---
title: "Child-directed speech — definition and key papers"
description: "Child-directed speech is the speech register adults and older children use when talking to young children."
canonical: "https://research.elamur.ai/free/en/concepts/child-directed-speech/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/concepts/child-directed-speech/)

[home](https://md.research.elamur.ai/free/en/) / [concepts](https://md.research.elamur.ai/free/en/concepts/) / Child-directed speech

# 🧸 Child-directed speech

Concept · 27 key papers · [all concepts](https://md.research.elamur.ai/free/en/concepts/)

Child-directed speech is the speech register adults and older children use when talking to young children. It typically has higher and more variable pitch, slower tempo, shorter and simpler utterances, frequent repetition and a restricted, concrete vocabulary. In many communities it makes up a large share of the linguistic input to early language acquisition.

Also: CDS, infant-directed speech (IDS), child-directed language (CDL), motherese, parentese, baby talk, caregiver speech

## What it is

Child-directed speech (CDS) is a recognisable way of talking that caregivers adopt when they address infants and toddlers. Descriptive work from the 1960s onward established its main features:

- Prosody: higher pitch, a wider pitch range and exaggerated intonation contours.
- Tempo: slower speech, longer pauses and clearer boundaries between phrases.
- Short utterances: fewer words per utterance and fewer embedded clauses.
- Repetition: the same words and frames are repeated, often with small variations.
- Vocabulary: a small set of frequent, concrete words tied to the here-and-now, plus special forms such as diminutives.
- Interaction: many questions, imperatives and attention-getters, with turns built around the child's focus.

These features shift as the child grows, so CDS addressed to a two-year-old differs from CDS addressed to a five-year-old. The register is widespread across cultures, but how much of it caregivers use, and in what form, varies considerably. In some communities, children hear much of their language as overheard adult talk rather than speech addressed to them.

For research, CDS matters because it is the actual input from which children build a grammar and lexicon, and any theory of acquisition has to say what can be learned from it. Dialogue between children and caregivers also has its own structure. One study recorded sessions from NannyME, a Japanese service in which human caregivers talk with children over an online video link. It compared these sessions with adult conversation and with parent–child corpora. In that dataset, overlapping speech was lower in child–caregiver talk than in adult dialogue, with overlap measured as cumulative duration per minute. Caregivers also steered which communicative intentions dominated each part of a session ([Yamashita et al. 2026](https://md.research.elamur.ai/free/en/papers/yamashita-2026-analysis-child-caregiver-interactions-developing/)).

## The key idea

The central claim is that the input children receive is not a random sample of adult language. It is shaped, simplified and organised in ways that may help learning. A standard measure of the "short utterances" feature, used both for CDS and for children's own speech, is mean length of utterance:

```
MLU = M / U
```

Here M is the total number of morphemes (or words, in a word-based variant) in a transcript sample, and U is the number of utterances in that sample. In Brown's procedure, MLU is computed over a fixed sample of utterances, with explicit rules for what counts as a morpheme. Word-based and morpheme-based values are therefore not directly comparable. MLU in CDS is low relative to adult-directed speech and tends to rise as the child's own MLU rises. This pattern is often read as caregivers finely tuning their input to the learner, a reading disputed by Newport, Gleitman and Gleitman (1977).

Two broad positions follow from this. On a facilitative view, simplified, repetitive, prosodically marked input makes word boundaries, categories and constructions easier to detect. On a sceptical view, CDS is simple on the surface but not tailored to the hard parts of grammar. It therefore leaves the poverty-of-the-stimulus problem intact: the argument that the input alone underdetermines the grammar children acquire, so some learning biases must be innate.

Computational modelling offers one way to test parts of this debate. A learner is trained on CDS, or on a matched amount of other text, and the test is what it generalises. Such models show what a given learner can extract from the input, not what children actually do.

