Everyday language input and production in 1001 children from 6 continents

DRAFT Everyday language input and production in 1001 children from 6 continents Elika Bergelsona,1, Melanie Soderstromb, Iris-Corinna Schwarzc,n, Caroline F. Rowlandd,e,f, Nairán Ramírez-Esparzag, Lisa Rague Hamrickh, Ellen Marklundc, Marina Kalashnikovai,o, Ava Guezj, Marisa Casillasd,f,k, Lucia Benettil, Petra van Alphenm, and Alejandrina Cristiaj,1 a Harvard University, Department of Psychology; b University of Manitoba, Department of Psychology; c Stockholm University, Department of Linguistics; d Max Planck Institute for Psycholinguistics, Language Development Department; eRadboud University, Donders Centre for Brain, Cognition and Behaviour; fARC Centre of Excellence for the Dynamics of Language (CoEDL); gUniversity of Connecticut, Psychological Sciences; hPurdue University, Department of Psychological Sciences; iBasque Center on Cognition Brain and Language (BCBL); j PSL University, Laboratoire de Sciences Cognitives et de Psycholinguistique, Département d’études cognitives, ENS, EHESS, CNRS; k University of Chicago, Comparative Human Development Department; l Ohio State University, School of Music; m Royal Dutch Kentalis; n Stockholm University, Department of Special Education; oIkerbasque, Basque Foundation of Science This manuscript was compiled on August 23, 2023 Language is a universal human ability, acquired readily by young children, who otherwise struggle with many basics of survival. And yet, language ability is variable across individuals. Naturalistic and experimental observations suggest that children’s linguistic skills vary with factors like socioeconomic status and children’s gender. But which factors really influence children’s day-to-day language use? Here we leverage speech technology in a big-data approach to report on a unique cross-cultural and diverse data set: >2,500 day-long, child-centered audio-recordings of 1,001 2to 48-month-olds from 12 countries spanning 6 continents across urban, farmer-forager, and subsistence-farming contexts. As expected, age and languagerelevant clinical risks and diagnoses predicted how much speech (and speech-like vocalization) children produced. Critically, so too did adult talk in children’s environments: Children who heard more talk from adults produced more speech. In contrast to previous conclusions based on more limited sampling methods and a different set of language proxies, socioeconomic status (operationalized as maternal education) was not significantly associated with children’s productions over the first four years of life, and neither were gender or multilingualism. These findings from large-scale naturalistic data advance our understanding of which factors are robust predictors of variability in the speech behaviors of young learners in a wide range of everyday contexts. 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 infancy | human diversity | language | socioeconomic status | speech Typically-developing children readily progress from coos to 1 complex sentences within just a few years, leading some to 2 hypothesize that the universal language abilities of humans 3 develop uniformly, with only incidental effects of individualor 4 group-level variation (1). And yet, studies using a variety of 5 proxies for language development find some evidence of such 6 variation in early language skills, with differences reported 7 between girls and boys (2), as well as those raised in socioeco8 nomically privileged compared to disadvantaged households 9 (3, 4).10 However interesting, these studies tend to rely on Western11 centric samples and methods, and may not reflect everyday 12 language use in children. Moreover, prior work often stops 13 after only considering individual predictors in a binary way 14 (i.e. do they significantly impact language development or 15 not), while failing to ask the more informative question of how 16 large their relative impact is (5), especially in freely-occurring, 17 everyday speech behavior.18 Recent research on mice and whales shows the promise of 19 machine learning for examining everyday animal behavior (6, 20 7). We leverage advances in wearables and machine-learning21 based speech technology to catalyze a similar breakthrough in 22 language development research. Our dataset is comprised of 23 >40,000 hours of audio from >2,500 days in the lives of 1,001 24 2to 48-month-olds from 6 continents and diverse cultural 25 contexts (Figure 1). Within this dataset, we focused on the 26 amount of speech or speech-like vocalization young children 27 produce in their everyday life. Critically, these automatically28 extractable “quantity” measures correlate robustly with gold29 standard “quality” measures of children’s language skills and 30 knowledge, like vocabulary estimates (see SI1D for relevant 31 evidence) (4). 32 We query and compare the effects of two types of factors. 33 First, there are factors with undeniable effects on early lan34 Significance Statement Harnessing a global sample of >40,000 hours of child-centered audio capturing young children’s home environment, we measured contributors to how much speech 0–4 year olds naturally produce. Amount of adult talk, age, and normative development were the sole significant predictors; child gender, socioeconomic status, and multilingualism did not explain how often children vocalized, or how much adult talk they heard. These findings (strengthened by our validation of existing automated speech algorithms) open up new conversations regarding early language development to the broader public, including parents, clinicians, educators, and policymakers. The factors explaining variance also inform our understanding of humans’ unique capacity for learning, and potentially large-scale applications of machine technology to everyday human behavior. EB, MC, and AC developed the initial conceptualization of the project and recruited corpus owners and co-authors. EB, MC, and AC curated the meta-corpus and meta-data and prepared them for analysis. EB and AC prepared materials for and/or led group decision-making. EB, MS, CR, NRE, AG, MC, LB, PvA, and AC contributed to the decision-making on the analytic approach, including selection of exploratory and confirmatory sets, selection of variables, identification of hypotheses and/or specification of models. AC, EB, and AG drafted the preregistrations. AC, EB, and AG designed and implemented the analyses. EB, CR, NRE, LRH, MK, and LB conducted and synthesized literature reviews on key topics related to the decision-making regarding literature review, hypotheses, and analyses. EB, EM, ICS, CR, LRH, MS, NRE, MK, MC, PvA, and AC provided corpus data and meta-data. See acknowledgments for non-author data contributors. EB, AC, and MS contributed to the initial manuscript draft writing. EB, MK, MC, and AC contributed to visualizations. EB, MC, and AC revised and responded to feedback and informal peer-review. MS, CR, LRH, LB, EB, AC, EM, ICS, and PvA contributed to supplementary materials, Open Science Framework project page and/or other documentation. Note: Other than first and last authors, middle authors are listed in reverse alphabetical order. The authors have no conflict of interest to declare. 