How to measure the onset of babbling reliably?*

INGE MOLEMANS; RENATE VAN DEN BERG; LIEVE VAN SEVEREN; STEVEN GILLIS

doi:10.1017/S0305000911000171

How to measure the onset of babbling reliably?*

Published online by Cambridge University Press: 06 September 2011

INGE MOLEMANS ,

RENATE VAN DEN BERG ,

LIEVE VAN SEVEREN and

STEVEN GILLIS

Show author details

INGE MOLEMANS: Affiliation:
University of Antwerp
RENATE VAN DEN BERG: Affiliation:
University of Antwerp
LIEVE VAN SEVEREN: Affiliation:
University of Antwerp
STEVEN GILLIS: Affiliation:
University of Antwerp

Article contents

Abstract
EXPERIMENT 1
EXPERIMENT 2
Footnotes
References

Rights & Permissions

Abstract

Various measures for identifying the onset of babbling have been proposed in the literature, but a formal definition of the exact procedure and a thorough validation of the sample size required for reliably establishing babbling onset is lacking. In this paper the reliability of five commonly used measures is assessed using a large longitudinal corpus of spontaneous speech from forty infants (age 0 ; 6−2 ; 0). In a first experiment it is shown that establishing the onset of babbling with reasonable (95%) confidence is impossible when the measures are computed only once, and when the number of vocalizations are not equal for all children at all ages. In addition, each measure requires a different minimal sample size. In the second experiment a robust procedure is proposed and formally defined that permits the identification of the onset of babbling with 95% confidence. The bootstrapping procedure involves extensive resampling and requires relatively few data.

Information

Type: Articles
Information: Journal of Child Language , Volume 39 , Issue 3 , June 2012 , pp. 523 - 552

DOI: https://doi.org/10.1017/S0305000911000171 [Opens in a new window]
Copyright: Copyright © Cambridge University Press 2011

INTRODUCTION

In every domain of science that relies on measuring phenomena in empirical data, there is a need for measuring reliably. In the domain of language acquisition, in which a primary research method consists of collecting naturalistic observational data, the use of valid methods for reliably measuring particular phenomena in those naturalistic observations should be a matter of great concern. Although phrasing this concern may sound like stating a truism, Tomasello and Stahl (Reference Tomasello and Stahl2004: 101) start their article with the observation that ‘[T]here has been relatively little discussion in the field of child language acquisition about how to best sample from children's spontaneous speech, particularly with regard to quantitative issues’. In that paper, the authors focus on – among other things – the size of the sample and the periodicity of sampling in longitudinal studies. They cogently argue that if a phenomenon has a certain incidence ‘in the real world’, a particular sample size is required in order to capture that phenomenon with a particular degree of confidence in observational data. For instance, suppose segment /x/ has an incidence of 1/100 (it occurs once every 100 tokens) and segment /y/ has an incidence of 1/1,000. It is straightforward to see that in a sample of, say, 100 tokens segment /x/ is expected to occur at least once, and segment /y/ is not expected to occur at all. Now suppose – for the sake of the argument – that a researcher collects one hour of observational data from a normally developing child and videotapes a so-called late talker also for one hour. The sample of the former may consist of 1,000 tokens of segments, while the sample of the latter may consist of only 100 tokens, not surprisingly because late talkers are well known to be less voluble or talkative. The researcher analyzes both samples and observes that in both samples the frequent segment /x/ occurs, but that segment /y/, the segment with a low incidence, only appears in the speech of the normally developing child and not in that of the late talker. The (hypothetical) researcher concludes from this observation that late talkers produce less low-frequency segments, and may develop a theoretical account of why this is the case, and may even formulate the clinical implications of this observation. But in this example the researcher's observation is simply due to a difference in the size of the samples that were analyzed, and the theoretical and clinical implications that our hypothetical researcher draws from them may be completely erroneous because of a methodological flaw.

