Volley

Your child's language grows with the back-and-forth. Volley counts it.

Volley hears the back-and-forth with your child on the Bee you already wear, counts it, throws the words away on your own computer , and sends a nudge to your Bee when your best hour goes quiet. The count researchers take with a recorder parents cannot buy, for a few cents a month.

Built for children of talking age, about two to five; the studies behind the number run from infancy to six.

A turn is you saying something and your child answering within five seconds, or the other way round.

See how it countsCount your own

A simulated family, labelled as one

157turns with Asha on 2026-09-28

Generated counts, not a day a Bee heard. On real recordings with no speaker labels, the median session's count was 9.6 percent from the true one, across 54 sessions.

best hour
6pm
longest exchange
7 turns
questions
32
  • 7am19
  • 8am15
  • 9amaway
  • 10amaway
  • 11amaway
  • 12pmaway
  • 1pmaway
  • 2pmaway
  • 3pmaway
  • 4pm11
  • 5pm16
  • 6pm38
  • 7pm31
  • 8pm27

Hours with no child speech are shown as away , not as zero. 1196 words were read and thrown away to make this .

Measured, not claimed: on 26.59 hours of public parent-child recordings we did not make, the rule counted 20,344 turns from the human speaker labels, zero inside any stretch where one person spoke alone ; on the builder's own Bee, hearing a real parent-child conversation voiced by Amazon Polly with no person in the room, Volley counted 66 of 75 turns with no speaker labels at all. Neither is a hand count by a person, and the page says what that leaves .

The weak point is two adults talking: without labels they gave 21 false turns in about three minutes, and the opt-in --only-with-child takes that to 0 and kept 66 of 66 real turns, at the cost of about a sixth of true turns on the corpus (15.7 percent); silence gave 1 false turn, which the check does not remove.

14 to 27%
of the variance in IQ and vocabulary at ages 9 to 14, explained by conversational turns at 18 to 24 months, after adjusting for family income (Gilkerson, Pediatrics 2018)
Turns, not words
predicted activity in the brain's language area, independent of income and IQ (Romeo, Psychological Science 2018)
Coachable
parents shown their own recordings raised turns and their children's language, and the gain held to 30 months (Ferjan Ramirez, PNAS 2020; Huber, Developmental Science 2023)

What it is, why it exists, how it runs.

What

A count of the back-and-forth between you and your child, by the hour, every day, and one thing to try tomorrow.

Why

This is the number that predicted language and IQ a decade later and survived the income adjustment that word counts did not . The instrument that measures it, LENA , is sold to programmes and clinics, not to parents. You already wear the device that can.

How

Your Bee transcribes the day. Volley reads the transcript on your computer, counts by the published rule, deletes the words, and shows the day. Only the counts can leave, and only if you ask.

How it counts, in twenty seconds.

Press play. Each bar is something somebody said; the band after it is the five-second window. The number in the corner is the same rule that runs on your computer, counting these bars as they arrive.

Twenty seconds, counted

5turns so far

adult
child
0s20s

The faint band after each bar is the five-second window . A reply that starts inside it is a turn; two adult bars close together merge into one; six seconds of silence ends the exchange.

  1. 0.0s, adult: a question
  2. 2.4s, child: an answer, 0.8s later
  3. 3.8s, adult: a follow-up
  4. 5.2s, adult: and another, merged into one block
  5. 7.6s, child: an answer, 1.0s later
  6. 14.4s, adult: six seconds of silence first: no turn
  7. 16.6s, child: a reply, 0.7s later
  8. 18.4s, adult: a reply, 0.6s later

A day, an hour at a time, and one thing for tomorrow.

This week, simulated

  • 09-22166
  • 09-23162
  • 09-24156
  • 09-25194
  • 09-26270
  • 09-27241
  • 09-28157

Best hour this week: 7pm, 309 turns. The published target is forty an hour; Volley shows it and sets none of its own .

Tomorrow, written from the latest day's counts

Reading the note from the service.

Rules pick the fact; Claude on Bedrock phrases it; every digit must come from the fact or the template is used . With your consent the sentence goes into your Bee as a reminder at that hour.