1 To whom correspondence should be addressed. E-mail: [email protected], [email protected] www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX PNAS | August 23, 2023 | vol. XXX | no. XX | 1–12 DRAFT guage production, namely, child age and language-relevant 35 clinical risks and diagnoses. Second, there are individualand 36 family-level factors that are reported to correlate with vari37 ability in early language skills: socioeconomic status (SES; 38 operationalized here as maternal education; SI2B), gender, 39 language input quantity, and multilingual background. Be40 cause small and homogeneous samples make universal claims 41 more questionable, a key novel contribution of this work is its 42 benchmarking of the level of stability and variability of every43 day language use in a heterogeneous, richly diverse participant 44 sample.∗ 45 Measuring Diverse, Real-life Language Use. Language skills 46 and knowledge are not directly observable. As a result, all 47 studies use a proxy when estimating them in individual chil48 dren. These proxies have variable validity and predictive power 49 relative to other measures, both concurrently and predictively, 50 and likely vary in the extent to which they reflect children’s 51 everyday language behavior. For instance, parental report 52 measures are indirect and—especially for receptive knowledge— 53 can be difficult for caretakers to estimate (9), even in relatively 54 homogeneous Western-centric contexts.55 Here, we adopt a very different approach. We employed 56 the LENA™system, which captures what children hear and 57 say across an entire day through small wearable recorders 58 (10); this ecologically-valid sampling method reduces observer 59 effects relative to, e.g., shorter video recordings (11). The 60 LENA™system uses standardized algorithms that estimate 61 who is speaking when, alongside automated counts of adult and 62 child linguistic vocalizations (4) (see definition and validation 63 in SI1C:E). The resulting LENA™measures correlate with 64 and predict other measures of language skills in children with 65 and without clinical risks or diagnoses, as revealed by manual 66 transcriptions, clinical instruments, and parent questionnaires 67 (12, 13). We use LENA™’s validated, automated estimates 68 to derive our measures of everyday language use: adult talk 69 and child speech (see detailed motivation in SI3B). We define 70 child speech as the quantity of children’s speech-related vo71 calizations (e.g., protophones (14), babbles, syllables, words, 72 or sentences, but not laughing or crying) per hour, and adult 73 talk as the number of near and clear vocalizations per hour 74 attributed to adults (both as detected by LENA™’s algorithm; 75 see Methods). Assuaging concerns that these measures are 76 merely capturing chattiness or repetition, both have a ≥. 7 77 correlation with measures of lexical diversity and language 78 “quality”: our child speech measure correlates with vocabulary 79 in an independent sample, and the adult talk measure corre80 lates with the number of word types from manual transcription 81 in a subset of the data (SI1D).82 Capitalizing on this standardized and deidentified numeric 83 output, we solicited LENA™datasets that researchers had 84 previously collected to study monoand multilingual children 85 (i.e. those learning >1 language) in urban, farmer-forager, 86 and subsistence-farming contexts worldwide (Figure 1). This 87 resulted in a dataset reflecting the state of current knowledge 88 in ecologically-valid speech samples from children’s daily lives 89 (SI3A; see Methods for more sample details).90 ∗ While these data collectively span living circumstances, geography, and family structure, some data donors were concerned that highlighting differences when minoritized communities are involved poses ethical challenges, in terms of honorable representation and potential harm. Individual data stewards are actively engaging in richer descriptions of included samples (see SI5), which may enable future work on meaningful population-level differences (e.g., 8). The dataset includes children from wide-ranging SES back91 grounds, based on maternal education levels spanning from no 92 formal education to advanced degrees (SI2B). This SES proxy 93 was selected not only because it was available in all 18 corpora 94 (only 3 had alternative SES proxies), but most importantly 95 because it is the most commonly employed SES proxy in lan96 guage acquisition research, as established in meta-analyses (15, 97 16). This allows our findings to inform ongoing discussions. 98 Theories of how SES relates to children’s language development 99 have proposed a wide range of pathways in which maternal 100 education is predictive of children’s language experiences, in101 cluding the connection between maternal education and the 102 tendency to employ verbal over physical responsiveness (17), 103 the diversity in mothers’ vocabulary (18), and the frequency of 104 verbally-rich activities (19). Maternal education also correlates 105 highly with other SES proxies (e.g. r=.86 in a study of children 106 growing up in 10 European or North American countries, 20), 107 suggesting it may also indirectly pick up on other pathways 108 linking SES to language development, through e.g. differential 109 access to resources and nutrition, or exposure to stress perina110 tally (21). At the same time, we recognize that comparing a 111 variable like education across countries, although commonly 112 done (22), is not straightforward. Therefore, we supplement 113 our pre-registered approach with numerous exploratory checks 114 and analyses examining alternative implementations (SI3G:H 115 described further below). 116 Crucially, by including children aged 2 to 48 months, we 117 span a wide range of linguistic skills, allowing us to better 118 capture the effects of our variables over a broad span of devel119 opment within our socio-culturally and geographically broad120 ranging participants. We also include children with a variety 121 of diagnoses of language delays and disorders, as well as those 122 at high risks for them (see Methods & SI2A for definitions and 123 detailed justification). Such children’s language development 124 is by definition non-normative. Thus, age and non-normative 125 status provide useful yardsticks for considering the significance 126 and effect size of other childand family-level factors (SES 127 through maternal education, child gender, monovs. multilin128 gual status, and how much adults talk to and around the child). 129 That is, if a factor (e.g., gender) has an effect far smaller than 130 that of age or non-normative development, it would suggest 131 that individual differences within it are relatively limited in 132 their connection to everyday language use. If the effects are 133 comparable in size, it would instead suggest that the amount of 134 speech humans produce in everyday interactions is undergirded 135 by substantial and structured individual differences, rather 136 than striking uniformity. Given that effects could vary as a 137 function of child age, we make sure to include key interaction 138 terms. For instance, we can expect age to interact with adult 139 talk if (as anticipated) older children are more sensitive to 140 adults’ talking to them than younger ones. 141 Predicting Children’s Speech Production. We employed a 142 hypothesis-testing approach: In a two-step preregistration, 143 we first established exploration and confirmation data subsets 144 (see Methods and SI3A for detailed explanation, and SI3D:E 145 for the procedure used to derive pre-registered hypotheses 146 and analyses). We then leveraged the held-out confirmation 147 subset to answer our key question: How well do specific 148 individualand family-level factors predict variation 149 in how much speech young children produce? At stake 150 in these analyses is whether systematic differences in children’s 151 2| www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX Bergelson et al. DRAFT Fig. 1. Geographical location, primary language, number of children (Nchild), number of recordings (Nrec) and data citation for each corpus. Table 1. Model results predicting child speech. q-values show FDR-corrected p-values. βSE q Intercept 0.109 0.128 .681 Child Gender(Male) 0.026 0.051 .852 SES(<H.S.