A related idea is that limits on the learner, and not only on the input, shape acquisition. One study models working-memory growth as y = b − aˣ with 0 \< a \< 1, where y is memory capacity, x is developmental time and b is the adult ceiling. Capacity starts at b − 1 when x = 0, and a smaller a makes it approach b faster ([Mita et al. 2025](https://md.research.elamur.ai/free/en/papers/mita-2025-developmentally-plausible-working-memory-shapes/)). The curve reflects rapid growth in early childhood, slower growth in late childhood and an adult plateau. It was turned into a training schedule in which a GPT-2 model's memory is tightly limited at first and relaxed exponentially as training proceeds. The model was trained on AO-CHILDES, an age-ordered collection of English child-directed transcripts. Under this schedule it outperformed both unconstrained and fixed-limit baselines on Zorro, a syntactic test suite designed for child-directed speech, with the clearest late gains on case and filler-gap items ([Mita et al. 2025](https://md.research.elamur.ai/free/en/papers/mita-2025-developmentally-plausible-working-memory-shapes/)).

## History and founding papers

Charles Ferguson (1964) described "baby talk" as a register found across languages. In his usage the term also covers special lexical forms such as *doggie* or *choo-choo*, a narrower sense than the register as a whole. Catherine Snow (1972) showed that mothers systematically simplify their speech to young children. Roger Brown (1973) popularised MLU as a developmental index. Newport, Gleitman and Gleitman (1977) questioned how much this simplification actually drives syntactic development. CHILDES was founded in 1984 by Brian MacWhinney and Catherine Snow as a shared database of transcribed child–caregiver interaction. Its manual, whose third edition MacWhinney published in 2000, documents it, and the database became the main empirical base for corpus studies of CDS. Anne Fernald (1985) highlighted infants' preference for the prosody of CDS.

Computational work built on CHILDES through structured annotation. An early example is high-accuracy syntactic annotation and parsing of CHILDES transcripts ([Sagae et al. 2007](https://md.research.elamur.ai/free/en/papers/sagae-2007-high-accuracy-annotation-parsing-childes/)). Automatic analysis of child transcripts was also applied to measuring development, for example through grammar checking in early-childhood data ([Hassanali & Liu 2011](https://md.research.elamur.ai/free/en/papers/hassanali-2011-measuring-language-development-early-childhood/)).

Computational learners then modelled acquisition directly from CDS. One line paired child-directed utterances with meaning representations and learned syntax and semantics jointly ([Kwiatkowski et al. 2012](https://md.research.elamur.ai/free/en/papers/kwiatkowski-2012-probabilistic-model-syntactic-semantic-acquisition/)). Another grounded word learning in multimodal child-directed input ([Lazaridou et al. 2016](https://md.research.elamur.ai/free/en/papers/lazaridou-2016-multimodal-semantic-learning-child-directed/)). Grammar induction from child-directed utterances examined how much category labels matter ([Jin & Schuler 2020](https://md.research.elamur.ai/free/en/papers/jin-2020-importance-category-labels-grammar-induction/)).

## How it is used today

Corpus resources. CHILDES transcripts are being enriched with standard annotation layers so that statistics about the input can be computed at scale. Several layers use Universal Dependencies (UD), a cross-linguistic scheme for annotating dependency syntax.

| Resource | What it adds |
| --- | --- |
| UD and semantics for English and Hebrew CDS ([Szubert et al. 2022](https://md.research.elamur.ai/free/en/papers/szubert-2022-universal-dependencies-semantics-english-hebrew/)) | Dependency trees and semantic representations |
| UD-English-CHILDES ([Yang et al. 2025](https://md.research.elamur.ai/free/en/papers/yang-2025-ud-english-childes-collected-resource/)) | Gold (manually verified) and silver (automatically parsed) UD trees |
| IPA CHILDES with G2P+ ([Goriely & Buttery 2025](https://md.research.elamur.ai/free/en/papers/goriely-2025-ipa-childes-g2p-feature-rich/)) | Phonemic transcriptions from grapheme-to-phoneme conversion, for cross-lingual phonology |
| Grammaticality annotation ([Nikolaus et al. 2024](https://md.research.elamur.ai/free/en/papers/nikolaus-2024-automatic-annotation-grammaticality-child-caregiver/)) | Automatic labels for child–caregiver conversations |

These layers support direct questions about what children hear. One example is an automated search of CHILDES for filler-gap dependencies, which estimates how often such constructions reach the learner ([Zhou et al. 2026](https://md.research.elamur.ai/free/en/papers/zhou-2026-what-exactly-do-children-receive/)). Filler-gap dependencies are constructions such as wh-questions and relative clauses, in which a displaced phrase is linked to an empty position elsewhere in the sentence.