This week's plan

The simulated family

Reading the week's plan from the service.

Coaching in the moment, not the morning after.

During the hour the family usually talks most, volley live --nudge watches the Bee's live stream, and if that hour is going quiet it puts a nudge into the Bee then; with --say, a speaker in the room says it aloud, a stand-in for an Echo.

The note above is written for tomorrow. This one arrives while the hour can still be had.

volley live --nudge
volley live --nudge --say

One evening, simulated

usual hour
7pm
usually
44 turns
nudges at
11 or fewer
so far, illustrative
9 turns
The rule, as volley live --nudge runs it: twenty minutes into the hour the family usually talks most, a quarter of that hour's usual turns or fewer puts the nudge in, once a day at most. Here that is 11 or fewer of the 44 usual at 7pm, the week ending 2026-09-28. The hours before it are the simulated family's 2026-09-21; the 9 turns so far are illustrative, chosen under the line, because on no simulated evening did the rule fire.

What it will never say.

It is not a test of your child.

A count, like steps. No score, no percentile, no comparison with other children, ever.

It does not compare you with other families.

The only comparison is with your own days and one published target, stated as its source's figure.

It never keeps a word.

The words are read on your computer, counted, and gone. The server refuses a transcript, and you can watch it do so above.

It does not know what the television was doing.

A 2009 study of 329 children found each hour of audible television cut turns by about a fifth of a standard deviation. Volley cannot see the set; the by-hour view is the only place a quiet hour will show, and it will not say why.

Questions a parent asks.

Is a low day bad?

No. It is a count of one day. The note names your best hour so you keep what already works, and one thing to try tomorrow. That is all it does.

Does it listen to my child?

The Bee on your wrist transcribes what it hears, including your child. Volley reads that transcript on your computer, counts, and deletes the words. Nothing your child said is ever stored by Volley.

What if my child does not talk yet?

Volley counts transcribed words, and a baby's babble may not be transcribed. Until that is measured, Volley is for children of talking age, about two to five. The research instrument counts sounds; this does not, and says so.

Do I need to talk all day?

No. Turns are exchanges, not volume. Ten good minutes at bath time count more than an hour of narration the child does not answer.

What is a good number?

LENA, whose recorder the studies used, publishes forty turns an hour as a target. Volley shows that figure as theirs and sets none of its own.

Who can see this?

You, with your key. The simulated family on this page is public and says so. There is no account with Volley beyond a name you chose.

What about the television?

Volley cannot see it. If an hour is quiet, the by-hour view will show it; what was happening is yours to notice.

Can I delete it?

Yes. The Bee app deletes the conversation; one request with your key deletes the household and every count.

A transcript, refused by the server, while you watch.

The API accepts counts and drops every field it does not know at the door. That is a claim until it is pressed. What the Bee itself keeps is Amazon's promise, in its own privacy notice, and not ours to restate.

Live, from this page

Send a day to the real intake route with a transcript, a name, a doctor, an address and an account number attached, then read back what was stored and search it for every one of them.

Your own day, in ten minutes.

Setup runs on the computer where your Bee is signed in, not on this phone. Send yourself the steps.

  1. 1

    Install Bee's command line tool 

    npm install -g @beeai/cli
  2. 2

    Turn on Developer Mode in the Bee app 

    Settings, then tap the version number five times.
  3. 3

    Sign in the tool to your Bee 

    bee login, then approve it in the app.
  4. 4

    Install Volley 

    pip install serve-return "volley-bee @ git+https://github.com/usv240/volley#subdirectory=apps/ingest"
  5. 5

    Count the day 

    volley setup --blind --child-name Asha
volley today
  6. 6

    See your days in a browser, if you want to 

    volley connect
volley today

Two setup options: volley setup --daily 20:30 counts every evening without typing, and --only-with-child yes sets aside stretches where no voice sounds like a child; the evidence says what that costs.

Recording other people is your responsibility under local law; the setup screen says so before anything runs. The child wears nothing. Your child's words never leave your computer.