The evaluation of sample size issues in the computation of formal measures of language acquisition and development was highlighted quite recently by, among others, Tomasello and Stahl (Reference Tomasello and Stahl2004) and Rowland, Fletcher and Freudentahl (Reference Rowland, Fletcher, Freudenthal and Behrens2008). But scattered throughout the literature are reports that phrase a similar concern with respect to often used measures such as MLU (mean length of utterance; Klee & Fitzgerald, Reference Klee and Fitzgerald1985) and type/token ratio (Richards, Reference Richards1987; Malvern, Richards, Chipere & Durán, Reference Malvern, Richards, Chipere and Durán2004). The basic message is that if a measure is applied to the data of two different children, the yardstick used for measuring should be the same in the two cases: the sample sizes should be equal as well as the unit in which those sizes are measured. For instance, lexical diversity crucially depends on the size of the sample, and that size may be expressed in terms of utterances or in terms of words, leading to very diverse results (Hutchins, Brannick, Bryant & Silliman, Reference Hutchins, Brannick, Bryant and Silliman2005). Only if measures are formally defined and if they can be reliably applied, can the outcomes for multiple participants in a study be compared. Moreover, these formal requirements are a condition sine qua none for comparing the results of different studies that claim to measure the same phenomenon. In this paper we will address these questions with regard to measures that have been proposed over the years to compute the age at onset of babbling on the basis of samples of spontaneous prelexical vocalizations.

There is general agreement in the literature that for typically developing children the onset of babbling occurs before age 0 ; 11 (e.g. Koopmans-van Beinum & van der Stelt, Reference Koopmans-van Beinum, van der Stelt, Lindblom and Zetterstrom1986; Nathani, Ertmer & Stark, Reference Nathani, Ertmer and Stark2006; Oller, Reference Oller, Yeni-Komshian, Kavanagh and Ferguson1980; Roug, Landberg & Lundberg, Reference Roug, Landberg and Lundberg1989; Stark, Reference Stark, Yeni-Komshian, Kavanagh and Ferguson1980). A delayed onset of babbling is even considered a predictor of later speech and language problems (Oller, Eilers, Neal & Schwartz, Reference Oller, Eilers, Neal and Schwartz1999). For instance, hearing-impaired infants start babbling considerably later than typically developing infants and they are frequently found to have a deviant speech and language development (Koopmans-van Beinum, Clement & van den Dikkenberg-Pot, Reference Koopmans-van Beinum, Clement and van den Dikkenberg-Pot2001; Oller & Eilers, Reference Oller and Eilers1988; Oller, Eilers, Bull & Carney, Reference Oller, Eilers, Bull and Carney1985; Stoel-Gammon & Otomo, Reference Stoel-Gammon and Otomo1986). The age at which infants start babbling is used in research and in clinical practice as a very early diagnostic marker of speech and language development. Care should therefore be taken regarding the reliability of sampling procedures with which babbling onset measures are applied, and regarding the comparability of babbling onset results with different measures for the same child.

In the literature several procedures have been proposed for identifying the onset of babbling in samples of spontaneously produced prelexical vocalizations. Perhaps the most well known is the one of Oller and Eilers (Reference Oller and Eilers1988). They view the onset of babbling as the emergence of mature-sounding or canonical syllables in the infant's vocal output, i.e. syllables with a fully resonant nucleus (a vowel) and at least one consonantal, non-glottal margin (Oller, Reference Oller2000). Babbling onset is credited if the percentage of canonical syllables computed over all utterances in the entire sample equals or exceeds 20%. Oller and Eilers (Reference Oller and Eilers1988) baptized this measure the ‘Canonical Babbling Ratio’ (henceforth: CBR^utt), and define it as in (1).

(1)

$\eqalign{\tab {\rm CBR}^{{\rm utt}} \equals {{{\rm \ }\# {\rm of\ canonical\ syllables}} \over {{\rm Total\ \# \ of\ utterances}}}$

Oller and Eilers (Reference Oller and Eilers1988: 444) report that for the identification of babbling onset for laboratory purposes a minimal sample size of 50 utterances ‘is commonly sought’, excluding vegetative or reflexive utterances from the sample. This sample size was deemed adequate for the calculation of CBR^utt by Rvachew, Creigthon, Feldman and Sauve (Reference Rvachew, Creighton, Feldman and Sauve2001), who found high correlations between CBR^utt values calculated on the first 50 utterances and calculated on 100 consecutive utterances of the same speech sample. For the 21 typically developing, normally hearing subjects in Oller and Eilers' study the age at onset of canonical babbling lay between 0 ; 6 and 0 ; 10 (mean: 7·6 months), while nine deaf subjects failed to start babbling before 0 ; 11 (range 0 ; 11−2 ; 1).

In the course of the 1990s the formula for the canonical babbling ratio was slightly changed (e.g. Oller, Eilers, Steffens, Lynch & Urbano, Reference Oller, Eilers, Steffens, Lynch and Urbano1994). Instead of computing the ratio of syllables over utterances, the proposal was to compute the ratio of the number of canonical syllables over the total number of syllables in the sample, giving the CBR^syl score defined in (2).