(1)) 0.001 0.111 .991 SES(H.S.(2)) -0.033 0.115 .932 SES(B.A.(4)) -0.064 0.079 .681 SES(>B.A.(5) -0.002 0.090 .991 Control -0.085 0.029 .035 * Norm -0.220 0.087 .036 * Adult Talk 0.260 0.037 <.001 * Age 0.647 0.024 <.001 * Mono 0.045 0.095 .852 Norm × Adult Talk -0.005 0.063 .991 Norm × Age -0.217 0.051 <.001 * Adult Talk × Age 0.125 0.022 <.001 * Adult Talk × Mono 0.092 0.072 .45 Mono × Age -0.048 0.056 .681 Norm × Adult Talk × Age 0.019 0.043 .852 Mono × Adult Talk × Age 0.137 0.065 .094 Note. Betas show deviation from the following baseline levels: Child Gender: female; SES: some university(3); Norm: Norm(ative development); Mono: Mono(lingual). SES = child SES based on maternal education (<H.S.(1) = less than high school, H.S.(2) = high school, B.A.(4) = college degree, >B.A.(5) = advanced degree); Control = overlap rate control; Adult Talk = adult vocalization count rate. lives have measurable links to their language production, and 152 if so, what the strength of these relationships is both overall, 153 and in relation to one another (see Table 1for results†). 154 As expected, we found that older children produced more 155 speech than younger ones (ß=0.647, SE=0.024). Children 156 with non-normative development produced less speech than 157 children with normative development (ß=-0.22, SE=0.087) ‡ , 158 an effect that strengthened with age (ß=-0.217, SE=0.051; see 159 Figure 2B). This is expected because for some groups in our 160 non-normative subset (e.g. those with familial risk of a speech 161 impairment) older children are more likely to have an actual 162 diagnosis (as opposed to risk factor) than younger ones (see 163 SI2A for details on non-normative classification). 164 Our results further revealed that young children’s speech 165 production correlated with the amount of adult talk they heard 166 (ß=0.26, SE=0.037). This correlation strengthened with age 167 (ß=0.125, SE=0.022; see Figure 2A), perhaps because variation 168 in adult talk rate has less effect on infants (whose early babbles 169 occur frequently even when infants are alone, 14). The effect 170 of adult talk is a substantial one. Taking the effects of age and 171 normativity as convenient (but unrelated) gauges for what 172 counts as a considerable effect, we see that the effect size of 173 adult talk is about a third of that for age and similar to that 174 for normativity (adult talk: 0.26; interaction adult talk by 175 age: 0.125; age: 0.647; non-normative development: -0.22; 176 interaction non-normative by age: -0.217; all effect size betas 177 expressed as SDs). 178 To provide these results in more intuitive units, we fit the 179 same model centering variables without scaling. Children 180 † All ßs in Tables and text are based on treatment-coded models. See SI3H for sum-coded models, which give the same pattern of results. ‡The normativity estimate is negative because normative development is the baseline. Bergelson et al. PNAS | August 23, 2023 | vol. XXX | no. XX | 3 DRAFT produced 66 more vocalizations per hour with each year of life. 181 For every 100 adult vocalizations per hour, children produced 182 27 more vocalizations; this effect grew by 16 vocalizations per 183 year. Relative to infants with typical development, those with 184 non-normative development produced 20 fewer vocalizations 185 per hour; this difference grew by 8 vocalizations per year.186 Surprisingly, and in contrast to previous results using 187 smaller and less diverse datasets and/or other language proxies, 188 we found that child gender, SES (indexed here by maternal 189 education), and monolingual status did not explain signifi190 cant variation in child speech. As our raw data figures and 191 model outcome results show, these null effects hold both when 192 considering covariates (as in our model; Table 1) and when 193 considering these variables individually (as in Figure 3; SI3F, 194 3G, 3H). In our full model controlling for other variables (Ta195 ble 1), the largest estimate for main effects or interactions 196 involving child gender, SES, and monolingual status was about 197 half of that for normativity, and one-sixth of that for age; none 198 reached thresholds for statistical significance.199 While our models are well-powered to estimate associations 200 of child speech with age, normativity, adult talk, gender, SES 201 (as measured by maternal education), and monolingual status, 202 this is predicated upon pooling the data and accounting statis203 tically for corpusand child-level variance via random effects, 204 as described in Methods. This makes it beyond this paper’s 205 scope to analyze language or population/cultural differences 206 in detail, i.e. in a way that might allow the consideration 207 of additional, culture-specific variables (hence their omission 208 in Figs 2–3); see SI5 for citations to research on individual 209 datasets, some of which tackle such differences directly.210 Noting that the results above have the strongest inferential 211 value thanks to being pre-registered, we also addressed certain 212 alternative hypotheses and interpretations that could have ren213 dered our conclusions unjustified through a series of follow-up 214 analyses. These checked for robustness of our key results with 215 different operationalizations and statistical implementations of 216 SES, when considering only children under or over 18 months, 217 when considering causal paths, and when incorporating speech 218 from other children as a predictor; our key results held in all 219 cases (SI3H).220 We highlight here the results that may run most counter to 221 many readers’ assumptions, namely, that in this large sample, 222 SES (indexed by maternal education) does not come out as 223 a significant predictor of child speech. This conclusion held 224 when declaring SES as an ordinal and as a continuous variable 225 based on levels or years of maternal education, when binarizing 226 SES levels based on individual countries’ average education 227 completion rate, and when declaring a random slope for SES 228 within corpus (which allows SES effects to vary across corpora). 