1. Introduction
In language acquisition research, naturalistic data have constituted a preferred way to observe children's speech. Naturalistic data are gathered by collecting children's speech in their natural environment – their home, their daycare centre or other familiar location – spontaneously interacting with those close to them, and with no specific instructions given by the researcher. Historically, the collection of these data was first carried out using parental diaries (e.g., in French, Grégoire Reference Grégoire1937), as direct transcriptions of children's utterances. With the advancement of technology, naturalistic data are now digitally audio and/or video-recorded and can be transcribed later on.
Naturalistic data are a useful source for language acquisition research, because they have a “high ecological validity as the recording situation closely approximates the real-life situation under investigation” (Eisenbess Reference Eisenbess, Blom and Unsworth2010:12). In a naturalistic setting, the linguistic behaviour of the child is less likely to change from her usual behaviour than it would in experimental conditions. Moreover, except for recording equipment, the collection of naturalistic data does not require specific conditions, nor the establishment of an experimental protocol, and it is accessible to any speaker.
Naturalistic data can be collected in two different ways: longitudinally and cross-sectionally. Longitudinal collection captures the continuous language development of one child, on the premise that this individual development might be generalized to the global language development of children who speak this particular language. Cross-sectional collection captures stages of language development in children of different ages, on the premise that these different stages might represent a continuous temporal development.
In both cases, the purpose of collecting spontaneous child speech is to open a window on the child's language development, in terms of stages. For this purpose, one has to decide how frequent or how long the recordings must be. For instance, in phonological development studies, naturalistic longitudinal recordings occur from every week (Fikkert Reference Fikkert1994) or two weeks (e.g., de Boysson-Bardies and Vihman Reference de Boysson-Bardies and Vihman1991, Demuth et al. Reference Demuth, Culbertson and Alter2006, Rose Reference Rose2000) to every month (e.g., Freitas Reference Freitas2003, Yamaguchi Reference Yamaguchi2012, Wauquier and Yamaguchi Reference Wauquier, Yamaguchi, Vihman and Keren-Portnoy2013). The length of a single recording session varies from 30 minutes to one hour.
But even at this frequency, recording sessions are still only a sample of the child's actual productions. Sampling data may have effects on findings on language development. Such misleading results can be illustrated with the example of overregularizations in child productions. Marcus et al. (Reference Marcus, Pinker, Ullman, Hollander, John Rosen, Xu and Clahsen1992) found that overregularizations of regular past tense to irregular verbs represented a small proportion of their data. In the reexamination of the same data, Maratsos (Reference Maratsos2000) found that a different sampling would yield a different conclusion; another study by Maslen et al. (Reference Maslen, Theakston, Lieven and Tomasello2004) showed substantial overregularizations, based on dense corpora. This example shows the importance of an adequate data sampling for a linguistic study. As Maratsos (Reference Maratsos2000) suggested, “fine-grained analyses” may be missed, because “these periods pass relatively quickly in time, or may be very sparsely sampled”.
More generally, Rowland et al. (Reference Rowland, Fletcher, Freudenthal and Behrens2008) examined the effects of data sampling on results, and they concluded that an inadequate data sampling would potentially lead to two types of misleading results. The first type is a miscalculation of errors, be they infrequent or occurring in infrequent structures. The second type is a misestimation of linguistic productivity, the chance that frequent structures are overrepresented in smaller samples.
In order to avoid data sampling issues, many studies have recently stressed the importance of denser corpora to capture an accurate picture of child language development. Denser is understood as more frequent sessions of short duration (Tomasello and Stahl Reference Tomasello and Stahl2004, Rowland and Fletcher Reference Rowland and Fletcher2006, Lieven and Behrens Reference Lieven, Behrens and Hoff2012), for a total of two to 10 hours per week, for example, or sessions of longer duration at specific points in the child's development (Gilkerson and Richards Reference Gilkerson and Richards2008, Chabanal et al. Reference Chabanal, Liegeois and Chanier2015), as a continuous five to 12 hours of recording every six months, for example.
