How SpeechStep’s AI Feedback Works — and What It Doesn’t Do

✓ Reviewed against ASHA Practice Portal, Peer-reviewed speech-technology researchEvidence level BLast reviewed September 29, 2026Published September 29, 2026

SpeechStep’s AI listens to what your child actually said and compares each sound against how it should be produced. Where that feedback is validated — the R sound today — your child gets an instant clear-or-try-again response. Everywhere else, practice is listen-and-compare with optional human review. It never diagnoses, and it never invents a score.

The short answer

When your child says a practice word, SpeechStep’s AI listens to the actual recording — not just whether something was said — and compares each sound in it against how that sound should be produced. Where that judgment is validated for children, your child sees an instant, encouraging clear-or-try-again response. Where it isn’t validated yet, SpeechStep shows no AI verdict at all: practice becomes listen-and-compare, and a human can review the recording instead.

📝

The rule behind every design decision here

No verdict without validation. A wrong “great job!” teaches a child the wrong sound, and a wrong “try again” teaches them to distrust practice. Where we cannot prove the AI reliable, we would rather say nothing than guess.

What the AI actually does

SpeechStep’s scoring builds on two well-established lines of speech research. The listening layer is a modern self-supervised speech model of the wav2vec 2.0 family, which learns the fine acoustic structure of speech from raw audio. On top of it, we use goodness-of-pronunciation style scoring — a method introduced by Witt and Young in 2000 and refined by two decades of pronunciation-assessment research — which asks, for each sound in the word: how closely does what this speaker produced match the target sound?^1,2

Because the analysis is per-sound rather than whole-word, the feedback can respond to what your child actually said. If a child says “wabbit,” the system isn’t just hearing “not rabbit” — it can locate the R and see that something W-like was produced in its place. That is also what powers the typed error analysis in our free S-sound check, which distinguishes a frontal (tongue-forward, “th”-like) pattern from a lateral (“slushy”) one — a distinction that matters, because the two are handled differently.^2,3

The example recordings your child practices against come in four pronunciation profiles — American, British, Indian and Australian English — so a family in Mumbai or Melbourne practices against the English they actually speak, and the target for “correct” is their own variety, not someone else’s.⁵

Where AI verdicts apply today

Instant AI feedback is on for a sound only when that sound has passed our validation for parent-facing use. Today that is the R sound. This list is deliberately short: it grows only when the evidence does, and the app reads it from a capability manifest rather than a marketing page — so what the product does and what this page says cannot quietly drift apart.

What feedback your child gets, by sound, today.
Practice target	What your child sees
R sound	Instant AI feedback: clear, or try again — plus listen-and-compare
Every other sound	Listen-and-compare: a clear example beside their own try — no AI verdict
Any recording you share	Optional human review; the reviewer’s judgment is shown beside the AI’s

What feedback your child gets, by sound, today.

Why so careful? Children’s speech is the hard case for speech technology. Vocal tracts are smaller, sounds are still developing on a schedule that runs to age 6 and beyond, and a typical 4-year-old error can look exactly like the “wrong” answer an adult model expects. Judging a child’s developing R against adult speech norms would be both technically easy and clinically wrong.^6,5

What the AI will never tell you

It will never diagnose. A speech sound disorder is diagnosed by a licensed speech-language pathologist after a comprehensive evaluation — not by an app, ours included.
It will never score sentences today. We built sentence-level scoring, measured it against a public dataset of children’s speech, and it did not meet our reliability bar — so it does not ship. Sentences your child tries are kept as a speech sample for human ears.
It will never invent feedback. No verdict appears for a sound the AI is not validated on; there is no fake percentage, no decorative star rating, and no score generated to fill space.
It will never be the last word. Any recording a family shares can be reviewed by a human, and when a reviewer judges it, their judgment is what the family sees — beside the AI’s, never silently replacing it.

These boundaries are not modesty; they are the ASHA scope-of-practice line drawn where it actually sits. Screening and practice support are things software can honestly do. Diagnosis and treatment planning are not.^4,7

Where humans come in

AI feedback is one voice in the loop, not the authority. Families can opt in to human review of practice recordings, and a connected speech-language pathologist can listen to whatever the family chooses to share, judge it themselves, set the practice plan, and write their own summary of progress. In every case the human’s judgment is displayed as the human’s, the AI’s as the AI’s — labeled, side by side.^4,9

And when what we hear suggests a professional evaluation — persistent errors past the expected age, speech that’s hard to understand, anything beyond a common sound error — SpeechStep’s job is to say so plainly and point you to one, for example through ASHA’s clinician directory.^8,10

Hear the difference yourself

Try one word with instant AI feedback and one with listen-and-compare — free with a parent account, in your browser.

Try it free →

Frequently asked questions

How accurate is SpeechStep’s AI?＋

Accurate enough to ship for the sounds where we can prove it, and honest enough to say nothing where we can’t. SpeechStep gives instant AI verdicts only where the sound-and-accent combination has passed our validation — the R sound today. For every other sound, the app does not pretend: practice is listen-and-compare, where your child hears a clear example next to their own try, with the option of a human review.

Which sounds get AI feedback?＋

Instant clear-or-try-again AI feedback is validated for the R sound today. Every word of every other sound still gets a full practice loop — a clear example recording, your child’s own attempt played back beside it, and optional human review — but no AI verdict, because we have not yet proven the AI reliable enough on those sounds to judge a child. As more sounds pass validation, they graduate to instant feedback.

Why doesn’t SpeechStep show an AI result on every sound?＋

Because an unproven verdict is worse than none. Children’s speech is genuinely hard for machines: small vocal tracts, developing sounds, and errors that differ from adult mistakes. Rather than show a confident-looking mark that might be wrong, SpeechStep limits AI verdicts to the sounds where validation supports them, and uses listen-and-compare — plus human review — everywhere else.

Does SpeechStep diagnose speech disorders?＋

No, and it never will by itself. Diagnosis of a speech sound disorder requires a comprehensive evaluation by a licensed speech-language pathologist — case history, structured testing, oral-motor examination and clinical judgment. SpeechStep screens, coaches, and organizes practice. When what it hears suggests a professional should be involved, it says exactly that.

Does the AI score my child’s sentences?＋

No. We built sentence-level scoring and measured it against a public dataset of children’s speech — and it did not meet our reliability bar, so we do not ship it. In the assessment, the sentences your child tries are collected as a speech sample a human can listen to, not scored by the machine. When we can score connected speech reliably, we will publish that change here first.