229 Some readers may wonder whether there were some corpora 230 for which SES did matter. If so, the analysis with random 231 SES slopes by corpus would have indicated this, but it did not 232 (SI3H). The relationship between SES and child speech was 233 weak and inconsistent across corpora (as evident in Fig. 4).234 Perhaps most convincingly, results also held when constrain235 ing our analysis to our largest homogeneous subset, the North 236 American subsample (642 daylong recordings from 206 infants 237 in 7 corpora; SI3G). We essentially replicated the full-sample 238 results in this subsample: adult talk and age were significant239 predictors, whereas gender and SES (based on maternal edu240 cation) were not. The significant adult talk × age interaction 241 also replicated. The main effect of normativity did not, likely 242 because normativity’s interaction with age was larger than 243 in the full-sample analysis. Finally, we also tested whether 244 removing the adult talk variable would result in an SES effect, 245 i.e. testing whether adult talk was absorbing variance that 246 would otherwise be accounted for by SES. This was not the 247 case: Removing the adult talk predictor, SES still does not 248 account for significant variance in child speech in our analysis. 249 A central contribution of this work is thus the clear lack of 250 evidence we find for effects of SES (under several operational251 izations focused on maternal education), on how much speech 252 young children produce in day-to-day life. 253 Another potential concern is that our conclusions hinge 254 on the use of LENA™’s particular algorithm; they do not. 255 The findings above successfully replicate in the subset of data 256 for which data stewards were able to share raw audio (11/18 257 corpora), which was analyzed with a wholly different algorith258 mic approach, the Voice Type Classifier or VTC (Methods; 259 SI3F). § Yet another worry is that our focus on adult talk may 260 mask other important contributions to children’s language 261 experiences, for instance, speech from other children. Testing 262 this in a supplemental analysis, we confirm that the level of 263 association found between adult talk and children’s speech 264 was unaffected by including other children’s talk measured by 265 LENA as a predictor variable (SI3H), confirming that our key 266 conclusions hold when factoring this other source of input in. 267 Finally, we also ran a model predicting adult talk (rather 268 than child speech). The amount of adult talk was not sig269 nificantly predicted by SES, child age, gender, monolingual 270 or normative status (Table 2, Figure 3E:H; SI3G:H). Impor271 tantly, these null results replicated in the North American 272 subset (SI3G) as well as in every other alternative analysis we 273 attempted (SI3H). Together, these analyses suggest that the 274 relationship we find between adult talk and child speech in the 275 child speech models is not attributable to childor family-level 276 factors affecting adult talk. 277 Speech and Other Early Vocal Behavior. While our central 278 query concerned variability within early speech production, 279 we conducted a further descriptive analysis examining how 280 much of children’s vocalizations were speech or speech-like, as 281 opposed to the two other classes of LENA™-identified vocal282 izations: crying and vegetative sounds (e.g. burps, hiccups). 283 We examined these vocalization types as a function of age, 284 monolingual status, and normative status. As Figure 2C shows, 285 for children with normative development, the proportion of 286 vocalizations that were speech increased from just over half to 287 the vast majority over 2–48 months. In contrast, the crying 288 proportion fell steeply over the same period, from nearly half 289 of vocalizations to a small fraction of them; the proportion 290 of vegetative sounds was low and constant. Convergent with 291 our speech analyses, monolingual status did not alter these 292 patterns but normative status did: While the same overall 293 patterns held for children with non-normative development, 294 their decrease in crying and increase in speech production with 295 age was less steep (see Figure 2C). 296 As with more narrowly-defined non-normative populations 297 (e.g. children with Autism Spectrum Disorder (23)), we find 298 clear divergences in language trajectories in our normative 299 vs. non-normative samples. This is notable because (a) our 300 §VTC too has been robustly validated relative to various gold standard manual measures (SI1E) 4| www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX Bergelson et al. DRAFT Low−AVCr Mid−AVCr High−AVCr 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0 50 100 150 200 250 300 350 400 450 500 550 600 Age (months) Child ling. voc. rate (per hr.) Corpus 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 A Non−Normative Normative 0 10 20 30 40 50 0 10 20 30 40 50 0 50 100 150 200 250 300 350 400 450 500 550 600 Age (months) Child ling. voc. rate (per hr.) B Speech Cry Vegetative 0 10 20 30 40 50 0 10 20 30 40 50 0 10 20 30 40 50 0.0 0.1 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Age (months) Prop. of vocal behavior Non−monoling. Non−norm. Monoling. & norm. C Fig. 2. Effects of adult talk, child age, and normative development on children's speech production. Points show each daylong recording; lines show linear regression with 95% Confidence Intervals (CI). Child speech is quantified as child linguistic vocalization rate; adult talk as adult vocalization count rate (AVCr). A: Child speech by age, split by low/mid/high tertiles of adult talk. Lines depict significant adult talk × age interaction. Color-shape combinations show each unique corpus, numbered to preserve anonymity. B: Child speech by age and normative status. Lines depict significant age × normative status interaction. C: Proportion of vocal behavior classified as speech, cry, or vegetative, by age. Line type/color indicate monolingual and normative statuses. N.B. Monolingual normative CI (blue) falls fully within that for multilingual children (pink) for all 3 types of vocal behavior, highlighting these groups' equivalent patterns. Table 2. Model results predicting adult talk (i.e. adult vocalization count rate). q-values show FDR-corrected p-values. βSE q Intercept -0.100 0.160 .778 Child Gender(Male) 0.174 0.148 .547 SES(<H.S.(1)) 0.239 0.173 .547 SES(H.S.(2)) -0.015 0.194 .939 SES(B.A.(4)) 0.148 0.131 .547 SES(>B.A.(5) 0.098 0.150 .778 Control 0.084 0.055 .547 Norm 0.013 0.103 .939 Age -0.030 0.029 .547 Mono -0.028 0.112 .939 Gender(Male) × SES(<H.S.(1)) -0.375 0.196 .547 Gender(Male) × SES(H.S.(2)) -0.263 0.252 .547 Gender(Male) × SES(B.A.(4)) -0.220 0.176 .547 Gender(Male) × SES(>B.A.(5)) 0.016 0.201 .939 Norm × Age -0.076 0.060 .547 Mono × Age 0.035 0.068 .804 Note. None of the variables in our model predicted adult talk. All abbreviations and baselines as in Table 1. non-normative sample is heterogeneous (SI2A) and (b) as 2–48301 month-olds, many children with non-normative classifications 302 here were at risk of (but not yet diagnosed with) language 303 delays or disorders. Automated speech analyses in naturalistic 304 recordings thus hold promise for future research into early 305 diagnostics (24, 25). 306 Adult Talk and Child Speech. Children who heard more adult 307 talk produced dramatically higher rates of speech, and this 308 effect increased with age. This result feeds into ongoing theo309 retical debates regarding the relevance of individual differences 310 (26). Although we cannot infer causality from our correlational 311 data, it is useful to consider possible causal paths that could 312 in principle have led to our results. A correlation between 313 child speech and adult talk is compatible with at least three 314 explanations: (1) Children who produce more speech elicit 315 more talk from adults; (2) Language-dense environments lead 316 children to produce more speech; or (3) A third variable causes 317 increases in both adult talk and child speech.