However, working with large amounts of dense corpora raises resource issues, since they are time-consuming to record and transcribe. For instance, an orthographic and phonetic transcription of a one-hour session may take up to 30 hours of work. At this point, one might ask whether denser corpora fit the purpose of the study.
All studies cited above dealt with syntactic or morphological analyses of children's production. In a half hour of speech, different morphological or syntactic events, such as the use of different tenses, different syntactic frames, or different morphological categories may occur very rarely, even in adult speech. But in the same amount of time, adult speech displays many exemplars of sounds (about 18,000, Rouas et al. Reference Rouas, Farinas and Pellegrino2004), syllables (from about 7000 to 12 000, depending on the speaking rate, Fougeron and Jun Reference Fougeron and Jun1998), and consequently as many stress patterns, and words (about 5200, Grosjean and Deschamps Reference Grosjean and Deschamps1975). The level of linguistic investigation is important in the sampling of data: while one needs more corpora in order to observe morphological or syntactic events, a phonological or lexical investigation can be performed on a smaller data sample.
The question of data sampling has not been as well documented for phonological or lexical development in child productions, as Demuth (Reference Demuth and Behrens2008) and Edwards and Beckman (Reference Edwards and Beckman2008) stressed. Lexical development is often analyzed through the evolution of vocabulary size, the composition of the lexicon, and the variability of the different words used (e.g., Bates et al. Reference Bates, Marchman, Thal, Fenson, Dale, Steven Reznicka, Reilly and Hartung1994, Bassano et al. Reference Bassano, Eme and Champaud2005, Kern Reference Kern2007). For the purpose of this study, two lexical variables produced by children were selected: word types and word tokens. Counting word types measures the diversity of the lexicon (that is, how many different words the child produces), while counting word tokens quantifies the frequency of occurrence of words. Phonological development concerns the acquisition of sounds and phonological structures, such as syllables, feet, stress, tones, etc. The analysis of these phenomena is linked to lexical development: the more different words a child produces, the more different phonological contexts there are. Phonological development may be analyzed through lexical production, but also through sound production (e.g., Demuth Reference Demuth and Archibald1995, Rose Reference Rose2000, Beckman et al. Reference Beckman, Yoneyama and Edwards2003, Demuth and Kehoe Reference Demuth and Kehoe2006, Fikkert Reference Fikkert and de Lacy2007, dos Santos Reference dos Santos2007, Yamaguchi Reference Yamaguchi2012). I focused here on sound development, by selecting three different variables: produced sound types, produced sound tokens and target sound types. Produced sound types indicate how many different sounds a child produced, and produced sound tokens measure the frequency of each sound type. Target sound types indicate the children's selectivity with respect to the targeted sound system.
This article tackles the issue of data sampling in terms of duration in the context of studying the development of words and sounds. The goal is to identify the ideal duration of a naturalistic recorded session for it to be considered a representative sample of children's linguistic behaviours, for phonological and/or lexical questions. In this sense, ideal should be understood as long enough to reflect as faithfully as possible the child's productions, but short enough to be transcribed in a reasonable amount of time.
The identification of the perfect session duration is done using two perspectives. Currently, if a researcher wants to analyze naturalistic child productions, two options are open: either using available corpora, or recording a new corpus. With the growth of available databases in the language acquisition research community, such as as CHILDES (MacWhinney Reference MacWhinney2000) or PhonBank (Rose and MacWhinney Reference Rose, MacWhinney, Durand, Gut and Kristoffersen2014), the first option has become a valid alternative. This is the first perspective: if we have access to already-recorded data, what do we need to transcribe? If, for example, recorded sessions are one hour long, is it possible to transcribe only part of them? The second option – recording a brand new corpus – takes more time, but may be necessary in order to study rare languages, for instance. In this case, I tried to identify what the adequate recording duration was in order to study the acquisition of words and sounds.
