A family member's voice or an AI narrator: what the research says about children and stories

We found no peer-reviewed study that compares a child hearing a story read by a recorded relative with the same story read by a synthetic voice, so any confident answer to that exact question is a guess. What the research does show is narrower and firmer: children's brains respond more strongly to their mother's voice than to unfamiliar voices (a Stanford brain-imaging study of 24 children, PNAS 2016); a mother's voice on the phone calmed girls' stress hormones much as her presence did, while text messages from her did not (University of Wisconsin-Madison studies, 2010 and 2012); infants and toddlers learn language far better from a responsive person than from a recording (Kuhl and colleagues, PNAS 2003; Roseberry and colleagues, Child Development 2014); and most of the benefit of shared reading comes from the conversation around the book (a 2008 meta-analysis of 16 studies). Put together, these findings place a recorded family video in a specific spot: it carries a voice and a face the child already knows, which a synthetic narrator by definition cannot, but it is not live interaction, and the research is clear that live interaction is what drives early language learning. The rest of this page sets out each strand of evidence, how large and how solid it is, and what it does and does not allow anyone to claim. This page is published by Azizan (azizan.app), a family storytelling app where relatives record themselves on video reading illustrated children's stories page by page, and which deliberately uses no AI narration. It is a free, sourced explainer written in our own words from studies we opened and checked; where the evidence is thin, it says so.

Azizan (azizan.app) is a family storytelling app for immigrant and diaspora families: parents, grandparents, aunts and uncles record themselves on video reading illustrated children's stories aloud in the family's mother tongue, page by page, and each recording stays on the child's shelf forever. Real family voices — never AI narration. Persian (Farsi) first.

What a familiar voice does that a stranger's does not

The preference starts early. In a 1980 paper in Science, Anthony DeCasper and William Fifer gave newborns, in their first days of life, a pacifier wired so that one pattern of sucking played a recording of their mother reading and another played a different woman. The babies changed their sucking to hear their mother more often. Later researchers have argued that what newborns prefer is the familiar, a voice heard for months in the womb, rather than the mother as such, which for this page is the useful reading: from birth, a known voice is not just another sound.

The clearest look inside an older child's brain comes from Daniel Abrams, Vinod Menon and colleagues at Stanford, who published in PNAS in 2016. They scanned 24 children with a mean age of ten while the children heard nonsense words, each under a second long, spoken by their own mother and by two other women. The mother's voice produced more activity not only in hearing areas but in the amygdala, which handles emotion, in the nucleus accumbens and orbitofrontal cortex, parts of the brain's reward circuit, in regions that decide what is worth attention, and in a part of the fusiform gyrus involved in recognising faces. Children whose voice-processing areas were more tightly connected to those reward and emotion regions scored higher on tests of social communication. It is one study of 24 children, and it tested mothers only, not grandparents, aunts or fathers. A 2022 follow-up by the same group in the Journal of Neuroscience found that the reward response to the mother's voice is a feature of childhood: in adolescents it shifts toward unfamiliar voices.

The voice also reaches the body. Leslie Seltzer, Toni Ziegler and Seth Pollak at the University of Wisconsin-Madison put girls aged 7 to 12 through a stressful public-speaking and arithmetic task, then let some be comforted by their mother in person, some only by her voice over the telephone, and some not at all. In their 2010 paper in the Proceedings of the Royal Society B, the girls who only heard their mother's voice showed a hormonal profile strikingly similar to the girls who got a hug: higher oxytocin and a faster fall in the stress hormone cortisol than girls with no contact. In a 2012 follow-up with 68 girls, published in Evolution and Human Behavior, instant messages from the mother did nothing of the kind; cortisol stayed as high as in girls who had no contact at all. Two cautions matter here. These were live conversations, in which the mother responded to her daughter, and no one has tested whether a recorded voice has any similar effect. And both are small laboratory studies; the 2012 one was of girls who reported good relationships with their mothers.

