Why ordinary speech recognition fails here
Commercial speech engines are trained on adults who pronounce things correctly, and they are built to repair errors β to infer the word you meant. In speech therapy that is exactly backwards. The error is the signal. If the engine helpfully decides the child said "sun" when they actually said "thun", the entire clinical value is gone.
Three genuinely hard engineering problems
Each one forced a decision that departs from the standard approach.
A child's voice is not a small adult voice
Fundamental frequency is nearly double, formants shift, and speaking rate is wildly inconsistent. Models trained on adults degrade sharply below age eight. That is not a calibration problem β it is a training data problem, and the only fix is the expensive one.
"Correct" is a range, not a point
Accent and regional dialect mean a sound can be produced several legitimate ways. Score against a single reference and you flag half of a perfectly typical region as disordered. The target has to be a distribution, and the engine has to know which one applies.
Recorded in a living room
Television in the background, a younger sibling, a tablet microphone. The engine has to separate genuine mispronunciation from a bad recording β and when it cannot tell the difference, stay quiet rather than deliver confident nonsense.
The pipeline, stage by stage
What happens between the child speaking and feedback appearing on screen.
Capture and gate
Voice activity detection, noise reduction, and a quality check. A recording that fails the threshold asks for another try instead of entering analysis and producing a wrong answer.
Transcribe and align
The acoustic model emits the phoneme sequence that was actually produced β not the word that was expected. Forced alignment maps it against the target transcription and marks where the two diverge.
Score each sound
Every phoneme gets a confidence score and an error classification: substitution, omission, insertion, or distortion. This is the difference between "not quite right" and "the /s/ becomes /ΞΈ/ in word-final position".
Choose the next task
An adaptive selector picks what comes next from the weakest sound, the most recent success rate, and time since last attempt β keeping difficulty at a level the child can still beat.
Summarise for the clinician
Rather than a stream of alerts, one weekly report: what improved, what stalled, and which three minutes of audio are worth listening to before the next session.
Privacy is an architectural decision
These are voice recordings of children. That shapes the design from day one β not at the end when a lawyer gets involved.
- On-device processing by default β acoustic analysis runs locally. What leaves the device is results, not audio
- Storing audio is an explicit choice β off by default, enabled only with parental consent for a clinical reason
- Deletion that means deletion β a deletion request clears backups too, within a defined window
- Identity separated from data β recordings carry a one-way identifier, never a name
- No training on patients β model improvement uses only data gathered under separate, purpose-specific consent
What the engine does not do
It does not diagnose, does not set a therapy plan, and does not decide when treatment ends. It measures and it suggests. Every clinical decision passes through a licensed clinician β and that is not legal hedging but a design choice: a tool that tries to replace professional judgement fails at both jobs.
What is still hard
Non-native accents, bilingual households, and children with atypical oral structure. Confidence drops in all three, and the engine is supposed to admit that rather than guess. This is the active work right now.
What gets measured
There are no success percentages to boast about here, because a system in development that reports outcomes has lied to someone. These are the metrics success is judged on β and when there is verified data, it will be published here in full, including the parts that disappoint.
Error detection agreement
How often the engine agrees with a clinician who listened to the same recording blind.
False alarm rate
How often an error is flagged that was not there. More damaging than a miss β it breaks a child's trust.
Practice streak
Consecutive days of practice. The single measure that actually predicts progress.
Time to feedback
Past about a second, a child loses the connection between what they said and what they are told.
Roadmap
Where this stands and what comes next. Current as of July 2026.
-
Done
Children's speech corpus, ages 4β10
Recordings labelled at the phoneme level, containing real disordered speech rather than simulated errors. Everything else rests on this.
-
Done
Alignment and scoring engine
Error classification into substitution, omission, insertion and distortion, with a confidence score per sound.
-
In progress
Clinic pilot
Controlled use with a small number of families, closely supervised. The goal is to find where the engine is wrong, not to prove it is right.
-
Next
Parent application
Daily practice and a plain summary, without gamification that pulls attention away from the speech itself.
-
Later
Clinician dashboard
Managing a caseload, building practice plans, and exporting summaries into clinical records.
-
Research
Additional languages
The same engineering problem in languages with almost no coverage. Gated entirely on collecting appropriate data.
Questions we get
Does this replace a speech therapist?
No, and it was not built to. The tool only works inside a plan a clinician designed. Without that it is a pronunciation game with no destination.
Are my child's recordings stored?
Not by default. Analysis runs on the device and only the result is sent. Storing audio requires explicit consent and can be withdrawn at any time.
What age does it work from?
Around four. Below that the engine's error rate is too high, and practice needs an adult leading it in any case.
My child speaks another language at home β is that a problem?
It reduces accuracy, and that is reported openly rather than hidden. For now the tool suits children whose main language is the one being assessed.
What device do we need?
A phone or tablet from the last few years, plus headphones with a microphone if the room is noisy. No special equipment.
Work in this field and want to talk?
Clinicians, researchers, engineers β a professional conversation is always welcome, with no commercial angle attached.