor variability in children. Some examples: [11] showed that oral-motor stability is below adult levels even at 14 years of age. [12] showed that, contrary to what one might expect from a frequency-based explanation, native English speaking children and adults produce iambs with more stability compared to trochees.

Continuing with Hewlett’s discussion of phonetic development, children’s formants tend to be more variable than adult’s formants [13]. Hewlett discusses the issue of whether children show more or less coarticuation than adults. A number of researchers, Susan Nittrouer being one example [14], have claimed that children actually show greater amounts of coarticulation. The implication is that children have less segmentalized speech, and therefore their early speech consists of unanalyzed whole words. This claim has been hotly debated (or was hotly debated 20 years ago), but it appears that coarticulation is often just different in children [15], without there being either more or less coarticulation in child speech.

Hewlett also discusses the issue of `covert contrasts’ or `incomplete neutralization’—cases where children appear to be producing two sounds the same but are actually producing them distinctly. For example, both /r/ and /w/ might be realized as something like a [w], but in fact, the productions are distinct, and children can reliably identify which word they intended from their own productions [16]. Elsewhere, I have argued that this is a systemic problem with analyses of child phonology. Because so much of the literature on `phonological processes’ in child speech is based on transcription data, it is unclear whether these cases reflect phonological processes or covert contrasts (in which case, `phonological’ must mean something entirely different than what it is usually taken to mean).

Hewlett concludes his review of phonetic development with three findings. First, sounds that appear in babbling may disappear from a child’s sound inventory after the onset of word production. Second, although adults are very good at compensating for a bite block and hitting acoustic targets, children may be less good at this [17]. Third, Hewlett notes that children seem readily able to acquire a foreign accent as well as a foreign language (although some more recent work [18] suggests that accent acquisition generally falls on a continuum based on age of acquisition). Regarding the last two findings, Hewlett concludes that children must be better than adults at learning to produce new sounds.


[1] Kiparsky, P. & Menn, L. (1977). On the acquisition of phonology. In Language Learning and Thought, J. Macnamara (Ed.). New York: Academic Press.

[2] Menn, L. (1983). Development of articulatory, phonetic, and phonological capabilities In Language Production, Vol II, B Butterworth (Ed.). London: Academic Press

[3] Locke, J. L. (1983). Phonological Acquisition and Change. New York: Academic Press.

[4] Boysson-Bardies, B. d., Halle, P., Sagart, L., & Durand, C. (1989). A crosslinguistic investigation of vowel formants in babbling. Journal of Child Language, 16(1), 1-17.

[5] Boysson-Bardies, B. d., & Vihman, M. M. (1991). Adaptation to language: Evidence from babbling and first words in four languages. Language, 67(2), 297-319.

[6] Menyuk P. & Menn, L. (1979). Early strategies for the perception and production of words and sounds. In Language Acquisition, P. Fletcher, M. Garman (Eds.). Cambridge, UK: Cambridge University Press. pp. 49-70.

[7] Priestly, T. M. S. (1977). One idiosyncratic strategy in the acquisition of phonology. Journal of Child Language, 4, 45-66.

[8] Schwartz, R. G., & Leonard, L. B. (1982). Do children pick and choose? An examination of phonological selection and avoidance in early lexical acquisition. Journal of Child Language, 9, 319-336.

[9] Jakobson, R. (1968). Child Language, Aphasia and Phonological Universals. The Hague: Mouton.

[10] Stampe, D. (1969). The acquisition of phonetic representation. Papers from the 5th Rebional Meeting of the Chicago Linguistic Society, 443-454.

[11] Smith, A. & Zelaznik, H. (2004) Development of functional synergies for speech motor coordination in childhood and adolescence. Developmental Psychobiology, 45, 22-33.

[12] Goffman, L. (1999). Prosodic influences on speech production in children with specific language impairment and speech deficits: Kinematic, transcription, and acoustic evidence. Journal of Speech, Language, and Hearing Research, 42, 1499-1517.

[13] Eguchi, S. & Hirsch, I. (1969). Development of speech sounds in children. Acta Otolaryngology Supplement, 257.

[14] Nittrouer, S., Studdert-Kennedy, M., & McGowan, R. S. (1989). The emergence of phonetic segments: Evidence from the spectral structure of fricative-vowel syllables spoken by children and adults. Journal of Speech and Hearing Research, 32, 120-132.

[15] Goodell, E. W. & Studdert-Kennedy, M. (1993). Acoustic evidence for the development of gestural coordination in the speech of 2-year-olds: A longitudinal study. Journal of Speech and Hearing Research, 36, 707-727.

[16] Kornfeld, J. R., & Goehl. (1974). A new twist to an old observation: Kids know more than they say. Chicago, IL: Chicago Linguistic Society.

[17] Oller, D. K. & MacNeilage, P. F. (1983). Development of speech production: Perspectives from natural and perturbed speech. In The Production of Speeech, P. F. MacNeilage (Ed.). New York: Springer Verlag, pp. 91-108.

[18] Flege, J. E., Munro, M. J. & MacKay, I. (1995). Factors affecting degree of perceived foreign accent in a second language, Journal of the Acoustical Society of America, 97, 3125-3134.