(2)

$\eqalign{\tab {\rm CBR}^{{\rm syl\ }} \equals {{{\rm \ }\# {\rm of\ canonical\ syllables}} \over {{\rm Total\ }\# {\rm of\ syllables}}}$

This change of the procedure ascertained equal potential ranges for the nominator and the denominator in the equation. Values obtained for CBR^syl are slightly lower than those of CBR^utt, and therefore the critical value indicating babbling onset was lowered to 0·15 for CBR^syl. Lynch, Oller, Steffens, Levine, Basinger and Umbel (Reference Lynch, Oller, Steffens, Levine, Basinger and Umbel1995) used this CBR^syl criterion for evaluating babbling onset in thirteen infants with Down Syndrome as compared to twenty-seven typically developing children. Results based on samples of at least seventy vocalizations − again excluding vegetative and reflexive sounds − showed that babbling onset for typically developing children occurred between 0 ; 6 and 1 ; 0 (mean: 8·1 months), while the age range for Down Syndrome children was between 0 ; 6 and 1 ; 2 (mean: 10 months).

In the original conception of the CBR, only non-glottal consonants were considered as margins of a canonical syllable. Inspired by Stoel-Gammon (Reference Stoel-Gammon1989), Chapman, Hardin-Jones, Schulte and Halter (Reference Chapman, Hardin-Jones, Schulte and Halter2001) narrowed down the definition of a canonical syllable to include only syllables containing combinations of vowels with true consonants. True consonants are, according to Stoel-Gammon (Reference Stoel-Gammon1989), all consonants except glottals and glides. Apart from excluding syllables containing exclusively glottal consonants, those containing glides are also not counted in the computation of what can be called the True Canonical Babbling Ratio (TCBR). Analogous to the CBR, the TCBR can be computed with the total number of utterances (TCBR^utt, as in (3)) or the number of syllables (TCBR^syl, as in (4)) in the denominator.

(3)

$\eqalign{\tab {\rm TCBR}^{{\rm utt\ }} \equals {{{\rm \ }\# {\rm of\ true\ canonical\ syllables}} \over {{\rm Total\ }\# {\rm of\ utterances}}}$

(4)

$\eqalign{\tab {\rm TCBR}^{{\rm syl\ }} \equals {{{\rm \ }\# {\rm of\ true\ canonical\ syllables}} \over {{\rm Total\ }\# {\rm of\ syllables}}}$

A ratio of 0·20 or greater for TCBR^utt is considered as the threshold for the canonical babbling stage. In practice, studies applying this criterion have worked with the TCBR^syl, as in (4), with the cut-off point for babbling onset at a ratio of 0·15 or greater. Chapman et al. (Reference Chapman, Hardin-Jones, Schulte and Halter2001), for example, studied prelexical vocal development at the age of 0 ; 9 in 30 infants with unrepaired cleft palate and 15 age-matched non-cleft controls. They counted the number of infants in each group who had reached a TCBR^syl of 0·15 or greater. Only 14 out of 30 cleft palate infants could be credited with babbling onset, while 14 out of 15 controls without cleft palate were babbling at that age. No sample size restrictions were reported, but the recordings used in the study lasted approximately one hour or until at least 100 utterances had been sampled (p. 1272).

The babbling onset measures introduced so far focus on the production of canonical syllables as the crucial aspect of babbled utterances. Another notable hallmark of babbling is its multisyllabic, often reduplicated nature. Van der Stelt & Koopmans-van Beinum (Reference van der Stelt, Koopmans-van Beinum, Lindblom and Zetterström1986) defined babbling as the production of vocalizations with continuous or interrupted phonation in combination with multiple articulatory movements in the course of one breath unit. In their study they instructed parents to identify the onset of this type of behaviour in their infant's vocal output: ‘The parents were to report the day on which they recognized babbling from their baby for the first time’ (p. 164). Parents of 51 children reported the onset of babbling as occurring between the ages of 0 ; 4 and 0 ; 11 (mean age: 0 ; 7). Likewise, Fagan (Reference Fagan2009: 504) determined the onset of reduplicated babbling ‘by parent report of two vocalizations containing CV syllable repetition and evidence of syllable repetition on audio or videotape’. She found a mean age of reduplicated babbling onset at 7·1 months in the 18 children in her study (range: 4·5–12 months).