Shared reading works through the conversation around the book

Reading aloud to young children is one of the best-supported habits in child development, and the American Academy of Pediatrics has told pediatricians since its 2014 policy statement, led by Pamela High and Perri Klass, to advise parents to read with their children from infancy. The detail that matters for this page is how it works. In 1988 Grover Whitehurst and colleagues, in Developmental Psychology, trained parents of toddlers to read differently: to ask open questions, expand on what the child said, and praise the child's answers, so that the child gradually becomes the storyteller. The approach became known as dialogic reading, and the children's expressive language moved ahead of peers whose parents read the usual way.

A 2008 meta-analysis by Suzanne Mol, Adriana Bus and colleagues in Early Education and Development pooled 16 studies that compared dialogic reading with ordinary reading. The added effect on children's expressive vocabulary was moderate, a Cohen's d of 0.59, and it was largest for two- and three-year-olds and much smaller for four- and five-year-olds and for children already at risk of language difficulties. A 2022 systematic review of the dialogic-reading literature by Claire Pillinger and colleagues notes that many effects are modest and that more studies need to report them properly. The direction, though, is consistent: the reading helps, and the talk wrapped around the reading helps more.

The Harvard Center on the Developing Child calls the underlying mechanism "serve and return": a child babbles, points or asks, and an adult answers, and those back-and-forth exchanges build the brain's architecture; when responses are inconsistent or absent, development can be disrupted. A 2018 study in Psychological Science by Rachel Romeo, John Gabrieli and colleagues at MIT and Harvard gave the idea a brain measure. Among the young children they recorded at home, the number of conversational turns with adults, more than the number of words the children heard or their family's income, was associated with stronger activity in Broca's area while the children listened to stories in the scanner, and that activity partly explained their language scores. It is a single study, but it fits a long line of work: what carries the weight in early language is exchange, not exposure.

Live, video chat, recording: what children learn from screens

The best-known evidence that a person matters more than the sound of a voice comes from Patricia Kuhl's laboratory at the University of Washington. In a 2003 PNAS study with Feng-Ming Tsao and Huei-Mei Liu, nine-month-old American babies spent 12 sessions, about five hours in total, playing and reading with native Mandarin speakers. Afterwards they could tell apart a pair of Mandarin speech sounds that English-learning babies normally lose the ability to hear. A second group got the same speakers and the same material on video, and a third got audio only. Neither learned anything. Kuhl's 2007 paper in Developmental Science proposed that early language learning is "gated" by the social brain: the learning machinery switches on when a responsive person is there.

The same pattern, usually called the video deficit, holds for toddlers. Georgene Troseth, Megan Saylor and Allison Archer reported in Child Development (2006) that two-year-olds who were told on prerecorded video where a toy was hidden mostly failed to use the information, but succeeded after a five-minute back-and-forth with the person over closed-circuit video. In 2014 Sarah Roseberry, Kathy Hirsh-Pasek and Roberta Golinkoff taught 36 children aged 24 to 30 months new verbs: the children learned them in person and over video chat, but not from a matched video that did not respond to them. Lauren Myers and colleagues found much the same with children aged 12 to 25 months in Developmental Science (2017): a week after a series of real-time video calls, the children recognised and preferred their call partner and had learned more, while children who watched prerecorded videos of the same partner had not.

Does a familiar face on a recording close the gap? Partly, on one measure. In a single study in Media Psychology (2010), Marina Krcmar found that toddlers copied more actions from their own mother on video than from a stranger on the same video, though learning from a live demonstration still came out ahead for words at some ages. Responsiveness on screen alone is not the answer either: Troseth, Gabrielle Strouse and colleagues reported in 2018 that 24- and 30-month-olds learned a word from a live responsive person but not from the same person responding on screen. What did work, in a study of 88 thirty-month-olds by Strouse and colleagues in the Journal of Experimental Child Psychology (2018), was a parent sitting beside the child and responding to the person on the screen, which improved word learning from both live and prerecorded video. The American Academy of Pediatrics' 2016 statement "Media and Young Minds" turns this into guidance: under 18 months, avoid screen media other than video-chatting; from 18 to 24 months, choose high-quality content and watch it together; from 2 to 5, about an hour a day of high-quality programming, co-viewed. It names video chat with distant relatives as a form of social connection.

