Skip to main content

The Talking Book

Status: in development. Reference implementation measured on the bench; no production units.

A page that answers back. A child presses the button in the page, asks their question out loud, and the book answers in plain language at their level. No screen. No cloud round trip. The answer comes from a small model on the chip inside the cover.

How it works​

PartWhat it isWhy
The bookPrinted picture book, hardware in the coverSlots into a shelf-space habit families already have
The buttonPush-to-talk, in the pageMic is off until pressed. Privacy and power in one part
The chipMicrocontroller-class (ESP32-S3), under ten dollarsRuns the model on-device. No server, no subscription. Measured numbers
The modelSmall, bounded to the book's contentNarrow scope means faster, cheaper, and more coherent answers
Voice in, voice outSpeech recognised and spoken on the deviceRaw audio is processed on the chip and discarded there

The one product law: off until a button is pushed. The microphone is physically dormant until the button is held. No wake word, no ambient listening. That single requirement delivers the privacy claim, months of battery life, and cheaper hardware at once.

Why a book that talks back​

The developmental value of reading to a young child comes from the back-and-forth around the book, not the book alone. That mechanism has 35 years of experimental evidence behind it.

Dialogic reading​

The founding experiments are Whitehurst and colleagues. A one-month intervention in which parents were trained to ask open-ended questions, expand on the child's answers, and let the child lead produced measurable expressive-language gains in two-year-olds over a control group that read the same books the ordinary way.

  • Whitehurst et al. (1988). Accelerating language development through picture book reading. Developmental Psychology 24(4). 10.1037/0012-1649.24.4.552
  • Whitehurst et al. (1992). Systematic extension to Mexican day care. Developmental Psychology 28(6). 10.1037/0012-1649.28.6.1106
  • Whitehurst et al. (1994). Replication and extension to a videotape training format. Journal of Educational Psychology 86(2). 10.1037/0022-0663.86.2.235

Meta-analysis. Mol, Bus, de Jong & Smeets (2008), Added Value of Dialogic Parent-Child Book Readings: A Meta-Analysis. Early Education and Development 19(1). 10.1080/10409280701838603. Sixteen studies. For expressive vocabulary (k = 9, n = 322) dialogic reading beat reading-as-usual with Cohen's d = 0.59, a moderate effect, strongest for 2 to 3 year olds and reduced for 4 to 5 year olds and for children at risk for language impairment.

Broader shared-reading base. Bus, van IJzendoorn & Pellegrini (1995), Joint Book Reading Makes for Success in Learning to Read. Review of Educational Research 65(1). 10.3102/00346543065001001. Frequency of parent-preschooler book reading predicts language growth, emergent literacy, and reading achievement, overall d = 0.59 (about 8% of outcome variance), independent of family SES.

d = 0.59dialogic reading vs reading as usual, expressive vocabulary (Mol 2008)
2 to 3the ages where the effect is strongest
16studies in the meta-analysis

Conversational turns, not word counts​

Two 2018 studies moved the field from "how many words a child hears" to "how many back-and-forth exchanges a child has."

  • Romeo et al. (2018). Beyond the 30-Million-Word Gap: Children's Conversational Exposure Is Associated With Language-Related Brain Function. Psychological Science 29(5). 10.1177/0956797617742725. In 36 SES-diverse 4 to 6 year olds, the number of conversational turns with adults, independent of SES, IQ, and adult word count alone, predicted greater Broca's area activation, which in turn explained the link between language exposure and verbal skill.
  • Gilkerson et al. (2018). Language Experience in the Second Year of Life and Language Outcomes in Late Childhood. Pediatrics 142(4). 10.1542/peds.2017-4276. Daylong recordings of 146 infants and toddlers. Conversational turn counts at 18 to 24 months accounted for 14% to 27% of the variance in IQ, verbal comprehension, and vocabulary at ages 9 to 14, after controlling for SES. Adult word counts weakened considerably once SES was controlled.

Harvard's Center on the Developing Child packages the same finding as serve and return: the child serves (points, babbles, asks), the adult returns.

What this means for the product. Every button press is a serve. Every answer is a return. The book is engineered around turns, which is the variable the research says matters. It works the language-comprehension strand of Scarborough's Reading Rope, the strand built mostly by talking and being read to.