With these two perspectives in mind, I first present the method used, detailing the corpora used and the linguistic variables analyzed: word types, word tokens, sound types and sound tokens produced, and sound types targeted. Comparisons of child productions in different recorded sessions are then exposed, and balanced with parental input. Finally, I discuss all these results and suggest what an ideal recording may be for the study of word and sound development.
2. Method
The data used in this article come from two distinct corpora: the PREMS corpus and the PSPT corpus. Both consist of longitudinal recordings of naturalistic interactions between children and their parents, all monolingual French-speakers. In what follows, I first give details about the specific participants, the collection and transcription of each set of data, and then introduce the different variables and predictions.
2.1 The PREMS corpus
This corpus was collected and transcribed within the research project PREMS, supported by French National Agency for ResearchFootnote 1. For the present study, the productions of four children from this corpus, three boys and one girl, were studied. They were recorded every two weeks at home, from the age of one to two years old. Sessions were recorded by an experimenter using a video camera and a digital audio recorder. This corpus is available on-line, as part of the CHILDES databaseFootnote 2 (MacWhinney Reference MacWhinney2000).
In this corpus, children's utterances were transcribed orthographically and phonetically using Logical International Phonetic Programs (LIPP). The transcriptions were then converted to the CLAN format (MacWhinney Reference MacWhinney2000) and then to the PHON format (Rose et al. Reference Rose, MacWhinney, Byrne, Hedlund, Maddocks, O'Brien, Wareham, Bamman, Magnitskaia and Zaller2006, Rose and MacWhinney Reference Rose, MacWhinney, Durand, Gut and Kristoffersen2014). Parents’ utterances were orthographically transcribed directly using PHON, but not all transcriptions included parental productions. Parental phonetic transcription was automatically generated with PHON. All transcriptions were made by trained Linguistics students. All phonetic transcriptions were checked, and corrected if necessary, by the author.
2.2 The PSPT corpus
This corpus was collected and transcribed within the research project “Psychological Significance of Production Templates in Phonological and Lexical Advance: A cross-linguistic study”, supported by the United Kingdom Economic and Social Research CouncilFootnote 3 (Wauquier and Yamaguchi Reference Wauquier, Yamaguchi, Vihman and Keren-Portnoy2013). For the present study, the productions of all seven children (4 boys and 3 girls) from this corpus were analyzed. Sessions were video-recorded using a camera and audio-recorded using a wireless microphone worn by the child. The children were recorded over a one-year period. The first session was recorded when they produced 20 different words on the basis of a parental questionnaire, namely the French adaptation (Kern and Gayraud Reference Kern and Gayraud2010) of the MacArthur-Bates Communicative Development Inventory (Bates et al. Reference Bates, Bretherton and Snyder1988, Fenson et al. Reference Fenson, Marchman, Thal, Dale, Steven Reznick and Bates2007). The ages of the children at the first recording session ranged from 17 to 23 months.
The corpus was transcribed directly using PHON (Rose et al. Reference Rose, MacWhinney, Byrne, Hedlund, Maddocks, O'Brien, Wareham, Bamman, Magnitskaia and Zaller2006, Rose and MacWhinney Reference Rose, MacWhinney, Durand, Gut and Kristoffersen2014). Parental productions were transcribed orthographically, and children's productions were transcribed orthographically and phonetically. All transcriptions were done by the author.
The data examined in this article is summarized in Table 1.
Table 1: Summary of the data analyzed in this article.
2.3 Comparing corpora
The aim of this article is to give researchers in language acquisition methodological tools to exploit longitudinal sessions without prior knowledge of the child's language development or of her communicative behaviour. The main factor is duration, and the comparison landmark between the children is age. In order to test the development of words and sounds, five variables were used: word types, word tokens, sound types, sound tokens and target sound types. Predictions about the influence of factors on these variables are presented.
2.3.1 Duration
In language acquisition studies, bi-weekly or monthly recordings vary from 30 minutes to one hour. In the data used here, the recordings from the PREMS corpus were 50 to 60 minutes long (mean duration = 54 minutes); these sessions are henceforth termed long sessions. The recordings from the PSPT corpus were 30 minutes long, and are henceforth termed short sessions. Long and short sessions were compared in terms of language production.