A more formal criterion for identifying babbling onset in a fragment of vocal output and based on the multisyllabicity requirement was suggested by Schauwers, Gillis, Daemers, De Beukelaer and Govaerts (Reference Schauwers, Gillis, Daemers, De Beukelaer and Govaerts2004). They collected monthly samples, and took the session in which a child for the first time evidenced two or more prelexical vocalizations with interrupted or uninterrupted phonation and multiple articulatory movements as indicative of babbling onset, provided that the number of babbled utterances did not decrease to less than two in the following two months (p. 266). This criterion for determining the onset of babbling will be called the multisyllabicity criterion (henceforth: MULTI). The multisyllabicity criterion does not only introduce a different angle on the description of children's vocal productions, it also introduces a longitudinal aspect: vocalizations from three consecutive months are to be taken into account instead of those from a single month. Schauwers et al. used selections of 20 relatively ‘voluble’ minutes out of the original recording of at least an hour, without further restrictions on sample size. Babbling onset for the 10 typically developing, normally hearing children in this study was credited between the ages of 0 ; 6 and 0 ; 8 (median: 0 ; 6).

In the studies reviewed, different methodologies were used and divergent criteria for establishing the onset of babbling were applied. Nonetheless, all studies agree on the finding that babbling onset typically occurs in the second half of the first year. This poses a number of interesting problems that will be dealt with in this paper. First of all, does applying the different criteria to the same data lead to identifying the onset of babbling at exactly the same point in time? For instance, Chapman et al. (Reference Chapman, Hardin-Jones, Schulte and Halter2001) used the TCBR^syl criterion and Schauwers et al. (Reference Schauwers, Gillis, Daemers, De Beukelaer and Govaerts2004) used the MULTI criterion. Both studies also report results for babbling onset based on the CBR^syl/CBR^utt criterion. However, they do not provide information about the comparability of the ages at babbling onset for individual children with each of the measures.

A second important issue, which logically precedes the previous one, concerns the size of the sample needed to reliably compute the different measures of babbling onset. Up till now that sample size has not been thoroughly validated for any of the criteria. Rvachew et al. (Reference Rvachew, Creighton, Feldman and Sauve2001) found high correlations among CBR^utt values calculated on a sample of 50 utterances and calculated on a sample of 100 utterances. However, they do indicate that there was a mean difference between the two CBR^utt values of 0·05. With the threshold for babbling onset at a CBR^utt value of 0·20, it remains unclear which sample size suffices to establish reliably whether a child scores under or above that threshold.

Sampling issues should be made a matter of great concern in the evaluation of babbling onset in order to increase the reliability, validity and comparability of results, as in other domains of child language research (Hutchins et al., Reference Hutchins, Brannick, Bryant and Silliman2005; Rowland et al., Reference Rowland, Fletcher, Freudenthal and Behrens2008; Tomasello & Stahl, Reference Tomasello and Stahl2004). However, in the studies reviewed, sample size was not an issue at stake: either an arbitrary number of utterances or an arbitrary number of minutes of recording was taken as the yardstick. But whether this yardstick led to a reliable measurement of the onset of babbling was not assessed. Moreover, not even all studies have ascertained that sample sizes were equal for all children. A comparison of age at babbling onset between children on the basis of unequal sample sizes may not be justified: the chance of finding 2 multisyllabic utterances in the course of a session containing 500 utterances is much larger than finding them in the course of a session containing only 50 utterances. The same holds for the criteria involving the calculation of a ratio of well-formed syllables to utterances or syllables (CBR and TCBR): the size of the denominator influences the results. Two children who produce the same absolute number of canonical syllables in the course of a session can still obtain very dissimilar canonical babbling ratios if that session contained 50 utterances for one child and 500 utterances for the other. It is very important to include equal amounts of data for all children and sessions in the procedure. This avoids differences between children in the result of a certain measure that are caused solely by differences in the volubility of those children. In this paper we will address the sample size issue for each of the five criteria for babbling onset outlined in this ‘Introduction’.

The aim of this paper is to answer two fundamental questions. First of all, we investigate whether a minimal sample size can be established that permits the reliable identification of the age at onset of babbling for each of the measures ((T)CBR^utt/syl and MULTI). This question will be explored in Experiments 1 and 2. A large longitudinal corpus of spontaneous speech from forty infants is used as an empirical database. In the first experiment it is investigated whether for each measure a minimal sample size can be established that permits identifying the onset of babbling reliably by computing the measure only once. In the second experiment, a more lenient method will be proposed that makes use of multiple randomly drawn samples and the computation of an average value. The second research question explores whether the various measures yield comparable results. Do all five measures indicate the same age for the onset of babbling? This question will be answered on the basis of the babbling onset results for each of the measures that were reached with a reliable sample size and procedure.