¶318 Our model predicting adult talk (see Table 2) can be 319 brought to bear on Explanation 1. If children talking more 320 elicited more talk from adults, then we would have expected to 321 ¶ Our analyses suggest that one such potential third variable, differences in activities across recordings, is not a likely candidate for the correlation between child speech and adult talk (SI4). Bergelson et al. PNAS | August 23, 2023 | vol. XXX | no. XX | 5 DRAFT 0 100 200 300 400 500 600 F M Child Gender Child ling. voc. rate (per hr.) A 0 100 200 300 400 500 600 <H.S. (1) H.S. (2) S.U. (3) B.A. (4) >B.A. (5) SES Child ling. voc. rate (per hr.) B Low−AVCr Mid−AVCr High−AVCr 0 1020304050 0 1020304050 0 1020304050 100 200 300 400 Age (months) Child ling. voc. rate (per hr.) Normative Y N C Low−AVCr Mid−AVCr High−AVCr 0 1020304050 0 1020304050 0 1020304050 100 200 300 400 Age (months) Child ling. voc. rate (per hr.) Monoling. Y N D 0 100 200 300 400 500 600 F M Child Gender Adult voc. count rate (per hr.) E 0 200 400 600 <H.S. (1) H.S. (2) S.U. (3) B.A. (4) >B.A. (5) SES Adult voc. count rate (per hr.) F 170 180 190 200 210 220 0 10 20 30 40 50 Age (months) Adult voc. count rate (per hr.) Normative Y N G 175 200 225 250 0 10 20 30 40 50 Age (months) Adult voc. count rate (per hr.) Monoling. Y N H Fig. 3. Factors that do not predict child speech or adult talk. Points = individual recordings, jittered horizontally. Lines = linear fit with 95% Confidence Intervals. Error bars = 99% bootstrapped CIs of sample means. Child speech is quantified as child linguistic vocalization rate; adult talk as adult vocalization count rate (AVCr). A & B: null effects of child gender (A) and socioeconomic status (SES) (B) on child speech. C: null 3-way effect of normative development × adult talk × age (N.B.: normative × age and adult talk × age are significant; see Fig. 2). D: null 3-way effect of age × adult talk × monolingual status. Eand F: null effects of child gender (E) and SES (F) on adult talk. G & H: null effect of normative development (G) and monolingual status (H) on adult talk. see that age and normative status were significant predictors322 of adult talk. Instead, we find that neither these (nor any 323 other variables in our model) predicted the quantity of adult 324 talk (Figure 3G). Nonetheless, the precise statistical analy325 ses we carried out do not allow us to directly rule out any 326 of the explanations, a combination of which may be jointly 327 true. Establishing a precise causal chain will require careful 328 consideration of a variety of proximal and ultimate pathways 329 through which child and adult behaviors are shaped. As one 330 example, given that most children here are genetically related 331 to their adult caregivers, we may be observing covariance in 332 amount of talk and its linguistic correlates (Explanation 3). 333 Evaluating these alternatives requires evidence from children 334 raised by unrelated caregivers or from genome-wide associ335 ation studies, as genetic and environmental factors remain 336 challenging to disentangle (27). In this vein, recent work 337 with adopted 15–73-month-olds provides evidence for input 338 effects (maternal utterance length and/or lexical diversity) on 339 adopted children’s vocabulary size (measured via caretaker 340 checklist) (28). This study suggests that shared genetics is not 341 the sole contributor to links between (at least these proxies 342 for) caretaker input and child language outcomes. Moreover, 343 shared genetics is just one of the ways in which adult and child 344 behavior may be independently shaped by an unmeasured 345 confounded variable (as per Explanation 3). For instance, 346 other third variables related to dimensions like personality, 347 neighborhood, and childcare context too may be contributors 348 (29, 30). These explanations can only be definitively teased 349 <H.S. (1) H.S. (2) S.U. (3) B.A. (4) >B.A. (5) <H.S. (1) H.S. (2) S.U. (3) B.A. (4) >B.A. (5) <H.S. (1) H.S. (2) S.U. (3) B.A. (4) >B.A. (5) 0 100 200 300 400 0 100 200 300 400 0 100 200 300 400 0 100 200 300 400 0 100 200 300 400 0 100 200 300 400 SES Child ling. voc. rate (per hr.) Fig. 4. Child speech as a function of SES within individual corpora. SES = maternal education levels as in Table 1. White lines = linear fit with 95% CIs in color, color = corpus. Black lines = 99% CIs of sample means bootstrapped separately from linear fit for each level of SES. These data (as well as our main models and further analyses in SI 3H/G) do not reveal an SES effect on child speech. apart by future work. 350 New Insight on Child and Family Factors. Our main models, 351 figures showing the raw data, and additional analyses (in 352 the North American subset of the data, as well as using an 353 alternative algorithm, see SI3F) reveal effects of normativity, 354 age, and adult talk but not SES (measured here through 355 maternal education), child gender, or monolingualism. To 356 illustrate the complexities involved in determining causal links 357 between child and family factors and child language skills, we 358 again consider how causal links might manifest, using SES as 359 a central example. 360 Our findings bear on debates regarding SES-associated 361 academic achievement differences in Western industrialized 362 societies (31, 32). Slower language development has often 363 been attributed to parents from lower-SES backgrounds pro364 viding less input to their children (viewed from a middle-class 365 Western-centric perspective (32)), leading to calls for behav366 ioral interventions aiming to increase it. Proponents of such 367 interventions might highlight our correlation between adult 368 talk and child speech; critics might instead underscore our 369 finding that SES was not significant in our main analyses nor 370 in every other re-analysis we attempted (SI3E:G). 371 A full understanding of how SES may relate to children’s 372 language input is complicated for empirical and conceptual 373 reasons, leaving strong conclusions premature. On the empiri374 cal side, two recent meta-analyses have investigated SES–input 375 correlations, one focused on LENA™measures (15), and the 376 other based on human-annotated measures (mostly from short 377 lab recordings) (16). The former finds evidence consistent 378 with a publication bias; correcting this bias statistically nearly 379 halves the association between SES and LENA™’s adult talk 380 measure (r= .19 versus .12). The latter finds a sizeable SES 381 6| www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX Bergelson et al. DRAFT effect when inspecting infant-directed speech (r = .34) and a 382 much smaller one when analyzing overall input quantities (r 383 = .09). Together, these studies suggest that our best estimate 384 of the association between overall input quantities and SES is 385 small (r = .1) and may not be detectable even with a sample 386 as large as ours (where the effect was estimated at |d| = .06, 387 or |r| = .03, which did not reach the threshold for significance). 