Related corpora add audio and dialect coverage. ChildTalk holds 112.5 hours of recorded conversation between Chinese children aged 2–8 and familiar adults, from 500 dyadic sessions released as complete dialogues. It spans standard Mandarin, eight Mandarin dialect subgroups, Southern Min and Jin, and is built for speech recognition ([Zhou et al. 2026](https://md.research.elamur.ai/free/en/papers/zhou-2026-childtalk-multi-dialect-chinese-child/)).

Language models at child scale. A major use of CDS is as training data for small language models whose data budgets resemble what a child hears. BabyBERTa showed that a compact masked model trained on small-scale child-directed language can learn substantial grammatical knowledge ([Huebner et al. 2021](https://md.research.elamur.ai/free/en/papers/huebner-2021-babyberta-learning-more-grammar-small/)). The same approach has been carried to Hebrew ([Gelboim & Sulem 2025](https://md.research.elamur.ai/free/en/papers/gelboim-2025-tafberta-learning-grammatical-rules-small/)) and to Italian, where recurrent and transformer architectures were compared ([Fusco et al. 2024](https://md.research.elamur.ai/free/en/papers/fusco-2024-recurrent-networks-are-linguistically-better/)).

The BabyLM challenge is a shared task in which models are trained on child-sized amounts of text. Work in this line adds developmentally inspired training schemes:

- curricula ordered by sentence complexity ([Oba et al. 2023](https://md.research.elamur.ai/free/en/papers/oba-2023-babylm-challenge-curriculum-learning-based/))
- masking schedules based on child language development ([Lucas et al. 2024](https://md.research.elamur.ai/free/en/papers/lucas-2024-curriculum-masking-based-child-language/))
- a pre-training scheme that advances step by step, inspired by child language development ([Yoshida et al. 2025](https://md.research.elamur.ai/free/en/papers/yoshida-2025-batch-wise-convergent-pre-training/))
- synthetic child-directed questions as extra input ([Poh et al. 2025](https://md.research.elamur.ai/free/en/papers/poh-2025-what-did-you-say-generating/))

Interactive settings model the caregiver as well. One referential-game study tests whether theory of mind (reasoning about what a partner knows) helps a simulated learner acquire language. A pretrained listener plays the caregiver, and the speaker reranks candidate utterances using its own internal model of that listener. The speaker is capped at a 200-token vocabulary and 20-token utterances to resemble an early learner ([Liu et al. 2023](https://md.research.elamur.ai/free/en/papers/liu-2023-computational-language-acquisition-theory-mind/)).

Testing acquisition theories. Position papers argue that neural networks can serve as testable learners for claims about human acquisition, provided the training data and conditions are plausible ([Warstadt & Bowman 2022](https://md.research.elamur.ai/free/en/papers/warstadt-2022-artificial-neural-networks-tell-us/)). Learning trajectories of models have been compared with children's developmental stages ([Evanson et al. 2023](https://md.research.elamur.ai/free/en/papers/evanson-2023-language-acquisition-children-language-models/)). Poverty-of-the-stimulus arguments have been probed by training networks on CDS and testing hierarchical generalisation ([Yedetore et al. 2023](https://md.research.elamur.ai/free/en/papers/yedetore-2023-how-poor-is-stimulus-evaluating/)).