EXPERIMENT 1

METHOD

Participants

Data were collected from a group of forty typically developing children living in the Dutch-speaking part of Belgium. All of these children were raised in monolingual homes, acquiring the standard variant of Dutch. Typical development was established through parent case history report and the administration of a checklist of the attainment of communicative and motor milestones (largely based on the checklist developed by Kind en Gezin, the Flemish infant welfare centre). Normal language development was monitored by the administration of the Dutch version of the CDI at 1 ; 0, 1 ; 6 and 2 ; 0 (N-CDI; Zink & Lejaegere, Reference Zink and Lejaegere2002).

Data collection and transcription

Monthly recordings were made in the children's home environments of spontaneous interactions between the children and their caretakers. The children were followed longitudinally from the age of 0 ; 6 onwards up to 2 ; 0. The recordings were made using a JVC digital video camera, and each recording lasted 60–90 minutes.

From each recording a selection of approximately 20 minutes was made by a member of the research team. The aim was to select uninterrupted stretches of delineated interactions in which the child was vocally active. These selections were transcribed according to the CHAT conventions (MacWhinney, Reference MacWhinney2000).

The children's lexical utterances were orthographically and phonemically transcribed. The coding of the children's prelexical utterances was done according to the procedure used in Schauwers et al. (Reference Schauwers, Gillis, Daemers, De Beukelaer and Govaerts2004). Non-vegetative, non-reflexive comfort sounds uttered within one breath unit and without a consistent sound−meaning relationship were considered prelexical utterances. Each prelexical utterance was coded on a number of different dimensions (coding tiers), three of which are of particular importance for the present paper:

(a) Each utterance was characterized in terms of a combination of phonation (no phonation, uninterrupted or interrupted phonation) and number of articulatory movements (no articulation or a sustained vowel, one articulation, or two or more articulations), in accordance with the sensorimotor model developed by Koopmans-van Beinum and van der Stelt (Reference Koopmans-van Beinum, van der Stelt, Lindblom and Zetterstrom1986). This coding allows the automatic distinction of babbles from other types of prelexical vocalizations, as required for MULTI.
(b) Each prelexical utterance was coded for its structure as a sequence of consonant- and vowel-like elements. This coding permits the automatic selection of CV syllables from the transcripts.
(c) Each consonant-like element was coded for various articulatory features. This coding permits us to automatically identify true consonants as required for the TCBR measures.

Visual information available from the video images guided decisions on segmental characteristics.

All transcribed prelexical utterances were subsequently syllabified with an automated procedure based on universal principles for syllabification (viz. the Core Syllabification Principle; Clements, Reference Clements, Kingston and Beckman1990). The syllabified materials were used for determining the number of canonical, non-canonical, and the total number of syllables for the calculation of the CBR and TCBR ratios.

Table 1 provides an overview of the data. For each child the numbers of prelexical utterances and syllables are provided in the samples from age 0 ; 6 up to 1 ; 0.

Table 1.

Overview of the numbers of utterances (Utts) and syllables (Syls) in the samples between 0 ; 6 and 1 ; 0 for each of the participants in the study (slashes indicate missing sessions)

Reliability

An extensive part of the large speech corpus underwent a reliability check to assess both between- and within-transcriber agreement. For the scoring of between-transcriber reliability, 10% of the original transcriptions in the database was retranscribed by a second, equally experienced transcriber who remained blind to the original transcription. For the coding of prelexical vocalizations relevant to the present paper, this procedure yielded the following reliability scores: for the decision if utterances contained multiple articulatory movements combined with interrupted or uninterrupted phonation, Cohen's kappa=0·65; for the coding in consonant-like and vowel-like segments, Cohen's kappa=0·58; and for the coding of manner of articulation of consonant-like segments as ‘true’ consonants as opposed to ‘other’ consonants, Cohen's kappa=0·52. All kappa scores for between-transcriber agreement fall within the range of ‘moderate’ agreement, according to the evaluation by Landis and Koch (Reference Landis and Koch1977).