Three doors, each for someone.

Parents: your Bee, on your computer

The steps above. The only door that reads a transcript, and it reads it where you are; the counts, and nothing else, can go to a page that is yours.

Anyone curious: a Bee export, in this tab

Drop a file below. It is counted in your browser and uploaded nowhere.

Developers: an API key 

curl -X POST https://6gm5fp5kuipgvjgr5zmhghq63u0hmmvd.lambda-url.us-east-1.on.aws/v1/households \
  -H 'content-type: application/json' \
  -d '{"name":"ours","child_name":"Asha"}'

Then POST counts to /v1/households/{id}/days with X-Api-Key, and read /days, /week, /note. Interactive docs. The schema drops what it does not know; the privacy section shows it.

Count in this tab, upload nothing

Run bee sync on your computer and drop one of the conversation files it writes here. It is read in your browser and counted with the same rule ; there is no upload in this panel at all.

Ask it out loud, and it cannot repeat a word.

An MCP server , spec 2025-11-25 over Streamable HTTP, with seven tools and an Agent Skill . It can say how the talking went today, what to try tomorrow, the week's direction and the best hour; it can record a day of counts; and it explains the rule. It has no tool that returns speech, because the server never receives any, and a test walks every tool to prove it.

Live, from this page

Open a session against the deployed server, list its tools, ask it what Volley counts, then ask how the talking went today for the simulated family, the way Alexa+ would. Every request below is real and timed, and no answer can hold a word of speech, because the server never had one.

By voice, in this browser

A stand-in for Alexa+'s voice, in this browser: the same MCP calls Alexa+ would make. The question is mapped to one of the server's tools, the tool is called live for the simulated family, and this browser reads the answer aloud.

Or tap a question; it makes the same call.

Speech recognition is the browser's own: Chrome sends the audio to its speech service to write it down. Volley's server receives only the tool call, and every tool answers for the public simulated family here.

The agents, the programmes that pay for this count, and every test: the evidence page.

What it refused to count, what it beat, and how close it is to a person.

Written for somebody who came here to find the weak point. Measured here, read from the literature, and not done yet are kept apart on purpose. Every figure below is read from the results the harness committed.

  • Human-labelled recordings

    20,344turns counted from the human labels

    Holds
    What it means

    Told who is speaking, the rule counts the back-and-forth and counts nothing while one person talks alone.

    54 sessions, 26.59 hours of public parent-child recordings we did not make

    0 turns inside 34 stretches where one person spoke alone

  • No speaker labels, families it never saw

    9.6%median error per session

    Holds
    What it means

    This is the count a parent actually gets, since the Bee labels everyone Unknown: close over a session, not exact for a single exchange.

    54 sessions; who is the child is read from the words and the pauses, as the Bee requires

    49 of 54 sessions within 20 percent of the true count

    One dot per session; filled when within 20 percent.

  • On the Bee itself

    66 of 75turns, with no labels

    Holds, once
    What it means

    The whole chain works on the device and runs a little low. One session is a check, not a study.

    One real conversation played to the band for 190 seconds, no person in the room

    -12 percent, on one session

  • Negative controls: no child there

    21false turns from two adults talking

    Weakness
    What it means

    Without speaker labels, adults talking to each other look like turns. This is the weak point, stated.

    About three minutes each, played to the Bee, against 66 turns from a real parent and child of the same length

    One adult voice 14, silence 1

    • Parent and child66
    • Two adults21
    • One adult voice14
    • Silence1

  • The child-present check, opt-in

    21 to 0false turns from two adults, with the check on

    Trade-off
    What it means

    It ends the false turns at a cost in real ones, so the parent chooses which error they would rather have.

    Off by default; on with volley setup --only-with-child yes

    One adult voice 14 to 0, and 66 of 66 real turns kept on the Bee. On the corpus, held out by family, it refused 70.6 percent of adults-only stretches and lost 15.7 percent of true turns.

    • Parent and child66
    • Two adults0
    • One adult voice0

  • Not measured yet

    A child under three on this device; a family other than the builder's; the Bee beside a LENA recorder on the same afternoon.