In order to compare the language productions of the same children, each long session was also divided into two equal parts, based on duration only, regardless of the number of utterances produced. Then the language production was compared in each half with that of the entire long session.
2.3.2 Age
As shown previously, (e.g., Bates et al. Reference Bates, Dale, Thal, Fletcher and MacWhinney1995), individual children of the same age do not obligatorily share the same language development stage. Nevertheless, age can be a predictor of linguistic productivity in relation to certain age ranges. For example, it has been shown that there is a correlation between age and mean length of utterances (MLU) produced by children (see Conant Reference Conant1987 for a review of studies about correlations between age and MLU).
Regarding lexicon development, one way to assess it is to use parental reports such as the MacArthur-Bates Communicative Development Inventory (Fenson et al. Reference Fenson, Dale, Steven Reznick, Thal, Bates, Hartung, Pethick and Reilly1993). This questionnaire has been standardized and used for many languages. Studies have shown a correlation between chronological age and lexical growth in production for English (Fenson et al. Reference Fenson, Dale, Steven Reznick, Bates, Thal, Pethick, Tomasello, Mervis and Stiles1994) as well as French (Kern Reference Kern2003, Reference Kern2007).
Even if there is individual variability between children, chronological age might give hints about children's linguistic development. Moreover, to avoid circularity, it is important to have an external, non-linguistic factor of the child's global development, in order to test the predictions about linguistic development.
2.3.3 Lexical and phonological variables
Five dependent variables were used in order to test the development of sounds and words according to the above factors.
As detailed above, two lexical variables were chosen: word types and word tokens produced by the children. The word type count measures the diversity of the lexicon, that is, how many different words a child produces, and the word token count quantifies the frequency of occurrence of words.
Three different variables were used for the evaluation of sound development: produced sound types, produced sound tokens and target sound types. Targeted sound types are to be understood as including the phonemes of the language, that is the 36 French phonemes that compose French words and that the children need to acquire. Produced sound tokens and produced sound types are to be understood as any sound produced by the children, even if it is not a phoneme of the French language. Phonetic transcriptions were done perceptually, but the transcribers were encouraged to use diacritics if needed. Thus, produced sound types can be a clue to the phonetic variability of the children and produced sound tokens indicate the frequency of each produced sound.
These different linguistic variables could be influenced by the duration of the recorded session, as stated in the predictions below. In these predictions, we collapse short sessions and halves of long sessions as 30-minute sessions, since we do not expect differences between the short sessions and the halves of long sessions.
(1) There would be more word types in a long session than in a 30-minute session, since the children are engaged in more and potentially more diverse activities.
(2) There would be more word tokens in a long session than in a 30-minute session, since the children have the time to produce more utterances.
(3) There would be no difference in the number of target sound types between long and 30-minute sessions, since thousands of instances of sounds may occur in 30 minutes, so every phoneme of the language has chances to be produced. The same applies for produced sound types, since the children have the chance to produce many instances of every sound they make.
(4) There would be more produced sound tokens in a long session than in a 30-minute session, since the children have the time to produce more utterances.
3. Results
In this section are presented the results relative to the predictions about the five linguistic variables in long, half and short sessions.
3.1 Focus on long sessions
In this analysis, I tried to determine whether it is necessary to transcribe a whole one-hour session in order to achieve the previously mentioned goals. Since many exemplars of words and sounds are produced in a half-hour, half a session may be sufficient. In this perspective, I tried to determine whether one half of a session is representative of the whole hour; and second, I tried to determine which half best represents the whole session.
Firstly, first and second halves are compared, to check whether one or the other was better in terms of linguistic productivity, using a Wilcoxon test with R, on the PREMS corpus, from 12 to 25 months old, for all four children. The means of each linguistic variable on the overall sessions, standard deviations, and the results of the Wilcoxon test are presented in Table 2.