For the assessment of within-transcriber agreement, the original transcriber fully retranscribed 5% of the material after a pause of at least two months. Within-transcriber reliability scores were: for the coding of utterances in terms of phonation and articulation, Cohen's kappa=0·75; for the coding of consonant-like and vowel-like segments, Cohen's kappa=0·76; and for the coding of manner of articulation of consonant-like segments as ‘true’ versus ‘other’ consonants, Cohen's kappa=0·67. All kappa scores for within-transcriber agreement are indicative of ‘substantial’ agreement.

Bootstrap

In order to establish the onset of babbling in a reliable way for all children in the study, a bootstrapping procedure was implemented (Baayen, Reference Baayen2008; Efron, Reference Efron1979). The aim of the procedure was to establish a single sample size that can be used as a general guideline for reliably establishing the onset of babbling. The procedure consists of extensively resampling and is written in pseudo-code in (5).

(5) For each CHILDBabblingOnset:=UNDEFINED;AGE:=0 ; 6;RepeatSampleSize:=0;RepeatSampleSize +=25;For i:=1 to 1,000Select a random sample (RS);ComputeMeasure (e.g. CBR^syl) for RS;If CBR^syl for RS >=0·15 then AboveThreshold +=1;If AboveThreshold >=950/1,000Then BabblingOnset:=AGE;Minimal_Sample_Size:=SampleSizeUntil condition (7) holdsIf BabblingOnset=UNDEFINEDAGE +=0 ; 1Until BabblingOnset=DEFINED

The bootstrapping procedure is used to establish the age at which each child for the first time reaches the babbling threshold, for instance CBR^syl>=0·15 (see (2)). This is accomplished by taking for the first age for which data are available (0 ; 6 in the present study) a random sample of 25 syllables from the transcript and computing the CBR^syl for that sample. The random sampling (after replacement) of 25 syllables is repeated 1,000 times, and each time CBR^syl is computed, so that after 1,000 iterations the proportion of samples that are equal to or above the threshold is established. If this threshold is not reached in 95% of the cases with samples of 25 syllables, the sample size is incremented with 25 syllables, and the procedure is repeated for 50 syllables, and so on. The age at which the child starts babbling is reached when in 95% (or more) of the samples the metric at least equals the preset threshold. If the data for a particular age (month) do not permit to establish the onset of babbling reliably, i.e. for no sample size a confidence level of 95% can be reached, the data for the next month are entered into the procedure.

The initial sample size of 25 items (syllables or utterances, depending on the measure) was set arbitrarily, as well as the increment of 25 items. The maximum sample size, i.e. the ceiling of the sampling procedure, was determined in a principled way: the number of combinations of size k from a set of size n was computed according to equation (6), the binomial coefficient:

(6)

$\eqalign {C\lpar n\comma k\rpar \tab \equals {{n\exclam } \over {k\exclam \lpar n \minus k\rpar \exclam }}$

This means, for instance, that for the smallest session (subject P29 at 0 ; 10: 79 utterances; see Table 1), a sample size of 75 items still leaves 1,502,501 unique (though (partially) overlapping) combinations to select from. Each of the 1,000 iterations can thus in principle be run with a different sample of 75 items. Hence, the sample size was increased under the conditions in (7):

(7) k <n and C(n,k) <r
with k=the sample size, n=the number of utterances/syllables (depending on the measure) in a session, and r=the number of runs (i.c. 1,000)

Thus for subject P29 mentioned above, samples sizes of 25, 50 and 75 were entered in the bootstrapping procedure: n=79 and k=75, so that k <n.

In each run of the bootstrap, the age at babbling onset is computed according to each of five criteria:

(a) Canonical Babbling Ratio >0·20 (CBR^utt): babbling onset is defined as the first age at which in at least 950 out of 1,000 runs of the bootstrapping procedure the number of canonical syllables (syllables containing one vowel-like element and at least one non-glottal consonant) divided by the total number of utterances equals or exceeds 0·20 (see (1));
(b) Canonical Babbling Ratio >0·15 (CBR^syl): babbling onset is defined as the first age at which in at least 950 out of 1,000 runs the number of canonical syllables (syllables containing one vowel-like element and at least one non-glottal consonant) divided by the total number of syllables equals or exceeds 0·15 (see (2));
(c) True Canonical Babbling Ratio >0·20 (TCBR^utt): babbling onset is defined as the first age at which in at least 950 out of 1,000 runs the number of canonical syllables (syllables containing one vowel-like element and at least one non-glottal, non-glide consonant) divided by the total number of utterances equals or exceeds 0·20 (see (3));
(d) True Canonical Babbling Ratio >0·15 (TCBR^syl): babbling onset is defined as the first age at which in at least 950 out of 1,000 runs the number of canonical syllables (syllables containing one vowel-like element and at least one non-glottal, non-glide consonant) divided by the total number of syllables equals or exceeds 0·15 (see (4));
(e) Multisyllabicity Requirement (MULTI): babbling onset is defined as the first age at which in at least 950 runs out of 1,000 two or more utterances combining continuous or interrupted phonation with multiple articulatory movements are present, provided the number of such babbled utterances does not decrease to less than two in samples from the two following months.