## Open questions

Is CDS actually better data? Controlled comparisons have asked whether child-directed speech is effective training data compared with other text of similar size ([Feng et al. 2024](https://md.research.elamur.ai/free/en/papers/feng-2024-is-child-directed-speech-effective/)). Results so far are mixed. Across languages and benchmarks, child-directed language does not consistently improve syntax learning ([Padovani et al. 2025](https://md.research.elamur.ai/free/en/papers/padovani-2025-child-directed-language-does-not/)). Another study suggests its benefit appears in production rather than comprehension ([Bunzeck & Zarrieß 2026](https://md.research.elamur.ai/free/en/papers/bunzeck-2026-child-directed-speech-facilitates-production/)). Whether the facilitative features of CDS matter for neural learners, as they are claimed to for children, remains unresolved.

What is missing from text? Most model studies use orthographic transcripts. These drop the prosody, tempo and situational context that define CDS for the infant. Phonemic resources restore some sound structure ([Goriely & Buttery 2025](https://md.research.elamur.ai/free/en/papers/goriely-2025-ipa-childes-g2p-feature-rich/)). Grounded setups that pair utterances with visual scenes restore some situational context ([Lazaridou et al. 2016](https://md.research.elamur.ai/free/en/papers/lazaridou-2016-multimodal-semantic-learning-child-directed/)). Neither captures prosody or tempo.

Which properties of the learner matter? The working-memory results ([Mita et al. 2025](https://md.research.elamur.ai/free/en/papers/mita-2025-developmentally-plausible-working-memory-shapes/)) suggest that the fit between input and learner constraints may matter more than the input alone. Curriculum methods ([Oba et al. 2023](https://md.research.elamur.ai/free/en/papers/oba-2023-babylm-challenge-curriculum-learning-based/)) approach that fit from the other side: they reorder the input by complexity rather than constraining the learner.

How representative are the corpora? CHILDES is dominated by a few languages and by particular settings. Evidence about rare constructions depends on reliable automatic detection ([Zhou et al. 2026](https://md.research.elamur.ai/free/en/papers/zhou-2026-what-exactly-do-children-receive/)) and on the quality of parsed annotation ([Yang et al. 2025](https://md.research.elamur.ai/free/en/papers/yang-2025-ud-english-childes-collected-resource/)). Wider coverage of languages, dialects and interaction types is still needed. Multi-dialect conversational speech corpora such as ChildTalk are one step in that direction ([Zhou et al. 2026](https://md.research.elamur.ai/free/en/papers/zhou-2026-childtalk-multi-dialect-chinese-child/)).

## Key papers

The works this page cites, the most cited first; citations as counted by OpenAlex plus the papers of this library that cite the work, read 2026-10-23.