    Open
    What it means

    The last is the study to run first with a grant. Until then the count is checked, not validated.

    Said here so nobody has to find it

The weak point in full: one voice, two adults, and the opt-in check

On the public corpus, 0 turns inside its 34 stretches of a minute or more where one person spoke alone , from the human speaker labels. Without labels it is the weak point: two adults from a real family dinner, no child, played to the Bee for 183 seconds, gave 21 false turns, and one adult voice gave 14, against 66 for a real parent and child of the same length. On the corpus, removing the child leaves false turns at 43.1 percent of the real count.

A check that a child is present now ships as an opt-in setting, volley setup --only-with-child yes. On the Bee it kept 66 of 66 turns of the real conversation, took two adults from 21 false turns to 0, and one adult voice from 14 to 0. On the corpus, held out by family, it refused 70.6 percent of adults-only stretches and lost 15.7 percent of true turns. That loss is the trade, and it is why the check is a choice and not the default. The Bee's own speaker label would end this.

Against the obvious alternatives: adult word count and a bare speaker-change count

Adult word count, which the last consumer device sold: across the corpus's 54 sessions it ranks with turns at Spearman 0.51, and two sessions with the same adult words carried turns 3.61 times apart . A speaker-change count with no window agreed with the rule on 99.3 percent of alternations, because the corpus's aligned timing has no gaps ; the window's work is in a day, not a lab session.

No speaker labels: how roles are read, and the run on the Bee

The Bee labels every speaker “Unknown”, so Volley reads who is the child and who the adult from the words and the pauses, with a model trained on CHILDES and scored only on families it was not trained on. Across 54 sessions it counted 20,673 turns against 20,344 true (+1.6 percent); the median session was within 9.6 percent, and turns per hour tracked the truth at a correlation of 0.874. Treating every pause as a change of speaker missed by 87.3 percent. Misread roles merge exchanges, so a raw blind count runs low by a steady 11 to 15 percent; Volley scales for it with a factor learned only from other families, which took the median session error from 12.2 to 9.6 percent. Each utterance's role was right 73.7 percent of the time, so a single exchange can be misread; a day, a week and a trend are what it is for.

On the Bee itself. A real mother-and-child conversation from the same corpus, voiced by Amazon Polly with an adult and a child voice at its original timing, played to the band for 190 seconds, no person in the room. The Bee wrote 124 utterances for 125 played and kept 75 of 75 turns: its segmentation loses none. Volley, with no labels and a model that never saw this family or its scale, counted 66 (-12 percent), with 76.6 percent of lines given the right speaker. One session; the corpus says how much one session varies.

The public corpus, every rate, measured 2026-09-30

The CHILDES Gleason corpus: 54 sessions, 26.59 hours, 24 children aged 2;1 to 5;2 with their mother, with their father and at the family dinner table, human speaker labels, aligned timing. The rule counted 20,344 turns: 171.2 to 1326.1 an hour by session, median 808.3. Dinner 672.4, with the mother 710, with the father 892 an hour. Children under three, 776.3 an hour across 13 sessions; three and over, 761.3 across 41. A fifty-second window instead of five changed the count to 20,483. 13 files had no timing and were skipped by name.

What this shows: the rule, the roles and the alternation logic hold on human-labelled recordings at twenty times the published daylong target, which is what half an hour of a parent's full attention looks like. What it does not show: the window, because the aligner's timing is contiguous; the device's transcription of a two-year-old, because these are human transcripts; and any family's day. No person has hand-counted a day on the Bee, and that count is what would answer the last two . volley verify recomputes every figure in this paragraph from the committed results, and from the corpus itself where it is present.

