> For the complete documentation index, see [llms.txt](https://api-docs.speechace.com/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://api-docs.speechace.com/use-cases/voice-ai-for-early-literacy.md). # Voice AI for Early Literacy Voice activities let young children read and speak out loud and get instant, low-stakes feedback — the kind of frequent, low-pressure practice that helps build early reading skills. This guide comes in two parts: **Part 1** covers *designing* children's voice activities responsibly (why voice, what AI should and shouldn't do, COPPA, and the realities of child speech), and **Part 2** covers *building* them with the Speechace API (endpoints, recording quality, interference handling, and feedback design). ## Part 1 — Designing voice activities for children ### Why voice, and what to build Reading starts as an oral skill, so speaking is the most natural way for a young child to practice it. Voice activities are engaging and playful, and — just as importantly — they let a child rehearse the exact thing reading assessments measure: saying letters, sounds, and words out loud. Research in the science of reading consistently points to the same foundations — phonemic awareness, phonics, and oral reading fluency — and to the value of **practice with immediate, corrective feedback, early and often**. Voice AI makes that practice scalable: every child can practice as much as they want and be screened frequently, without waiting for scarce one-on-one adult time. For accuracy metrics on children's speech, see [Speechace's results on child datasets](https://www.speechace.com/using-the-speechace-api-as-voice-ai-for-kids/). Typical activity types: * **Phonics** — letter names and letter sounds. * **Word practice** — real words and nonsense (decodable) words. * **Multiple choice** — the child picks and says one option. * **Oral reading fluency (ORF)** — reading a passage aloud. Part 2 maps each of these to a Speechace endpoint. ### What AI should — and shouldn't — do Voice AI is a **low-stakes helper**. Use it to: * let children practice as much as they want, with instant feedback; * screen **every** child frequently, not just those already flagged; * give parents and teachers more signal than they could ever gather one child at a time. Do **not** use it to make **high-stakes decisions on its own** — a child's grade, their reading level, or whether they need intervention. An automated score is one input among many. A **human — teacher, reading specialist, or parent — is always the reviewer and the decider.** Build the product so the AI *surfaces and flags*, and a person *confirms and decides*. {% hint style="info" %} A useful rule of thumb: the AI can **recommend a second look**; only a person should **make a determination**. {% endhint %} ### A word on COPPA and compliance An app that records a child's voice is collecting **personal information from a child**, which carries legal obligations. In the US, the **Children's Online Privacy Protection Act (COPPA)** generally requires **verifiable parental consent** before collecting personal information from children under 13 — or, in a school setting, the school may provide that consent on parents' behalf for educational use. A recording of a child's voice counts as personal information. Practical implications for your app: * Obtain and record **parental or school consent** before you capture any audio. * Collect **only what you need**, and be clear with families about what you collect and why. * Have a **retention and deletion policy** (see [Data minimization, retention, and deletion](#data-minimization-retention-and-deletion)). * Remember that **your app is the party responsible** for the child's data; Speechace processes audio on your behalf under the terms of your contract. {% hint style="warning" %} This section is general information, **not legal advice**. Requirements vary by jurisdiction and change over time — consult qualified counsel and review the [FTC's guidance on children's privacy](https://www.ftc.gov/business-guidance/privacy-security/childrens-privacy). {% endhint %} ### The challenges of child speech Child speech is harder to capture cleanly than adult speech — not only because young voices have different acoustic characteristics, but because of **how and where children record**. Expect the following, and plan for them deliberately rather than being caught off guard: * **Noisy classroom** — other children speaking right next to the recorder. * **Noisy household** — a sibling nearby, or the child working in a busy family room. * **A helping adult** — a parent or aide talking to the child, or whispering the answer, while the mic is on. * **The child says more than intended** — "I don't know," "this word is…," repeating the word several times, or giving both the letter *name* and the letter *sound* when only one was asked for. Each of these puts **extra speech** into the recording that the scoring engine did not expect. Part 2 covers how to design around them and how to detect them with the API. ## Part 2 — Building it with the Speechace API ### Activities and endpoints
ActivityRecreate withTry a demo
Wordsscore/worddemo
Letter namesscore/word (word="B")demo
Letter soundsscore/word (arpa_mark, e.g. [b]{b ah0})demo
Nonsense words
e.g. "shuzzle"		score/word (arpa_mark)demo
Multiple Choicescore/word (Single word options)
score/text (Phrase/Sentence options)		demo
Oral Reading Fluency (ORF)score/text with include_fluency=1demo
Fluency scoring reports words-correct-per-minute, speaking rate, pausing, hesitations, repetitions, and intonation — see [score/text](/api-reference/score-text.md#fluency-metrics). ### Recording guidelines and best practices Good scores start with good recordings. Given the child-speech challenges above, apply a few rules to the **capture experience**: * **Explicit mic On/Off.** Let the child (or a helping adult) start and stop recording deliberately. Don't auto-start or auto-stop on silence — a young child's pauses, false starts, and noisy surroundings make voice-activated capture unreliable. * **Don't over-record.** Match the recording window to the task. If the activity is a single word or a letter, capture a second or two — not a 15–25-second window that invites extra talk. * **Give clear audio *****and***** text instructions.** Tell the child exactly what to say, and tell any nearby adult to let the child say it **once, on their own**. Pair spoken instructions with on-screen text so both the child and the adult know what's expected. * **Add a practice / calibration step.