RESULTS

Repeatedly computing the measures

In the literature the various measures for the identification of babbling onset ((T)CBR^utt/syl and MULTI) are computed on one single speech sample for each age. For instance, Oller and Eilers (Reference Oller and Eilers1988) propose a sample size of 50 utterances to compute CBR^utt. In Table 2 the results are shown of computing CBR^utt 1,000 times using randomly drawn samples of 50 utterances. For this simulation the data of child P40 are used from 0 ; 6–1 ; 0.

Table 2.

CBR^utt computed on 1,000 randomly drawn samples of 50 utterances from 0 ; 6–1 ; 0 of child P40

The results in Table 2 show that the mean value of CBR^utt at 0 ; 9 is larger than 0·20 (the threshold for babbling onset). However, at that age the CBR^utt values range from 0·06 to 0·60, which means that if only one sample had been taken to determine the onset of babbling, CBR^utt could have pointed in either direction: the child could or could not have been credited with onset of babbling. Depending on the hazards of selecting the speech sample, the child could have started babbling at age 0 ; 8 – the age at which the maximum value of CBR^utt exceeds 0·20 for the first time − or the onset of babbling could not have been reached at 1 ; 0: the minimum values up to age 1 ; 0 are all below 0·20. Similar results hold for the other babbling measures. Hence it does not suffice to compute a measure on a single isolated sample in order to arrive at a reliable decision as to the onset of babbling. In other words, it is necessary to repeatedly compute the various measures so that the variance in the data can be estimated.

Results of bootstrapping

The results of the bootstrapping procedure using the algorithm outlined in (5) are displayed in Table 3. For each measure, Table 3 shows the median age at onset of babbling and the sample size needed to reach 95% confidence in the result. The number of items (syllables or utterances) required for reliably establishing babbling onset depends on the child and on the particular measure. The amounts range from a minimum of 25 to 300 utterances for TCBR^utt, and maximally 350 utterances for MULTI, 375 utterances for CBR^utt, 425 syllables for CBR^syl and 500 syllables for TCBR^syl. This means that, based on the results for the forty children in our study group, for TCBR^syl samples of at least 500 syllables are needed in order to decide with 95% confidence that a child has started babbling using only a single computation of the measure.Footnote ¹

Table 3.

Overview of the ages at onset of babbling and the required minimal sample sizes using bootstrapping procedure (5)

Effect of sample size

In Figure 1, the results of the bootstrapping procedure outlined in (5) are illustrated for the computation of TCBR^syl with sample sizes increasing from 25 syllables up to 600 syllables for child P36 from 0 ; 6 up to the onset of babbling.

Fig. 1.

Proportion of samples at or above the critical threshold of TCBR^syl=0·15 for sample sizes 25–600 syllables for child P36.

The graph shows that at 0 ; 6 a sample size of 25 utterances yields only 93 out of 1,000 TCBR^syl values at or above the critical threshold for babbling onset which was set at 0·15. That number decreases with an increase of the sample size. The child can thus not be credited with babbling onset at age 0 ; 6. At 0 ; 7 and 0 ; 8 the threshold for babbling onset is not reached in enough samples either. At 0 ; 8, for instance, there are 309 samples of 25 syllables reaching the threshold. That number decreases to 173 for 50 syllables, 93 for 75 syllables and further down to 0 from 300 syllables onwards. At 0 ; 9, 593 out of 1,000 samples of 25 syllables are at or above TCBR^syl=0·15. However, with increasing sample size the 95% confidence level is eventually reached, viz. when samples of 500 syllables are used. Hence in the case of P36 the onset of babbling according to the TCBR^syl criterion can be set at age 0 ; 9 (for a sample size of 500 syllables).