Table 2: Comparison of the mean number of occurrences for the linguistic variables in each of the two halves of the long sessions.
As shown in Table 2, even if it seems that the second half of each long session is more productive in terms of word types and tokens as well as sound types and tokens than the first half, the differences found are not statistically significant for all five variables. There seems to be no effect of tiredness or habituation on the children's linguistic productivity. It is worth noting that standard deviations are extremely high, showing great variability in the data.
I then compared second halves with whole long sessions, using a Wilcoxon test with R, on the PREMS corpus, from 12 to 25 months old, for all four children. The means and standard deviations of each linguistic variable, and the results of the tests on the comparison between second halves and overall sessions, are presented in Table 3.
Table 3: Comparison of the mean number of occurrences for the linguistic variables in half #2 and in the whole sessions.
As shown in Table 3, all linguistic variables are greater in the whole long sessions than in their second half. These differences are highly significant; standard deviations are also high, showing variability in the data.
These results confirm predictions 1, 2 and 4. In the one hour sessions, children have more time to produce more utterances. Prediction 3 is invalidated by these results: there are more sound types, produced or targeted, in a whole session than in half of it. These results suggest that, with a one-hour recorded session, it is better to transcribe the whole session.
3.2 Comparing long and short sessions: what should I record?
This second comparison is different from the last one. In long sessions, nearly one hour of parent-child interactions was recorded. The question in the preceding section was about the efficiency of transcribing the whole session. In the following comparison, interactions were recorded during 30 minutes only. The parents were told from the beginning of the recording period that the sessions would be 30 minutes long. The question here is whether the duration of recording is correlated to the linguistic productivity of the child.
I compared children from the PREMS and the PSPT corpora, by selecting data within the same age range (17–24 months old). If we follow the results of the first set of comparisons, then we should expect all linguistic variables to be greater in the long sessions than in the short sessions. Word types, word tokens, sound types, sound tokens are presented longitudinally according to the duration of the recordings sessions. Results are then compared to the parental input.
3.2.1 Children's productions
The comparison of the mean number of word types between long and short sessions is presented in Figure 1, along with standard deviation bars. As displayed in this figure, the number of word types is comparable in long and short sessions. In short sessions, the range of word types goes from 13 (at 18 months old) to 213 (at 24 months old). In long sessions, the range of word types goes from 5 (at 17 months old) to 284 (at 24 months old). Contrary to the prediction in 1, there is no significant difference between the mean number of word types in long sessions (67.42) and the mean number of word types in short sessions (84.41), as confirmed by a Mann-Whitney test, with U = 822 and p = 0.060.
Figure 1: Number of word types in long and short sessions
However, the standard deviation bars on figure 1 indicate great variability among children, with data overlapping at each age point.
The comparison of the mean number of word tokens in long and short sessions is presented in Figure 2 along with standard deviation bars. In short sessions, the range of word tokens goes from 49 (at 18 months old) to 831 (at 24 months old). In long sessions, the range of word tokens is wider, and goes from 11 (at 17 months old) to 1357 (at 24 months old). Surprisingly, there is no significant difference between the mean number of word tokens in long sessions (309.3) and the mean number of word tokens in short sessions (297), as confirmed by a Mann-Whitney test, with U = 976.5 and p = 0.491. This result means that, even if the child has twice the time to produce words, she does not produce more words in a 54-minute recording session than in a 30-minute recording session.
Figure 2: Number of produced word tokens in long and short sessions
But, as displayed in Figure 2, this result conceals a great variability depending on age. Until the age of 20 months, there are slightly more word tokens in short sessions than in long ones. But from the age of 21 months, word tokens seem to be fewer in short sessions than in long ones, and this difference increases until the age of 24 months. It seems that the prediction in 2 is invalidated until the age of 20 months, but is validated from the age of 21 months.
Moreover, as for word types, word tokens show great individual variability. The extended standard deviation bars indicate that the individual productions of the children overlap regardless of the duration of the session.