Read from the literature: every study on this page
  • Gilkerson J et al. Language Experience in the Second Year of Life and Language Outcomes in Late Childhood. Pediatrics 2018. PMID 30201624. Three authors work for the LENA Foundation.
  • Romeo RR et al. Beyond the 30-Million-Word Gap. Psychological Science 2018. PMID 29442613.
  • Ferjan Ramirez N, Lytle SR, Kuhl PK. Parent coaching increases conversational turns and advances infant language development. PNAS 2020. PMID 32015127.
  • Huber E et al. Parent coaching from 6 to 18 months improves child language outcomes through 30 months of age. Developmental Science 2023. PMID 36999222.
  • Christakis DA et al. Audible television and decreased adult words, infant vocalizations, and conversational turns. Archives of Pediatrics and Adolescent Medicine 2009. PMID 19487612.
  • Romeo RR et al. Language Exposure Relates to Structural Neural Connectivity in Childhood. Journal of Neuroscience 2018. PMID 30104336.
  • Romeo RR et al. Neuroplasticity associated with changes in conversational turn-taking following a family-based intervention. Developmental Cognitive Neuroscience 2021. PMID 34052580.
  • Suskind DL et al. A parent-directed language intervention for children of low socioeconomic status: a randomized controlled pilot study. Journal of Child Language 2016. PMID 26041013.
  • Leung CYY, Hernandez MW, Suskind DL. Enriching home language environment among families from low-SES backgrounds: A randomized controlled trial of a home visiting curriculum. Early Childhood Research Quarterly 2020. doi 10.1016/j.ecresq.2018.12.005.
  • Elmquist M et al. Parent education to improve early language development: A preliminary evaluation of LENA Start. Journal of Child Language 2021. PMID 32921333.
  • Wong K, Thomas C, Boben M. Providence talks: A citywide partnership to address early childhood language development. Studies in Educational Evaluation 2020. doi 10.1016/j.stueduc.2019.100818.
  • Gomez E, Strasser K. Language and socioemotional development in early childhood: The role of conversational turns. Developmental Science 2021. PMID 33825290.
  • Sperry DE, Sperry LL, Miller PJ. Reexamining the Verbal Environments of Children From Different Socioeconomic Backgrounds. Child Development 2019. PMID 29707767; and the reply, Golinkoff RM et al., PMID 30102419.
  • Reed J, Hirsh-Pasek K, Golinkoff RM. Learning on hold: Cell phones sidetrack parent-child interactions. Developmental Psychology 2017. PMID 28650177.
  • Radesky JS et al. Patterns of mobile device use by caregivers and children during meals in fast food restaurants. Pediatrics 2014. PMID 24616357.
  • Cristia A et al. A thorough evaluation of the Language Environment Analysis (LENA) system. Behavior Research Methods 2020. PMID 32728916.
  • Ferjan Ramirez N, Hippe DS, Kuhl PK. Comparing Automatic and Manual Measures of Parent-Infant Conversational Turns: A Word of Caution. Child Development 2021. doi 10.1111/cdev.13495.
  • Norbury CF et al. The impact of nonverbal ability on prevalence and clinical presentation of language disorder. Journal of Child Psychology and Psychiatry 2016. PMID 27184709.
  • Council on Early Childhood. Literacy promotion: an essential component of primary care pediatric practice. Pediatrics 2014. PMID 24962987. Garner A, Yogman M. Preventing Childhood Toxic Stress: Partnering With Families and Communities to Promote Relational Health. Pediatrics 2021. PMID 34312296.
  • Harvard University Center on the Developing Child. Serve and Return.
  • The corpus: Gleason JB, CHILDES English Gleason Corpus, doi 10.21415/T5101R, used under the TalkBank ground rules , and cited as they require through Masur EF, Gleason JB. Parent-child interaction and the acquisition of lexical information during play. Developmental Psychology 1980. doi 10.1037/0012-1649.16.5.404. The database, cited as the rules also ask: MacWhinney B. The CHILDES Project: Tools for analyzing talk. Third edition. Mahwah, NJ: Lawrence Erlbaum Associates, 2000; CHILDES is supported by NICHD grant HD082736. Retrieved and analysed 2026-09-30. Playlogue (Kalanadhabhatta et al., IMWUT 2024), the corpus first planned, is gated behind the same rules on Hugging Face; the harness for it is in the repository and runs the day access opens.

Every test, and the commands that run it: the evidence page.