As Figure 1 illustrates, by applying the measure on samples of increasing size, a curve can be drawn that exhibits a clear slope. This slope indicates unambiguously whether the measure is coming closer to and will eventually reach the threshold or not. But a sufficient number of data points is needed in order to fit the data well: at 0 ; 9, the first three data points show decreasing values (a negative slope), while eventually, the slope of a linear fit is positive. Suppose the data for P36 at 0 ; 9 only permitted samples of 75 syllables. Less than 95% of the samples would have reached criterion, and onset of babbling would not have been credited. Or, if in that case a regression was applied, the direction of the slope would have been negative, with the same result: no babbling onset at 0 ; 9.

The effect of the size of the sample is illustrated more generally in Figure 2, which depicts for TCBR^utt per age the cumulative number of children that start babbling. For reasons of clarity, the graph is restricted to samples of 25 to 400 utterances. At 0 ; 6, 20% of the children have started babbling, when samples of 25 utterances are considered. That number increases to 23% when samples of 75 or 100 utterances are drawn, and it further increases until samples of 200 utterances are drawn. After that point, no further increase is noted, so that it can be concluded that for TCBR^utt at age 0 ; 6 at least 200 utterances are needed in order to reliably compute the onset of babbling. At 0 ; 7 33% of the children reach the babbling threshold with samples of 25 utterances, but that number increases to over 50% by the time samples of 200 utterances are used in the procedure. At 0 ; 8 it even takes samples of 300 utterances to reach the maximal proportion of children who can be considered to babble according to TCBR^utt. Thus, we can conclude that for TCBR^utt at least 300 syllables are needed in order to compute the measure reliably at 0 ; 8. In more general terms: the larger the sample, the more children reach the critical threshold that marks onset of babbling at any given age. This was true for each of the five measures under investigation.

Fig. 2.

Cumulative proportion of children reaching the babbling onset threshold for TCBR^utt.

For none of the measures was a significant correlation found between the age at onset of babbling and the required sample size (Spearman's rho: CBR^syl: rho=0·1, p=0·33; CBR^utt: rho=0·26, p=0·10; TCBR^syl: rho=−0·16, p=0·32; TCBR^utt: rho=−0·002, p=0·99; MULTI: rho=−0·20, p=0·23). Hence, there is no clear-cut relationship between the age at onset of babbling and the size of the sample required to compute it.

The results in Figure 2 imply that an equal amount of data (syllables or utterances) should be considered for each subject and each age. The bootstrap procedure shows that otherwise this may lead to the attribution of babbling onset to a particular child and not to another child solely because of a difference in the amount of data available. Both could have been credited for babbling if an equal amount of data had been considered.

Discussion

The aim of the first experiment was to establish a minimally required sample size for reliably computing the onset of babbling in one single pass given a relatively limited recording of a sample of spontaneous speech. In other words, the question addressed was: How large should a sample be so that the computation of (T)CBR^syl/utt or MULTI yields a reliable indication of whether a child has started babbling or not?

A first finding was that in order to establish the minimal sample size the measures should be applied repeatedly on the data so that the variance can be measured. A bootstrapping experiment was designed implementing a specific algorithm for repeatedly measuring (T)CBR^syl/utt and MULTI for increasing sample sizes. The main result of the experiment was that if a single computation of the measures is aimed at, and a confidence of 95%, reasonably large samples are required: at least 300 utterances for TCBR^utt and at least 500 syllables for TCBR^syl, for example (see Table 3). In order to put these findings into practice, these sample sizes are required for all children, otherwise erroneous conclusions with respect to babbling onset can be drawn in individual cases.

This poses a serious problem: inspection of Table 1 reveals that, especially at the younger ages, such amounts of utterances/syllables are not available in the transcriptions. Are there solutions for this impasse?

A first solution is rather pragmatic in nature: corpora such as the one used in this study are rather limited. That is, each recording consists of a limited number of data (utterances, syllables). If an equal number of utterances or syllables is required for further analysis, the size of the samples can be determined by the recording with the least data. A quick overview of Table 1 reveals that the smallest number of utterances and syllables is produced by P19 at 0 ; 10: 79 utterances and 127 syllables. Thus the smallest sample can be used as benchmark, taking into account the fact that in so doing the number of children with babbling onset in a particular month will be (vastly) underestimated.

A second solution will be explored in Experiment 2: the quest for finding a minimal sample size for which a single computation of the five measures suffices is abandoned. Instead, a ‘lenient’ bootstrapping procedure – as opposed to the strict version in (5) – is introduced, which permits establishing a minimal sample size that is sufficient to identify the onset of babbling given that the (T)CBR^syl/utt and the MULTI measures are computed in 1,000 runs in a bootstrapping procedure.