388 Similarly, descriptive plots of the potential correlation between 389 our SES proxy and children’s speech (Figure 4) did not suggest 390 a strong or stable relationship across the 18 corpora, leading 391 to our conclusion that, in the sample as a whole, on average, 392 maternal education does not predict how much adults and 393 children talk.394 On the conceptual side, SES differences in input and lan395 guage skills may depend on how language is measured (33). For 396 instance, we speculate that SES effects may be magnified by 397 measures like prevalence of low-frequency words and complex 398 sentence structures common in written text. Such words and 399 structures may occur more in the input to Western, higher-SES 400 children because of parenting practices stereotypical in these 401 groups (34). Moreover, such measures may predict academic 402 achievement better than others, because of the importance 403 literacy has in Western schooling today. In contrast, SES 404 differences in input may be minimized by holistic measures of 405 speech quantities. Indeed, a strength of daylong recordings 406 is that they provide a relatively neutral (rather than West407 ern, high SES-centric) measure, as they tap into how much 408 children are contributing (via speech) to their community’s 409 conversational interactions instead of how many rare words or 410 complex constructions they have been taught.411 An exclusive focus on word counts or speech quantities 412 likely misses certain behaviors. As machine learning advances 413 (35), it may soon be possible to automatically transcribe 414 conversations happening in daylong recordings (at least in 415 monolingual high-resource language contexts). We suspect 416 that analysis of conversational content may reveal SES dif417 ferences in, e.g., rare word use or family practices around 418 book-reading even in naturalistic samples (36). Future work 419 with a high-density longitudinal lens is also needed to assess 420 the predictive value of global quantitative measures of speech 421 (like those we employ) relative to more specialized measures 422 (e.g. book-reading practices) with respect to culturally-relevant 423 outcomes (e.g. academic achievement, pragmatic competence 424 in multi-party conversation, etc.)425 In our view, causal links between parental behavior and chil426 dren’s outcomes can best be illuminated by randomized control 427 trials. Discovering and leveraging such links to change long428 term language outcomes depends on community partnership429 based approaches that are informed by the role that structural 430 inequalities play in these outcomes and engage with culturally 431 informed perspectives (37). The present results should not 432 be used to deny families access to resources that evidence 433 suggests are linked with better outcomes for children and their 434 families.435 Complicated causal effects are integral to all developmental 436 processes. While we illustrated this with our SES null results, 437 we also found no differences in child speech or adult talk as 438 a function of child gender or multilingual status. Regarding 439 multilingualism, we could not examine relative input in each 440 language the child was exposed to. Future machine learning 441 advances will permit the separate quantification of different 442 languages in daylong recordings, but this must happen along443 side reflection on how to fairly measure input and outcomes 444 in such heterogeneous populations (38–40). 445 Automated Tools and What They Count. A key benefit of our 446 approach is that we were able to pool and identically process 447 40,933 hours of independently-collected data (SI3A). Moreover, 448 unlike parental surveys, clinical assessments, lab instruments, 449 or hand-annotated data, current published evidence suggests 450 that the LENA™algorithm’s results do not vary systematically 451 by language (though they do vary somewhat across samples, 452 12). More relevant here, in analyzing the algorithm’s accuracy 453 as a function of samples grouped by language and cultural 454 features, we found no significant differences (Methods, SI1E). 455 While children’s language skills grow dramatically over 2–48 456 months, our measure is not an index of comprehension (which 457 can show quite a different trajectory, 41) but rather of ob458 servable linguistic behavior, focusing exclusively on children’s 459 rate of linguistic vocalizations (SI3B). These results certainly 460 do not deny effects found on proxies of more narrow-scoped 461 linguistic developments (e.g. vocabulary, processing efficiency, 462 or syntactic complexity), given that some predictors that fail 463 to explain variance here may nonetheless be significant there 464 (3, 42). 465 The same holds for our measure of adult talk, which is 466 quantitative and holistic; additional research is needed to dis467 tinguish child-directed from child-available speech, with the 468 latter including all speech the child hears. Although some 469 research suggests child-directed speech shows tighter correla470 tions with children’s vocabulary than child-available speech 471 does (43, 44), the importance of the latter has not been as fully 472 studied for other types of language knowledge (45); and, as far 473 as we know, this paper is the first to document a significant 474 link for everyday child speech behavior. Therefore, it would 475 be relevant to further investigate the strength of the predictive 476 value of overall adult talk (which was a significant predictor 477 here) versus child-directed talk, in a similarly large and diverse 478 sample as the present one. Unfortunately, automated tools for 479 separating child-directed from overheard speech are not yet 480 sufficiently accurate to make this possible (46). Future work 481 could also develop promising new approaches for considering 482 other sources of speech (e.g., other children) given their rele483 vance as a function of family structure (47). These approaches 484 were not possible here due to both technical algorithmic con485 straints and family structure information not being available 486 in our data-subsets. Another fruitful future direction could 487 consider conversational dynamics, studying both children’s 488 tendency to vocalize around adults and the complexity of such 489 vocalizations. Recent work (that is critically reliant on human 490 annotation of social intent) raises particularly interesting ideas 491 in this domain (14, 48). Relatedly, novel exploratory analyses 492 describing the acoustics of children’s vocalizations (49) hold 493 promise for driving future hypothesis-testing work building on 494 the present results. 495 Whatever measures are employed in the future as proxies 496 of child language production and input, we strongly encourage 497 researchers to consider psychometric properties and ecological 498 validity. The current approach demonstrates measure validity 499 that is comparable to that of other standard infant instruments 500 (SI1D:E). As context, measures used as proxies for infant 501 language and cognitive knowledge are inherently noisier than 502 the best batteries used to assess highly educated adults in 503 Bergelson et al. PNAS | August 23, 2023 | vol. XXX | no. XX | 7 DRAFT Western-centric settings. Notably, even there, reliabilities can 504 fall well below r = 1.