- [BabyBERTa: Learning More Grammar With Small-Scale Child-Directed Language](https://md.research.elamur.ai/free/en/papers/huebner-2021-babyberta-learning-more-grammar-small/) — Philip A. Huebner, Elior Sulem, Fisher Cynthia et al. · 2021 · 91 citations
- [What Artificial Neural Networks Can Tell Us About Human Language Acquisition](https://md.research.elamur.ai/free/en/papers/warstadt-2022-artificial-neural-networks-tell-us/) — Alex Warstadt, Samuel R. Bowman · 2022 · 72 citations
- [Language acquisition: do children and language models follow similar learning stages?](https://md.research.elamur.ai/free/en/papers/evanson-2023-language-acquisition-children-language-models/) — Linnea Evanson, Yair Lakretz, Jean-Rémi King · 2023 · 33 citations
- [How poor is the stimulus? Evaluating hierarchical generalization in neural networks trained on child-directed speech](https://md.research.elamur.ai/free/en/papers/yedetore-2023-how-poor-is-stimulus-evaluating/) — Aditya Yedetore, Tal Linzen, Robert Frank et al. · 2023 · 27 citations
- [A Probabilistic Model of Syntactic and Semantic Acquisition from Child-Directed Utterances and their Meanings](https://md.research.elamur.ai/free/en/papers/kwiatkowski-2012-probabilistic-model-syntactic-semantic-acquisition/) — Tom Kwiatkowski, Sharon Goldwater, Luke Zettlemoyer et al. · 2012 · 27 citations
- [Computational Language Acquisition with Theory of Mind](https://md.research.elamur.ai/free/en/papers/liu-2023-computational-language-acquisition-theory-mind/) — Andy Liu, Hao Zhu, Emmy Liu et al. · 2023 · 21 citations
- [High-accuracy Annotation and Parsing of CHILDES Transcripts](https://md.research.elamur.ai/free/en/papers/sagae-2007-high-accuracy-annotation-parsing-childes/) — Kenji Sagae, Eric Davis, Alon Lavie et al. · 2007 · 20 citations
- [Is Child-Directed Speech Effective Training Data for Language Models?](https://md.research.elamur.ai/free/en/papers/feng-2024-is-child-directed-speech-effective/) — Steven Y. Feng, Noah D. Goodman, Michael C. Frank · 2024 · 16 citations
- [Child-Directed Language Does Not Consistently Boost Syntax Learning in Language Models](https://md.research.elamur.ai/free/en/papers/padovani-2025-child-directed-language-does-not/) — Francesca Padovani, Jaap Jumelet, Yevgen Matusevych et al. · 2025 · 9 citations
- [Developmentally-plausible Working Memory Shapes a Critical Period for Language Acquisition](https://md.research.elamur.ai/free/en/papers/mita-2025-developmentally-plausible-working-memory-shapes/) — Masato Mita, Ryo Yoshida, Yohei Oseki · 2025 · 9 citations
- [BabyLM Challenge: Curriculum learning based on sentence complexity approximating language acquisition](https://md.research.elamur.ai/free/en/papers/oba-2023-babylm-challenge-curriculum-learning-based/) — Miyu Oba, Akari Haga, Akiyo Fukatsu et al. · 2023 · 9 citations
- [Multimodal Semantic Learning from Child-Directed Input](https://md.research.elamur.ai/free/en/papers/lazaridou-2016-multimodal-semantic-learning-child-directed/) — Angeliki Lazaridou, Grzegorz Chrupała, Raquel Fernández et al. · 2016 · 9 citations
- [IPA CHILDES & G2P+: Feature-Rich Resources for Cross-Lingual Phonology and Phonemic Language Modeling](https://md.research.elamur.ai/free/en/papers/goriely-2025-ipa-childes-g2p-feature-rich/) — Zébulon Goriely, Paula Buttery · 2025 · 6 citations
- [Automatic Annotation of Grammaticality in Child-Caregiver Conversations](https://md.research.elamur.ai/free/en/papers/nikolaus-2024-automatic-annotation-grammaticality-child-caregiver/) — Mitja Nikolaus, Abhishek Agrawal, Petros Kaklamanis et al. · 2024 · 5 citations
- [UD-English-CHILDES: A Collected Resource of Gold and Silver Universal Dependencies Trees for Child Language Interactions](https://md.research.elamur.ai/free/en/papers/yang-2025-ud-english-childes-collected-resource/) — Xiulin Yang, Zhuoxuan Ju, Lanni Bu et al. · 2025 · 4 citations
- [TafBERTa: Learning Grammatical Rules from Small-Scale Language Acquisition Data in Hebrew](https://md.research.elamur.ai/free/en/papers/gelboim-2025-tafberta-learning-grammatical-rules-small/) — Anita Gelboim, Elior Sulem · 2025 · 3 citations
- [What did you say? Generating Child-Directed Speech Questions to Train LLMs](https://md.research.elamur.ai/free/en/papers/poh-2025-what-did-you-say-generating/) — Whitney Poh, Michael Tombolini, Libby Barak · 2025 · 3 citations
- [Recurrent Networks Are (Linguistically) Better? An (Ongoing) Experiment on Small-LM Training on Child-Directed Speech in Italian](https://md.research.elamur.ai/free/en/papers/fusco-2024-recurrent-networks-are-linguistically-better/) — Achille Fusco, Matilde Barbini, Maria Letizia Piccini Bianchessi et al. · 2024 · 3 citations
- [Measuring Language Development in Early Childhood Education: A Case Study of Grammar Checking in Child Language Transcripts](https://md.research.elamur.ai/free/en/papers/hassanali-2011-measuring-language-development-early-childhood/) — Khairun-nisa Hassanali, Yang Liu · 2011 · 3 citations
- [Using Curriculum Masking Based on Child Language Development to Train a Large Language Model with Limited Training Data](https://md.research.elamur.ai/free/en/papers/lucas-2024-curriculum-masking-based-child-language/) — Evan Lucas, Dylan Gaines, Tagore Rao Kosireddy et al. · 2024 · 2 citations
- [What Exactly do Children Receive in Language Acquisition? A Case Study on CHILDES with Automated Detection of Filler-Gap Dependencies](https://md.research.elamur.ai/free/en/papers/zhou-2026-what-exactly-do-children-receive/) — Zhenghao Zhou, William Dai, Maya Viswanathan et al. · 2026 · 1 citation
- [Batch-wise Convergent Pre-training: Step-by-Step Learning Inspired by Child Language Development](https://md.research.elamur.ai/free/en/papers/yoshida-2025-batch-wise-convergent-pre-training/) — Ko Yoshida, Daiki Shiono, Kai Sato et al. · 2025 · 1 citation
- [Universal Dependencies and Semantics for English and Hebrew Child-directed Speech](https://md.research.elamur.ai/free/en/papers/szubert-2022-universal-dependencies-semantics-english-hebrew/) — Ida Szubert, Omri Abend, Nathan Schneider et al. · 2022 · 1 citation
- [The Importance of Category Labels in Grammar Induction with Child-directed Utterances](https://md.research.elamur.ai/free/en/papers/jin-2020-importance-category-labels-grammar-induction/) — Lifeng Jin, William Schuler · 2020 · 1 citation
- [Analysis of Child-Caregiver Interactions for Developing a Caregiver Spoken Dialogue System](https://md.research.elamur.ai/free/en/papers/yamashita-2026-analysis-child-caregiver-interactions-developing/) — Sanae Yamashita, Shota Mochizuki, Yuko Kuma et al. · 2026
- [Child-directed speech facilitates production, not comprehension, in BabyLMs](https://md.research.elamur.ai/free/en/papers/bunzeck-2026-child-directed-speech-facilitates-production/) — Bastian Bunzeck, Sina Zarrieß · 2026
- [ChildTalk: A Multi-Dialect Chinese Child Speech Corpus with Full-Length Child–Caregiver Conversations for Speech Recognition](https://md.research.elamur.ai/free/en/papers/zhou-2026-childtalk-multi-dialect-chinese-child/) — Jiaming Zhou, Yujie Guo, Shiwan Zhao et al. · 2026