Why no screen​

  • Print beats the tablet for toddlers. Munzer et al. (2019), Differences in Parent-Toddler Interactions With Electronic Versus Print Books. Pediatrics 143(4). 10.1542/peds.2018-2012. Counterbalanced lab study of 37 parent-toddler pairs across print, basic electronic, and enhanced electronic formats. Parents produced more dialogic talk with print (11.9 vs 6.2 enhanced, p < .001); toddlers produced more book-related talk and scored higher on collaboration with print. Enhanced e-books did worst.
  • Four pillars of learning. Hirsh-Pasek et al. (2015), Putting Education in "Educational" Apps. Psychological Science in the Public Interest 16(1). 10.1177/1529100615569721. Anything that claims to teach should be active, engaged, meaningful and socially interactive. A physical book with a spoken back-and-forth hits all four without a screen.
  • Guided play. Zosh et al. (2018), Accessing the Inaccessible: Redefining Play as a Spectrum. Frontiers in Psychology 9. 10.3389/fpsyg.2018.01124. Guided play, where the child leads within a structure an adult designed, outperforms both free play and direct instruction for learning goals.
  • Paediatric guidance. AAP Council on Communications and Media (2016), Media and Young Minds. Pediatrics 138(5). 10.1542/peds.2016-2591. Superseded in 2025 by Digital Ecosystems, Children, and Adolescents (10.1542/peds.2025-075320), which keeps the framing that digital products must be designed around children's developmental needs rather than engagement.

Why on the chip​

The two best-known AI toys for this age group show the two ways cloud-dependent design fails.

Curio (Grok, Gabbo, Grem plush). From Curio's published privacy policy, read 2026-09-06:

"Audio files generated from a Voice Interaction are not stored; the audio streams through our systems for transcription and is then discarded. Transcriptions of Voice Interactions, including any personal information your child provides to the Device, are retained for up to 90 days."

Listed third-party operators that "may collect or maintain personal information from children through the App or Device": Kids Web Services, Azure Cognitive Services, OpenAI, Perplexity AI. That is a cloud-dependent toy, by design. The child's voice leaves the house every time.

Moxie (Embodied). Shut down December 2024 after funding was withdrawn. Because the robot ran its conversation through cloud language models, devices stopped working within days of the company closing. No refunds. Coverage.

Our architecture. The raw audio of a child never leaves the book. Inference happens on the chip and the audio is discarded immediately. There is no server, so there is nothing to shut off and nothing to subscribe to. Full data path: How the data path works.

Where the evidence stops

The dialogic-reading and conversational-turn literature is about adults talking with children. No published trial yet tests an on-device generative book with preschoolers. We claim the mechanism and we design to it. Measuring the result is what the pilot is for.

Bounded models are small. The book will not answer every question a child asks. It is designed to answer well inside its scope and to say "let's ask a grown-up" outside it. The book does not replace the adult; it extends the number of turns a book can generate when an adult is not free.

Sources​

  1. Whitehurst et al. 1988, Developmental Psychology. https://doi.org/10.1037/0012-1649.24.4.552
  2. Whitehurst et al. 1992, Developmental Psychology. https://doi.org/10.1037/0012-1649.28.6.1106
  3. Whitehurst et al. 1994, Journal of Educational Psychology. https://doi.org/10.1037/0022-0663.86.2.235
  4. Mol, Bus, de Jong & Smeets 2008, Early Education and Development. https://doi.org/10.1080/10409280701838603
  5. Bus, van IJzendoorn & Pellegrini 1995, Review of Educational Research. https://doi.org/10.3102/00346543065001001
  6. Romeo et al. 2018, Psychological Science. https://doi.org/10.1177/0956797617742725
  7. Gilkerson et al. 2018, Pediatrics. https://doi.org/10.1542/peds.2017-4276
  8. Harvard Center on the Developing Child, Serve and Return. https://developingchild.harvard.edu/key-concept/serve-and-return
  9. Munzer et al. 2019, Pediatrics. https://doi.org/10.1542/peds.2018-2012
  10. Hirsh-Pasek et al. 2015, Psychological Science in the Public Interest. https://doi.org/10.1177/1529100615569721
  11. Zosh et al. 2018, Frontiers in Psychology. https://doi.org/10.3389/fpsyg.2018.01124
  12. AAP 2016, Pediatrics. https://doi.org/10.1542/peds.2016-2591 · AAP 2025. https://doi.org/10.1542/peds.2025-075320
  13. Curio privacy policy. https://heycurio.com/privacy
  14. Aftermath, Embodied/Moxie shutdown. https://aftermath.site/moxie-robot-ai-dying-llm-embodied/