‖ 505 Moreover, standardized tests face ecological validity threats, 506 particularly when applied cross-culturally. If our goal is to mea507 sure and understand the human mind, we need implementable, 508 culturally sensitive and appropriate ways of measuring human 509 behavior on a large scale. To our knowledge, there are no 510 such measures whose reliability has been examined, driving 511 us to conduct extensive quantification of the reliability of the 512 metrics we employed here (SI1D:E). We found that our mea513 sures show levels of reliability that are consistent with those 514 already in use for research and clinical purposes in infant pop515 ulations. For example, the MacArthur-Bates Communicative 516 Development Inventory (a parental report instrument used 517 largely as a proxy for vocabulary size) has been the basis for 518 cross-linguistic, demographic, and clinical research (9, 51–53), 519 and reports a median correlation between itself and labora520 tory measures of .61 (54). Our median accuracy comparing 521 automated and manual annotation for each of our algorithms 522 (LENA™and VTC) is .74, squarely in line with field standards 523 (SI1E). Indeed, converging evidence across these two wholly 524 separate algorithms regarding overall accuracy of our measure 525 serves to increase confidence in the validity of our results.526 In sum, rather than eliciting knowledge or caregiver-child 527 interaction in a constrained lab setting, or using checklists 528 in contexts where they make little sense socio-culturally, we 529 measure everyday language use en masse. Our measure of 530 early speech production is global, since we simply measure 531 more versus less speech or speech-like production on the part 532 of adults and children as they go about their daily life. And 533 yet, these measures have important advantages, which led us 534 to select them as proxies here, including comparable relia535 bility to other measures of language development commonly 536 used in both research and applied settings (Methods, SI1D:E); 537 reported correlations between them and finer-grained, “quali538 tative” measures of language development (SI1D), and conver539 gent validity with respect to standardized language tests (13). 540 Most importantly, our speech measure merits consideration as 541 one of many possible proxies of language development thanks 542 to its cross-cultural adaptability, observer-free sampling vol543 ume, and sheer ecological validity. Indeed, our results raise 544 the possibility that more ecologically-valid lexical, phonetic, or 545 grammatical measures will also reveal stability across factors 546 like SES (55), gender, and multilingualism. Exploring these 547 factors, however, awaits machine-learning developments that 548 can extract such fine-grained linguistic measures from the raw 549 audio collected with child-worn devices.550 Conclusion. Our analysis of speech behavior in daily life 551 around the world evinces scientific progress on two fronts. 552 First, by revealing substantial variation in young children’s 553 speech, we provide evidence against a monolithic picture of 554 language development. Instead, this work reveals individ555 ual variation as fundamental to our understanding of this 556 species-wide ability. Second, by tapping into natural speech 557 interactions at unprecedented scale and diversity, we are able 558 to move beyond prior work by simultaneously considering the 559 interlocking factors that affect speech production over early 560 development. Our results reveal not only the expected cor561 relations with age and clinical factors, but also substantial 562 ‖ For instance, prior work finds test-retest reliabilities as low as r = .6 for certain sections of the widely used Wechsler Adult Intelligence Scale among North American English-speaking adults (50). associations with adult talk. All other factors paled in compar563 ison with these three, the null effect of our SES proxy being 564 of particular noteworthiness. These findings open exciting av565 enues for both theoretical research and potential applications, 566 including the prospect of behavioral interventions to harness 567 adult talk in the context of speech and language diagnoses. 568 Small-scale experimental and observational research has been 569 fundamental to our understanding of language, development, 570 and the human mind. Machine learning (like that in speech 571 technology) promises to extend our scientific reach by explod572 ing the range of everyday interactions we are able to capture 573 and analyze. Just as recent technological innovations have 574 opened new vistas in understanding the vocalizations of mice 575 and whales (6, 7), so too does speech technology have the 576 potential to reveal how everyday human communication gives 577 rise to language learning in children around the world. 578 Methods 579 All code used to generate our analysis and the 580 manuscript is available at https://osf.io/9v2m5/?view_only= 581 50df17fcf0844145ae692c35b78c6b08.582 Data Discovery and Integration. We took steps to counter a preva583 lent bias for normative North American data (see SI3A for 584 corpus constitution procedure). Included data were indepen585 dently collected by 18 stewards (56–77); see SI5 for list of 586 publications based on individual datasets. We note that while 587 our corpora covered a much greater variety of participants 588 than prior work, it would not be appropriate to interpret our 589 samples as comprehensively representative of the country or 590 language community from which they are drawn. 591 Socioeconomic status and normative development were 592 streamlined for cross-corpus consistency (SI2A:B, SI3A, Fig593 ure S3A.1). For socioeconomic status we use maternal ed594 ucation, a reliable proxy for SES in previous research on 595 language development (18, 78). Maternal education was avail596 able across all datasets, and could be converted into a 5597 point maternal education scale with levels corresponding to 598 less than high school degree, high school degree or equiva599 lent, some college/vocational/associate degree level training, 600 university/college degree, and advanced degree (SI2B; Table 601 S2B.1). 602 For non-normative development, data stewards had tagged 603 a wide variety of infant or familial characteristics as poten604 tially non-normative. We confirmed that the classification 605 was backed up by extant literature (SI2A). Infants ultimately 606 classified as having non-normative development in the present 607 sample include those who met one or more of the following 608 criteria: preterm birth (<37 weeks); diagnosed speech or lan609 guage delay; global developmental delay; low birth weight 610 (<2500g when specified); hearing loss, hearing aids or cochlear 611 implants; familial risk of Autism Spectrum Disorder, specific 612 language impairment and/or dyslexia; other relevant genetic 613 syndromes. Notably, our child vocalization rate measure is 614 not a standardized normed clinical evaluation, and thus non615 normative status may not necessarily translate to behavior 616 that falls >1 standard deviations below the norm in these 617 naturalistic recordings. 