The comparison of the mean number of produced sound types in long and short sessions is presented in Figure 3 along with standard deviation bars. In short sessions, the range of produced sound types goes from 20 (at 18 months old) to 46 (at 23 months old). In long sessions, the range of produced sound types and goes from 25 (at 17 months old) to 38 (at 19 months old). As displayed in this figure, there are more sound types in short sessions than in long sessions. The mean number of sound types is 33.9 in short sessions and 30.92 in long sessions. This difference is significant, as confirmed by a Mann-Whitney test, with U = 622 and p = 0.001. The prediction in 3 stated that there would be no difference in the number of sound types in long and short sessions, so this result – the children producing more varied sounds in a shorter session – is surprising. However, this result is to be taken with caution, since there is great individual variability exhibited by the extended standard deviation bars in figure 3, especially for children in short sessions. Moreover, recall that produced sound types do not necessarily correspond to phonemes of the target language, but to phones that the children produced. Since the transcribers were different for short and long sessions, it could be that the transcribers of the short sessions were more specific in the phonetic transcriptions than the transcribers of the long sessions. To support this hypothesis, the total number of phones used by the transcribers was counted, and it was found that indeed, the transcribers of the short sessions used 129 phones (including diacriticized phones), compared with 73 total phones used by the transcribers of the long sessions.
Figure 3: Number of produced sound types in long and short sessions
The comparison of the mean number of target sound types in long and short sessions is presented in Figure 4 along with standard deviation bars. In short sessions, the range of target sound types goes from 14 (at 18 months old) to 35 (at 24 months old). In long sessions, the range of target sound types and goes from 10 (at 17 months old) to 35 (at 24 months old). As displayed in this figure, the number of target sound types is similar in long and in short sessions. There is no significant difference between the mean number of sound types in long sessions (27.61) and the mean number of sound types in short sessions (30.44), as confirmed by a Mann-Whitney test, with U = 825.5 and p = 0.083. This result confirms the prediction in 3. As with the previous results, there is a great deal of individual variability, reflected in the nearly overlapping standard deviation bars for each session type.
Figure 4: Number of target sound types in long and short sessions
The comparison of the mean number of produced sound tokens in long and short sessions is presented in Figure 5 along with standard deviation bars. In short sessions, the range of produced sound tokens goes from 269 (at 18 months old) to 2249 (at 24 months old). In long sessions, the range of produced sound tokens goes from 445 (at 17 months old) to 3697 (at 24 months old). As displayed in this figure, there are more sound tokens in long sessions than in short sessions. The mean number of sound tokens is 1507.5 in long sessions and 1058.2 in short sessions. This difference is significant, as confirmed by a Mann-Whitney test, with W = 1468, p-value = 0.002. As expected in prediction 4, there are more sound tokens in a 54-minute session than in a 30-minute session. Nevertheless, as was shown in figures 1 and 2, 54-minute sessions do not display more word types and word tokens globally. This fact, like the preceding result, suggests that the number of sound tokens may not be related to the number of word types or tokens.
Figure 5: Number of sound tokens in long and short sessions
3.2.2 Parents’ productions
In order to explain these different results, an analysis of parental input was performed. This was done on fewer sessions, since not all parental utterances were transcribed in the PREMS corpus. In this corpus, only 23 sessions out of 52 were transcribed for parental input. The variables studied are word types and word tokens, since the phonetic transcription is missing for almost all sessions in both corpora.
The comparison of the mean number of parental word types and word tokens in long and short sessions is presented in Figures 6 and 7. As displayed in these figures, there are more word types and word tokens in long sessions than in short sessions. The mean number of word types is 563.43 in long sessions and 420.34 in short sessions. This difference is significant, as confirmed by a Mann-Whitney test, with W = 798, p-value > 0.001. The mean number of word tokens is 3855.52 in long sessions, and 2684.85 in short sessions. This difference is significant, as confirmed by a Mann-Whitney test, with W = 740, p-value > 0.001.