## Related concepts

- [Psycholinguistics](https://md.research.elamur.ai/free/en/concepts/psycholinguistics/) — Psycholinguistics is the study of the mental processes by which people produce, comprehend and acquire language. It combines experimental…
- [Construction grammar](https://md.research.elamur.ai/free/en/concepts/construction-grammar/) — Construction grammar is a family of linguistic theories holding that knowledge of a language consists of constructions. Constructions are…
- [Language model](https://md.research.elamur.ai/free/en/concepts/language-model/) — A language model is a probability distribution over sequences of tokens (words, subwords or characters). It is usually factored into…
- [Cognitive linguistics](https://md.research.elamur.ai/free/en/concepts/cognitive-linguistics/) — Cognitive linguistics is an approach to language that treats grammar and meaning as products of general human cognition, including…

## More papers

Topics: [🧒 Psycholinguistics & language acquisition](https://md.research.elamur.ai/free/en/topics/psycholinguistics/)

Shelves: [🗣️ Linguistics & languages](https://md.research.elamur.ai/free/en/shelves/linguistics/)

- [← previous concept BLEU score](https://md.research.elamur.ai/free/en/concepts/bleu/)
- [next concept → Cognitive linguistics](https://md.research.elamur.ai/free/en/concepts/cognitive-linguistics/)
- [↑ back to concepts](https://md.research.elamur.ai/free/en/concepts/)