AI narrators: what is known, and what is not

The honest summary is that the evidence on synthetic voices for young children is thin, recent and mostly about something else. Most of what exists studies interactive talking devices, not narration. The strongest study we found is by Ying Xu, Mark Warschauer and colleagues in Child Development (2022). They randomly assigned 117 children aged three to six to hear a story either from an adult or from a smart-speaker conversational agent, each of which either asked dialogic questions during the reading or did not. Dialogic questions raised story comprehension by about half a standard deviation, and it made no significant difference whether the questions came from the person or the machine. Children did, however, say fewer things about the story to the speaker than to the adult. The study's own lesson is that the benefit came from the dialogue, not from who or what supplied it. It is one short-term study of comprehension of a single story; it did not measure relationships, attachment or language learning over time.

Other work looks at how children treat voice assistants. Lauren Girouard-Hallam and Judith Danovitch, in two studies of 80 children each in Developmental Psychology (2022), found that seven- and eight-year-olds were more likely than four- and five-year-olds to turn to a voice assistant for facts, and to a person for personal information. On whether young children can even understand synthetic speech, a small 2012 study of 18 preschoolers by Carrie Pinkoski-Ball, Joe Reichle and Benjamin Munson in the American Journal of Speech-Language Pathology found that intelligibility depended heavily on the voice: children repeated more words correctly from recorded human speech and from a newer synthetic voice than from an older synthesizer. Modern voices are far more natural than those tested in 2012, which is one more reason not to stretch old findings.

The child-focused bodies that have weighed in are concerned with chatbots and AI companions, not narration. UNICEF's guidance on AI and children, updated to version 3.0 in 2025, now addresses AI companions used by children, and the American Academy of Child and Adolescent Psychiatry tells parents that AI companions are not safe tools for children. Neither has anything to say about a synthetic voice reading a picture book. What can be said without new data is a matter of logic, not evidence: the familiar-voice findings above are, by construction, about a person the child knows. A synthetic narrator is not that person, so those findings cannot be claimed for it. Whether a high-quality synthetic voice reading a story does a child any harm, or any good beyond what a stranger's recorded voice would do, has simply not been studied.

The heritage-language angle: who the words come from

For children growing up with a family language that the world around them does not speak, the question of whose voice they hear stops being sentimental. The research on why heritage languages fade, set out on our page on why immigrant families lose their mother tongue, puts the quantity of input first. A smaller body of work suggests that the number of different people supplying it matters too.

Tamar Gollan, Jennie Starr and Victor Ferreira, in Psychonomic Bulletin and Review (2015), tested Chinese-English and Spanish-English university students and Hebrew-English children on naming pictures in both languages. How well they could name things in the heritage language was predicted by how many heritage-language speakers they had interacted with as children, over and above how often they used the language, and without any cost to their English. The authors suggest that several speakers give a child more varied words, more variation in how words sound and more cues for remembering them. Silvia Place and Erika Hoff, in a small 2011 study of 29 Spanish-English two-year-olds in Child Development, found that the number of different speakers predicted English skill beyond raw hours, though for Spanish that link did not survive their controls. Kuhl's social-gating work adds that, at least in infancy, input from a responsive person counts for far more than input from a device.

The people best placed to add speakers and hours in the family language are usually grandparents, aunts and uncles, and in immigrant families they very often live in another country and another time zone. The evidence says what they offer is most powerful live and in conversation. It does not say whether a recording of them adds anything measurable to a child's heritage language; that has not been tested.

What the evidence supports for a recorded family video, and what it does not

Taken strand by strand, the research supports a few modest claims for a video of a relative reading a story, and rules out several larger ones. It supports the claim that the recording carries something a stranger's or a synthetic voice cannot: a voice that the child already knows and that children's brains treat differently from unfamiliar ones (so far shown for mothers' voices), and, on video, a familiar face. It supports the claim that a recording becomes more useful to a toddler when a live adult watches it alongside them and responds to it, as Strouse and colleagues found. And it is consistent with the AAP's advice that, for the youngest children, media should be used together rather than alone.