Figure 6: Number of word types in parental productions, long & short sessions
Figure 7: Number of word tokens in parental productions, long & short sessions
This result seems logical: parents produce more words in a 54-minute recording session than in a 30-minute recording session. Nevertheless, it should be noticed that parents do not produce twice as many words in a long session as in a short session.
4. Discussion and conclusion
The aim of this article was to identify the ideal duration of naturalistic parent-child interactions in order to have insights about children's acquisition of sounds and words.
The first question was about the efficiency of transcribing a whole one-hour session (if these sessions are already recorded). The first set of results suggested that, as expected, transcribing the whole session would give more data on word types, word tokens, produced sound types, produced sound tokens, and target sound types.
The second question was upstream of the question of transcription. The second set of results showed first that the development of the studied linguistic variables follows the same pattern for short and long sessions. As for quantitative results, the global results seemed to show that, as expected, the number of produced sound tokens is greater in long sessions than in short ones. As for the number of word types, word tokens, and target sound types, there was no difference between long and short sessions, and there were more produced sound types in short sessions than in long ones. Nonetheless, these surprising results need to be viewed with caution. Several hypotheses are offered to explain these results.
Age. Children of the same age may be at different levels of language development (for instance, in word productions, see Kern Reference Kern2007). This is supported by the fact that there is a great deal of variability in the data examined, as shown by the standard deviation bars, which overlap at each point. As for word tokens, the results seem to indicate that from the age of 21 months there is a difference in favour of long sessions. This suggests that age should be taken into account when deciding on the ideal duration for a recording session. Before 20 months, the difference between a 30-minute and a 60-minute session may not be relevant, but it could be significant at a later stage.
Transcription. The results for sound types are interesting, because they suggest that they are the same or greater in short sessions as in long ones. As we have seen, these results may be due to a transcription bias, since many more phonetic symbols were used in the transcription of the short sessions. This suggests that the comparison of data should be done using inter-transcriber reliability and agreement (Vihman et al. Reference Vihman, Macken, Miller, Simmons and Miller1985).
Context. The great variability in the results may also be explained by the variability of the situations in the recordings. One hypothesis is that parents may feel more involved in shorter sessions than in longer ones. It has been shown that the global involvement of parents favours children's linguistic skills (Tamis-LeMonda et al. Reference Tamis-LeMonda, Shannon, Cabrera and Lamb2004). This involvement may be reflected in the type of activities proposed during the recording session. While the children may be left alone for some time in a one-hour session, this almost never occurs in a 30-minute session. This difference may affect the linguistic production of the children. Glas and Kern (Reference Glas and Kern2015) have shown that child language use is favoured in maintenance (health care, eating time) and social activities, compared to solitary activities. Since in a one-hour session, this last type of activity is more likely to occur than in a 30-minute session, it may explain the unexpected results for word types in short versus long sessions.
Finally, it should be noted that this study focused on the question of the quantity of data needed to study the development of sounds and words. Perhaps there is a need to investigate the question of the quality of the data, in the sense of diverse kinds of production. Previous studies have shown that children's productions are different in terms of speech acts (Leaper and Gleason Reference Leaper and Gleason1996), lexicon (Gleason et al. Reference Gleason, Ely, Phillips, Zaretsky, Guo, Lieven, Budwig, Ervin-Tripp, Nakamura and Ozcaliskan2009), or referential expressions (Salazar Orvig et al., Reference Salazar Orvig, Marcos, Heurdier, Da Silva, Hickmann, Veneziano and Jisain press) depending on the types of activity they are engaged in. In this perspective, recording different activities may help analyze how, how often and when children use the different linguistic resources available to them.
At first glance, some of these results seem to go against the generalization of dense corpora in language acquisition. Actually, if dense corpora are used in the perspective of recording multiple activities and situations, the chances of recording rare events, such as rare phonemes, rare combinations of phonemes, and rare words are multiplied, which could help to provide a fuller picture of child language development.