618 Analysis Details. We first randomly partitioned the data within 619 each corpus such that 35% of monolingual, normative chil620 dren were placed in an exploration set (N children = 264; N 621 8| www.pnas.org/cgi/doi/10.1073/pnas.XXXXXXXXXX Bergelson et al. DRAFT recordings = 850), and all others in a confirmation set (N 622 children = 737; N recordings = 2025) (SI3A). The exploration 623 set was used to study the psychometric properties of potential 624 language input and output variables (SI3B), resulting in the 625 selection of the output variable referred to as child speech 626 above, and CVCr (Child Vocalization Count rate) in anal627 ysis and supplementary files (SI3B, Table S3B.1); and the 628 input variable referred to as adult talk above, and AVCr 629 (Adult Vocalization Count rate) in analysis and supplemen630 tary files (SI3B, Table S3B.2). Note that this includes both 631 child-directed and child-available speech.632 In addition, we used the exploration set to check the ro633 bustness of results to variation in random effect structure, and 634 explored diverse model structures using mixed models in R’s 635 lme4 package (79), checking whether the addition of effects or 636 interactions explained additional variance (SI3C). This led us 637 to (a) include overlap rate as a covariate (see Figure S3C.1), 638 to control for the fact that in noisy environments, more child 639 speech and adult talk within the same recordings may be 640 labeled as “overlap” by LENA (and thus not attributed to 641 either speaker type) and (b) to not include random slopes 642 for any of the predictors. Regarding the latter choice, our 643 exploration of random effect structure revealed that models 644 including random slopes for any of the predictors (notably 645 including gender and SES) as a function of corpus led to 646 non-convergent models. While such non-convergence could 647 be due to various reasons, the most likely explanation is that 648 the model is overparametrized (80), i.e., variance cannot be 649 reliably attributed to predictors within each corpus (see SI3H 650 for additional checks, including one including random slopes 651 for SES, and SI2B for discussion of alternatives to our SES 652 implementation).653 Evaluation against human annotations. To assess the validity of 654 our child speech and adult talk measures, we evaluated them 655 against human annotations (see SI1D:E for further informa656 tion). The median correlation of human to algorithm perfor657 mance for the algorithms is >.7, i.e. comparable reliability to 658 established developmental clinical and research instruments 659 (81–83). As far as we know, the present multi-cultural val660 idation exceeds those from prior research instruments. For 661 example, the Ages and Stages Questionnaire (84) is a standard 662 instrument used at well-child visits in the U.S. It is also recom663 mended by the World Bank as one of the most popular tools 664 to measure child development, used in at least 20 countries 665 (85). And yet, a recent systematic review (83) reports only 6666 reliability analyses (averaging, e.g., .7 for internal consistency 667 at 24mo.). Relative to this, our validation effort containing es668 timates for 14/18 corpora and finding strong validity is notable. 669 Finally, one may wonder whether the LENA™algorithm per670 forms less well for languages and cultures that diverge from 671 its training set, which was English-learning children growing 672 up in an urban/suburban U.S. setting. Although we observe 673 considerable corpus variation, this variation is not attributable 674 to whether children were learning English or growing up in 675 an urban setting, as assessed by Welch’s t-tests, for either 676 our child speech measure (CVCr; English versus non-English677 medians 0.785 vs. 0.71, t (6.04) = -0.5, p = 0.637; urban versus 678 rural medians 0.77 vs. 0.71, t (8.11) = -0.46, p = 0.661), or 679 for our adult talk measure (AVCr; English versus non-English 680 medians 0.75 vs. 0.74, t (7.91) = 0.42, p = 0.686; urban ver681 sus rural medians 0.75 vs. 0.74, t (3.07) = -0.23, p = 0.835). 682 Instead, our results suggest that corpus variation more likely 683 reflects how the human annotation was done rather than how 684 well the algorithm worked, since the corpora with lower reli685 abilities were also those in which the human annotation was 686 more coarse-grained (see SI1E). 687 Additional algorithm. To make sure that key conclusions were 688 robust to methodological details, we reanalyzed the subset of 689 the data for which data stewards shared audio with a newer, 690 open-source alternative to LENA™: the Voice Type Classifier 691 (VTC) (86). Like the LENA™algorithm, VTC returns an 692 estimation of child and adult vocalization counts. A total 693 of 1065 audio files from 11 corpora were available for this 694 reanalysis (SI3F). 695 The VTC algorithm employs a completely different ap696 proach than the proprietary algorithm developed by LENA™, 697 including the use of neural networks running directly from the 698 audio (rather than from MFCC features). VTC allows multi699 ple talker classes to be activated at the same time, whereas 700 in the LENA™algorithm, overlap between talkers (or be701 tween a talker and noise) is tagged as “Overlap,” which is 702 not counted towards children’s input or output. VTC also 703 differs from LENA™in its training set. While LENA™was 704 trained entirely on data from North American, monolingual 705 English-learning, urban children, VTC was developed using 706 the combination of various corpora of children residing in 707 urban or rural settings and learning one or more of several lan708 guages (including the tonal language Minn, French, Ju|’hoan, 709 Tsimane, English, and several others, in rough order of quan710 tity of data). Further information on accuracy is provided in 711 SI1E; both algorithms render similar accuracy when compared 712 to human annotation as noted above. 713 Models. We used linear mixed regressions (Gaussian family), 714 and established model structure from the exploration data 715 (SI3C). Hypotheses were derived from exploratory models and 716 systematic reviews of literature on monolingualism and nor717 mativity (SI3D). The model predicting the rate of children’s 718 linguistic vocalizations (i.e. child speech) was: child_gender + 719 SES + child_normative∗AV Cr ∗age + child_monolingual∗720 AV Cr ∗age + overlap + (1 + overlap + AV Cr|corpus ) + 721 (1 |corpus : child_id ). The model predicting the rate of adult 722 linguistic vocalizations (i.e. adult talk) was: child_gender + 723 SES + child_normative ∗age + child_monolingual ∗age + 724 overlap + (1 + overlap|corpus ) + (1 |corpus : child_id ). Full 725 model details and a link to model diagnostics are provided 726 in SI3E. We report estimates (standardized, which serve as 727 effect sizes), standard errors of the estimates, and q-values 728 (FDR-corrected p-values); see Tables 1and 2.729 Participants. Table 3lists participant characteristics noting both 730 (1) the exploration/confirmation split (SI3A), and (2) that 731 some children provided multiple recordings. We excluded 732 2/850 recordings from 1/264 children from the exploration set 733 and 8/2025 recordings from 5/737 children in the confirmation 734 set from our models because data regarding their maternal 735 education was missing. For child gender, there were slightly 736 more boys than girls. This was in part because corpora with 737 children with non-normative development also include children 738 with normative development matched in gender, leading to an 739 over-representation of boys since more boys than girls have 740 non-normative development. See Table 3and Figure 5for 741 specific numbers and visualized distributions. 742 Bergelson et al. PNAS | August 23, 2023 | vol. XXX | no. XX | 9