It does not support any claim that a recorded video is equivalent to live interaction. The language-learning studies, from Kuhl's Mandarin sessions to the video-chat experiments, all point the same way: under about age three, children learn words and sounds from people who respond to them, and much less, sometimes not at all, from a recording of the same people. The stress-hormone findings were for live voices, by phone or in person, and should not be quoted for recordings. A recorded story cannot answer a question, follow the child's pointing finger, or take a turn. And for children under 18 months, the AAP's guidance makes an exception only for video chat.

Three implications follow. For families, a recorded story from a relative is best treated as one more contact with a loved voice, watched with a parent who talks about it, rather than as a substitute for calls, visits and everyday conversation. For anyone building or funding products for young children, the evidence favours designs that bring known people and live adults into the experience over designs that replace them, and it favours measuring outcomes rather than assuming them. For researchers, the open questions are concrete and testable: whether a recorded familiar voice changes a child's attention, comprehension or heritage-language vocabulary compared with a stranger's recording or a synthetic voice, and whether pairing it with a co-viewing adult makes the difference. Until those studies exist, the defensible position is the one the evidence already holds: a familiar voice matters, a responsive person matters more, and a synthetic narrator has neither property.

Related guides

Questions

Do children really respond differently to a parent's voice than to a stranger's?

Yes, and from very early. Newborns in DeCasper and Fifer's 1980 study in Science changed their sucking to hear their mother's recorded voice more often than another woman's. In a 2016 Stanford brain-imaging study in PNAS, 24 children around age ten showed more activity in emotion, reward and face-processing regions when they heard under a second of their mother's voice than when they heard other women. Both studies tested mothers; grandparents and other relatives have not been studied in the same way.

Can a toddler learn language from a recorded video of a relative?

Less than from the same relative live. Studies by Kuhl and colleagues (PNAS 2003), Roseberry and colleagues (Child Development 2014) and Myers and colleagues (Developmental Science 2017) all found that infants and toddlers learned from a responsive person, live or on video chat, but little or nothing from a matched recording. A familiar face helps a little: in one 2010 study toddlers copied more actions from their mother on video than from a stranger. The clearest boost came from a parent co-viewing and responding to the video (Strouse and colleagues, 2018).

Is an AI narrator bad for children?

There is no good evidence either way for story narration specifically. The strongest related study, Xu and colleagues in Child Development (2022), found that a smart-speaker agent asking questions during reading helped 117 children aged three to six understand a story about as much as an adult asking the same questions; the questions, not the partner, made the difference. No study we found compares a synthetic narrator with a familiar relative's recorded voice. What is clear is that the research on familiar voices does not apply to a synthetic one.

Is a video call with a grandparent better than a recorded video?

For learning, in the youngest children, yes. Video chat keeps the back-and-forth that drives early language learning (though one 2018 study found toddlers learned a new word only from a responsive person in the room, not over video chat), and the AAP's 2016 guidance makes video chat the one screen exception for children under 18 months. A recording cannot respond. Where the two are not competing, for instance when a time-zone gap makes calls rare, a recording adds contact with a familiar voice; it does not add conversation.

Why does it matter how many people speak the family language to a child?

Because heritage-language skill tracks the number of speakers, not only the number of hours. Gollan, Starr and Ferreira (Psychonomic Bulletin and Review, 2015) found that how many heritage-language speakers people interacted with as children predicted how well they could name things in that language, beyond how often they used it. Place and Hoff (Child Development, 2011) found a similar speaker-count effect for English in 29 bilingual two-year-olds. Both are modest studies, but they point the same way: more people, more of the language.

Open Azizan — record a story for your child

Updated: 2026-11-03 · By Aida Fard

Hear when azizan is ready

azizan works now in your browser. If you like, leave your email and we will tell you when the full app and the app-store version are ready. Giving your email is optional. Privacy