** Before a scored attempt, run a quick practice turn and verify the child is speaking the intended utterance — and *only* that — using interference detection (below). {% hint style="info" %} **The shorter the utterance, the more a little interference hurts.** A stray background word barely moves the score on a sentence, but it can dominate a single sound. Cost of interference, lowest to highest: **sentence → multi-syllable word → single-syllable word → letter name → letter sound** For the shortest utterances — letter names and especially letter sounds — be strictest about a clean, single-voice recording. {% endhint %} ### Detecting interference and acting on it Speechace can measure how much **competing speech** a recording contains. Pass `include_interference_metrics=1` to [score/word](/api-reference/score-word.md) or [score/text](/api-reference/score-text.md) and the response includes an `interference_ratio` — a higher value means more speech beyond the intended utterance. See [Detecting Speech Interference](/interpreting-results/detecting-speech-interference.md) for the value bands and recommended code patterns. Use it to: * **Verify the calibration step** — confirm the child is speaking only the intended utterance before you begin scoring. * **Caution or retry on a noisy attempt** — when interference is high, prompt a redo in a quieter setting instead of trusting the score. * **Flag patterns, not just single recordings.** If a child shows **high interference across many or all** of their recordings, that usually reflects their *environment* — a loud classroom, a talking adult — rather than their reading. Flag that child's results for human review rather than acting on the scores. ### Handling no speech and invalid attempts Children will sometimes submit nothing usable — silence, a cough, "I don't know," or a tap that captures an empty clip. Treat these as **invalid attempts, not wrong answers**: * Detect empty or no-speech responses (for example, an [`error_no_speech`](/getting-started/error-handling.md) result) and offer a **graceful retry** with a friendly re-prompt. * **Never let an invalid attempt lower a child's activity or quiz score.** A missed recording is a capture problem, not a reading mistake — exclude it, don't average it in. ### Feedback: what to show a child vs. an adult reviewer The same result should be presented very differently depending on who is looking. **To the child** — keep it simple, encouraging, and immediate: * positive, specific praise plus a single thing to try next ("Great reading! Let's try the /b/ sound one more time"); * let them **replay their own recording** and hear the target pronunciation; * friendly, visual cues rather than numbers, score bands, or phoneme jargon; * never a bare score or a big red "wrong." **To the adult reviewer (teacher, specialist, or parent)** — give the full picture: * word-, syllable-, and phoneme-level quality; and for ORF, words-correct-per-minute and the other fluency metrics; * **interference levels** (especially if high) and the ability to **replay the audio** and judge for themselves; * results **across sets of activities and over time**, not a single attempt (see [Measuring progress](#measuring-progress)); * the ability to **override the AI, redo an activity, or re-run a screening**. ### Review dashboard: keep a human in the loop Give teachers and parents a place to **review, not just receive** results. A useful reviewer dashboard lets them: * see a child's results across activities, with the AI's flags surfaced; * **replay the original audio** for any attempt; * **redo an activity or re-run a screening** when a recording was compromised; * make and record the **final decision**. This is what keeps low-stakes screening low-stakes: the AI does the volume, and a person makes the calls. ### Equity and fairness Children speak with different accents and dialects, with developing articulation, and many are multilingual. Don't let a valid pronunciation be marked wrong: * Use [**markup language**](/reference-data/markup-language.md) (`arpa_mark`) to specify the exact phonemes you'll accept — including **alternate pronunciations** of a word, so a legitimate dialectal variant is scored as correct. * Use **multiple choice** ([score/word](/api-reference/score-word.md#request-examples) / [score/text](/api-reference/score-text.md#request-examples)) when more than one spoken answer should count as correct. * Route **flagged or borderline** results to a human reviewer rather than deciding automatically. ### Measuring progress Judge a child on a **body of work, not one recording.** A single attempt is noisy — a cough, a distraction, or a loud room can sink it. Score across **sets of activities and over time**, watch **trends** rather than one number, and prefer frequent low-stakes checks to a single high-stakes attempt. This is fairer to the child and more useful to the teacher. ### Data minimization, retention, and deletion You are responsible for the children's data your app collects. Keep the two sides straight: * **Speechace's side.** Data you send to the API is **yours** — Speechace only processes and retains it to serve your request and support queries, and you can set a specific retention period, **including zero retention**, in your contract (see [Data Retention](/resources/data-retention.md)). Retention expiry runs automatically and continuously, so a per-record **deletion request to Speechace is effectively a no-op** — data ages out on its own under your policy. * **Your side.** You are responsible for retaining and deleting the audio and results **in your own systems** — storage, dashboards, and backups. Honor parent and school **deletion requests** on your side, collect only what you need, and delete (or only briefly retain) audio once it has been scored. ### Device and environment recommendations You can't control every setting, but you can nudge toward clean audio: * Recommend a **headset or close-talk microphone** over a far-field laptop or tablet mic. * Have the child record in the **quietest available spot**, away from other speakers. * **Check the microphone during onboarding** so a bad device is caught before it skews scores. ### Be transparent Tell parents, teachers, and schools that scoring is **AI-assisted** and that a human makes the final call. Transparency builds trust — and reinforces the human-in-the-loop model that keeps low-stakes screening low-stakes.